Nellie Liang is a senior fellow in the Hutchins Center on Fiscal and Monetary Policy in the Economic Studies program at Brookings. From July 2021 to January 2025, she served as Under Secretary of the Treasury for domestic finance, the office that oversees Treasury debt issuance. This post is drawn, in part, from her April 8, 2025, testimony at a House Financial Services Committee hearing, which you can read in full here.

Why is the market for US Treasury debt so important?

There is no larger thoroughfare for global capital than the market for U.S. Treasury debt securities. It averages around $900 billion in transactions per day, with high volume days in recent years around $1.5 trillion. In addition, there is roughly $4 trillion in Treasury repurchase agreement—or repo—financing each day. And average daily trading volume in U.S. Treasury futures was $645 billion in notional in 2023 and higher in 2024.

The Treasury market serves several critical functions. It is key for financing the U.S. government at the lowest cost to the taxpayer. It is an important channel for the Federal Reserve’s monetary policy. It provides the benchmark risk-free yield curve for pricing risky assets. And it serves as a key source of safe and liquid assets for investors and is used for liquidity risk management by many financial firms, both banks and nonbanks. To serve these critical functions, the Treasury market needs to be deep and liquid. Market participants need to be able to sell Treasury securities without materially impacting the price and at low transaction costs, not only under normal economic conditions but also during periods of uncertainty and stress.

What fragilities in the Treasury market are concerning?

The Treasury market has changed significantly in the past couple of decades in ways that make market liquidity less resilient to big shocks. The amount of Treasury market debt has increased enormously because of large federal deficits. At the same time, intermediation has changed significantly. Traditional securities dealers have pulled back from market making after capital standards and risk management practices were strengthened after the Global Financial Crisis (GFC) in 2008. Electronic trading has increased and principal trading firms, typically with smaller capital cushions than securities dealers, now represent most of the trading in electronic, inter-dealer markets. And the investor base is more price sensitive, as private funds with leverage or redemption pressures have increased their holdings of Treasuries, and the share held by foreign official entities who are less price-sensitive has fallen. The share of Treasuries held by money market funds, mutual funds, and hedge funds has risen to more than 27%, while the share held by foreign holders has fallen from about 50% in 2015 to 30% now.

What lessons did we learn from what happened in the Treasury market in March 2020 at the onset of COVID?

When COVID hit, it was far from clear what it would mean for the economy and for day-to-day life. In the Treasury market, market liquidity deteriorated by much more than expected; dealers and other intermediaries’ capacity to buy Treasury debt was overwhelmed by selling of Treasuries by investors who faced funding pressures or others who wanted ultra-safe cash—the “dash for cash.” Open-end bond mutual funds and hedge funds needed cash to meet margin calls or to satisfy investor redemptions. They chose to sell their most liquid securities—Treasuries. This surge in desired selling exceeded the ability or willingness of dealers to supply liquidity in the face of unprecedented risks and disruptions to normal practices, as many traders were sent to work from home. Treasury prices fell and interest rates rose sharply, especially for off-the-run securities (those issued before the most recent issue and which are still outstanding). The sharp rise in rates contrasted sharply with past episodes of high uncertainty when the flight of investors to the safe haven of Treasury securities would drive down interest rates.

Market functioning was restored only after the Federal Reserve began purchasing huge amounts of Treasury securities to provide liquidity. To put this in perspective, the Fed bought $80 million a month on long-term government debt in the third round of Quantitative Easing (QE3) in 2012. In the week of March 25, 2020, the Fed bought $360 billion in Treasuries. The Fed ultimately committed to “purchase Treasury securities and agency mortgage-backed securities purchases in the amounts needed to support smooth market functioning and the effective transmission of monetary policy to broader financial conditions.”

Fortunately, the Fed’s purchases to restore market functioning in March 2020 were aligned with its monetary policy objectives at the time—to stimulate the economy and raise inflation to its 2% target. It is possible, however, that the Fed may someday confront the need to purchase Treasury securities at a time when doing so would conflict with achieving its mandate of maximum employment and price stability. Avoiding this conflict underscores the importance of regulatory reforms to strengthen Treasury market resilience.

What happened in the Treasury market in early April 2025?

The tariffs President Trump announced on April 2 were bigger and broader than expected and created massive uncertainty for the economy. Financial market volatility soared. Treasury yields fell initially amid growing concerns about a recession, but the yield on longer-term Treasury debt began to rise later that week as investors reassessed the prospects for higher inflation and weaker growth over a longer horizon.

With the escalation of tariffs and no clear signs of an off-ramp, financial asset prices became even more volatile. The rise in the 10-year Treasury yield accelerated with unusual speed from less than 4% percent on Friday, April 4, to spike to 4.5% intra-day on Tuesday, April 8, and the 30-year yield topped 5%. Market liquidity declined as intermediaries often pull back from risk-taking and raise transaction costs when volatility is high. The spread of Treasury yields-to-OIS (also known as the interest rate swap) at longer maturities, a measure of the demand for Treasury securities relative to the fixed rate on the interest rate swap, widened sharply late Monday, April 7, raising concerns that market liquidity for Treasury securities was deteriorating further. Amid the turmoil, the Treasury’s auction of 3-year Treasury notes was weaker than usual. Investors began questioning whether markets would become dysfunctional, as they were in March 2020.

But the worst fears about market functioning eased on Wednesday, April 9, when demand was strong at the Treasury’s auction of 10-year bonds and after President Trump announced a 90-day delay in tariffs for some countries, though he increased them with China. The 30-year auction also went well the next day. The Treasury yield curve, however, still remains steeper than before the tariffs were announced, and yields on 10-year and 30-year Treasury securities are up around 25-50 basis points from their recent lows on April 4.

Some investors speculated that the sharp rise in yields of longer-term Treasuries indicated that leveraged hedge funds were facing funding pressures and that resulting sales would force Treasury yields even higher. Indicators from the repo market suggest some funding pressures but not of the magnitude to explain the swap spread, though some believe they could have worsened materially if uncertainty remained high or increased further. Others speculated that some of the rise in rates came from increasing doubts about Treasury securities as the pre-eminent global safe-haven asset, consistent with the decline in the dollar. A re-pricing of Treasury debt for this reason would be very consequential, forcing the U.S. government to pay more to borrow to finance deficits and raising the costs of borrowing for businesses and households. It will take time with detailed data that the financial regulators have to disentangle how much of the rise in longer-term yields was due to fundamental revisions in the outlook for the economy and inflation, to market illiquidity because of deleveraging pressures that may have forced sales of Treasuries to raise cash, or how much could be from sales because investors repriced the safe-haven quality of Treasury securities. But available evidence suggests that the current episode so far is not a repeat of the market dysfunction in March 2020 from a cash-basis unwind by hedge funds and redemptions from bond funds.

What are the most significant steps taken in the past four years to strengthen the resilience of the Treasury markets?

The work of the Inter-Agency Working Group on Treasury Market Surveillance (IAWG)—composed of staff from the Treasury, the Fed, and financial regulatory agencies—to strengthen Treasury market resilience has focused on expanding dealer intermediation capacity, to reduce surges in demand for liquidity (sales of Treasuries) in periods of high stress, to modernize the infrastructure, and to improve transparency and visibility into the market.

There have been some significant accomplishments. To highlight a few, much more data on Treasury securities transactions and on hedge funds are being disclosed to the public. New data are being collected on a key segment of the repo market—the bilateral uncleared repo market, where dealers finance their hedge fund clients. This market represents about one-half of the total repo market but has been opaque to authorities since before the GFC. In addition, Treasury initiated a buyback program to allow dealers to sell off-the-run securities on a predictable basis to help free up their balance sheets. The Fed put in place a standing facility to finance Treasury repo with pre-authorized dealers and banks which could encourage dealers to invest in market-making capacity and support liquidity in times of market stress.

What’s left to be done?

Of special importance is complete implementation of the Securities and Exchange Commission’s rule to mandate more central clearing of Treasury and repo. Central clearing is used for other assets and can reduce risk by standardizing risk management requirements and increase intermediation capacity through multilateral netting. There are many complicated operational, regulatory, and accounting issues to resolve, and industry groups are actively engaged and committed to addressing them. They recently received an extension of the deadlines by one year for central clearing Treasury securities in December 2026 and Treasury repo in June 2027. As I noted in my recent Congressional testimony, I don’t believe they should delay further.

Changes should be considered to the supplementary leverage ratio (SLR) put in place following the GFC. The SLR requires banking firms to hold the same amount of capital for riskless reserves at the central bank as they would for risky assets. One change could be to exclude central bank reserves and perhaps Treasury securities in the trading book (but not all Treasuries because they have interest rate risk) from the SLR calculation, but importantly, with an adjustment so that there would not be a reduction in bank capital. This adjustment would preserve safety and soundness and would let the SLR function as intended as a backstop.

To reduce surges in selling in periods of stress, open-end bond funds should be required to reduce significant liquidity mismatches, which force Treasury sales. In addition, supervisors should prevent excessive leverage of hedge funds in trades, such as the cash-futures basis trade and others, that can force rapid, disorderly unwinds of positions. The new data on bilateral repo transactions and from Form PF filings, as well as from when central clearing is implemented, should provide much greater visibility into leverage in Treasury markets.

These reforms are complementary and interconnected, and, in my view, all should continue to be pursued. Prudent changes to the SLR would not on its own significantly improve intermediation and strengthen resilience in stress periods, especially as the amount of Treasury debt continues to grow. Central clearing as well as incentives from the SRF are needed to increase intermediation capacity, and stronger rules and supervision to reduce liquidity mismatch and leverage of investors also are needed.

Superagency in the workplace: Empowering people to unlock AI’s full potential (47 pages)

Artificial intelligence has arrived in the workplace and has the potential to be as transformative as the steam engine was to the 19th-century Industrial Revolution. With powerful and capable large language models (LLMs) developed by Anthropic, Cohere, Google, Meta, Mistral, OpenAI, and others, we have entered a new information technology era. McKinsey research sizes the long-term AI opportunity at $4.4 trillion in added productivity growth potential from corporate use cases.

Therein lies the challenge: the long-term potential of AI is great, but the short-term returns are unclear. Over the next three years, 92 percent of companies plan to increase their AI investments. But while nearly all companies are investing in AI, only 1 percent of leaders call their companies “mature” on the deployment spectrum, meaning that AI is fully integrated into workflows and drives substantial business outcomes. The big question is how business leaders can deploy capital and steer their organizations closer to AI maturity.

This research report, prompted by Reid Hoffman’s book Superagency: What Could Possibly Go Right with Our AI Future, asks a similar question: How can companies harness AI to amplify human agency and unlock new levels of creativity and productivity in the workplace? AI could drive enormous positive and disruptive change. This transformation will take some time, but leaders must not be dissuaded. Instead, they must advance boldly today to avoid becoming uncompetitive tomorrow. The history of major economic and technological shifts shows that such moments can define the rise and fall of companies. Over 40 years ago, the internet was born. Since then, companies including Alphabet, Amazon, Apple, Meta, and Microsoft have attained trillion-dollar market capitalizations. Even more profoundly, the internet changed the anatomy of work and access to information. AI now is like the internet many years ago: The risk for business leaders is not thinking too big, but rather too small.

Register for upcoming McKinsey Live Amid the AI boom, compute power is emerging as one of this decade’s most critical resources, influencing the pace at which AI is deployed. Data centers train the foundation models and machine learning applications that underpin all AI technology. The hardware, processors, memory, storage, and energy needed to operate these data centers are collectively known as compute power—and there is a seemingly unquenchable need for more. Meeting this demand is not simply a race to invest trillions of dollars. Join McKinsey senior partners Humayun Tai and Pankaj Sachdeva on April 23 at 10:30a.m. EDT / 4:30p.m. CEST as they share how business leaders can navigate this complex environment.

This report explores companies’ technology and business readiness for AI adoption (see sidebar “About the survey”). It concludes that employees are ready for AI. The biggest barrier to success is leadership.

About the survey To create our report, we surveyed 3,613 employees (managers and independent contributors) and 238 C-level executives in October and November 2024. Of these, 81 percent came from the United States, and the rest came from five other countries: Australia, India, New Zealand, Singapore, and the United Kingdom. The employees spanned many roles, including business development, finance, marketing, product management, sales, and technology. All the survey findings discussed in the report, aside from two sidebars presenting international nuances, pertain solely to US workplaces. The findings are organized in this way because the responses from US employees and C-suite executives provide statistically significant conclusions about the US workplace. Analyzing global findings separately allows a comparison of differences between US responses and those from other regions.

Chapter 1 looks at the rapid advancement of technology over the past two years and its implications for business adoption of AI.

Chapter 2 delves into the attitudes and perceptions of employees and leaders. Our research shows that employees are more ready for AI than their leaders imagine. In fact, they are already using AI on a regular basis; are three times more likely than leaders realize to believe that AI will replace 30 percent of their work in the next year; and are eager to gain AI skills. Still, AI optimists are only a slight majority in the workplace; a large minority (41 percent) are more apprehensive and will need additional support. This is where millennials, who are the most familiar with AI and are often in managerial roles, can be strong advocates for change.

Chapter 3 looks at the need for speed and safety in AI deployment. While leaders and employees want to move faster, trust and safety are top concerns. About half of employees worry about AI inaccuracy and cybersecurity risks. That said, employees express greater confidence that their own companies, versus other organizations, will get AI right. The onus is on business leaders to prove them right, by making bold and responsible decisions.

Chapter 4 examines how companies risk losing ground in the AI race if leaders do not set bold goals. As the hype around AI subsides, companies should put a heightened focus on practical applications that empower employees in their daily jobs. These applications can create competitive moats and generate measurable ROI. Across industries, functions, and geographies, companies that invest strategically can go beyond using AI to drive incremental value and instead create transformative change.

Chapter 5 looks at what is required for leaders to set their teams up for success with AI. The challenge of AI in the workplace is not a technology challenge. It is a business challenge that calls upon leaders to align teams, address AI headwinds, and rewire their companies for change.

Chapter 1 An innovation as powerful as the steam engine

Imagine a world where machines not only perform physical labor but also think, learn, and make autonomous decisions. This world includes humans in the loop, bringing people and machines together in a state of superagency that increases personal productivity and creativity (see sidebar “AI superagency”). This is the transformative potential of AI, a technology with a potential impact poised to surpass even the biggest innovations of the past, from the printing press to the automobile. AI does not just automate tasks but goes further by automating cognitive functions. Unlike any invention before, AI-powered software can adapt, plan, guide—and even make—decisions. That’s why AI can be a catalyst for unprecedented economic growth and societal change in virtually every aspect of life. It will reshape our interaction with technology and with one another.

Scientific discoveries and technological innovations are stones in the cathedral of human progress. Reid Hoffman, cofounder of LinkedIn and Inflection AI, partner at Greylock Partners, and author

Many breakthrough technologies, including the internet, smartphones, and cloud computing, have transformed the way we live and work. AI stands out from these inventions because it offers more than access to information. It can summarize, code, reason, engage in a dialogue, and make choices. AI can lower skill barriers, helping more people acquire proficiency in more fields, in any language and at any time. AI holds the potential to shift the way people access and use knowledge. The result will be more efficient and effective problem solving, enabling innovation that benefits everyone.

AI superagency What impact will AI have on humanity? Reid Hoffman and Greg Beato’s book Superagency: What Could Possibly Go Right with Our AI Future (Authors Equity, January 2025) explores this question. The book highlights how AI could enhance human agency and heighten our potential. It envisions a human-led, future-forward approach to AI. Superagency, a term coined by Hoffman, describes a state where individuals, empowered by AI, supercharge their creativity, productivity, and positive impact. Even those not directly engaging with AI can benefit from its broader effects on knowledge, efficiency, and innovation. AI is the latest in a series of transformative supertools, including the steam engine, internet, and smartphone, that have reshaped our world by amplifying human capabilities. Like its predecessors, AI can democratize access to knowledge and automate tasks, assuming humans can develop and deploy it safely and equitably.

Over the past two years, AI has advanced in leaps and bounds, and enterprise-level adoption has accelerated due to lower costs and greater access to capabilities. Many notable AI innovations have emerged (Exhibit 1). For example, we have seen a rapid expansion of context windows, or the short-term memory of LLMs. The larger a context window, the more information an LLM can process at once. To illustrate, Google’s Gemini 1.5 could process one million tokens in February 2024, while its Gemini 1.5 Pro could process two million tokens by June of that same year. Overall, we see five big innovations for business that are driving the next wave of impact: enhanced intelligence and reasoning capabilities, agentic AI, multimodality, improved hardware innovation and computational power, and increased transparency.

Image description begins: The text-based exhibit illustrates the evolution of capabilities of several gen AI large language models, or LLMs, from select frontier labs between 2022 and 2025. The is presented as a table comparing two time periods: 2022-2023 and January 2025. For each of five LLMs—Anthropic’s Claude, Google’s Gemini, Meta’s Llama, Microsoft’s Phi, and OpenAI’s GPT—the exhibit shows a list of capabilities for each time period. In 2022-2023, all five platforms lacked multimodal capabilities, functioning primarily with text only. Anthropic’s Claude, for example, showed limited contextual understanding and no tool usage. Google’s Gemini, similarly, had limited real-time data integration and low personalization. Meta’s Llama 1 exhibited fair reasoning but had difficulty with complex conversations and lacked API access. Microsoft’s Phi-1 had fair reasoning limited to coding tasks, with focused training on a smaller dataset. OpenAI’s GPT-3.5 demonstrated fair reasoning, scoring well on the SAT but poorly on the bar examination, while also displaying limited contextual understanding in complex conversations, though it did offer standard API access for text generation. By January 2025, a significant shift is apparent. Claude 3.5, Gemini 2.0 Flash, Llama 3.3, Phi-4, and OpenAI’s model o1 all gained multimodal capabilities, incorporating text, audio, and images. Advanced reasoning capabilities, capable of multistep problem-solving and nuanced analysis, became common across most of the platforms. Enhanced contextual understanding, maintaining coherence during long dialogues, is also highlighted as an improvement. Furthermore, real-time data integration and advanced personalization features were added to some platforms. Finally, several platforms highlight improved or advanced API access, allowing for tools related to model and agent development and multimodal inputs. Source: Company websites and press releases. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Intelligence and reasoning are improving

AI is becoming far more intelligent. One indicator is the performance of LLMs on standardized tests. OpenAI’s Chat GPT-3.5, introduced in 2022, demonstrated strong performance on high-school-level exams (for example, scoring in the 70th percentile on the SAT math and the 87th percentile on the SAT verbal sections). However, it often struggled with broader reasoning. Today’s models are near the intelligence level of people who hold advanced degrees. GPT-4 can so easily pass the Uniform Bar Examination that it would rank in the top 10 percent of test takers, and it can answer 90 percent of questions correctly on the US Medical Licensing Examination.

The advent of reasoning capabilities represents the next big leap forward for AI. Reasoning enhances AI’s capacity for complex decision making, allowing models to move beyond basic comprehension to nuanced understanding and the ability to create step-by-step plans to achieve goals. For businesses, this means they can fine-tune reasoning models and integrate them with domain-specific knowledge to deliver actionable insights with greater accuracy. Models such as OpenAI’s o1 or Google’s Gemini 2.0 Flash Thinking Mode are capable of reasoning in their responses, which gives users a human-like thought partner for their interactions, not just an information retrieval and synthesis engine.

AI in Action: An interactive learning journey

Agentic AI is acting autonomously

I’ve always thought of AI as the most profound technology humanity is working on . . . more profound than fire or electricity or anything that we’ve done in the past. Sundar Pichai, CEO of Alphabet

The ability to reason is growing more and more, allowing models to autonomously take actions and complete complex tasks across workflows. This is a profound step forward. As an example, in 2023, an AI bot could support call center representatives by synthesizing and summarizing large volumes of data—including voice messages, text, and technical specifications—to suggest responses to customer queries. In 2025, an AI agent can converse with a customer and plan the actions it will take afterward—for example, processing a payment, checking for fraud, and completing a shipping action.

Software companies are embedding agentic AI capabilities into their core products. For example, Salesforce’s Agentforce is a new layer on its existing platform that enables users to easily build and deploy autonomous AI agents to handle complex tasks across workflows, such as simulating product launches and orchestrating marketing campaigns. Marc Benioff, Salesforce cofounder, chair, and CEO, describes this as providing a “digital workforce” where humans and automated agents work together to achieve customer outcomes.

Multimodality is bringing together text, audio, and video

Today’s AI models are evolving toward more advanced and diverse data processing capabilities across text, audio, and video. Over the last two years, we have seen improvements in the quality of each modality. For example, Google’s Gemini Live has improved audio quality and latency and can now deliver a human-like conversation with emotional nuance and expressiveness. Also, demonstrations of Sora by OpenAI show its ability to translate text to video.

Hardware innovation is enhancing performance

Hardware innovation and the resulting increase in compute power continue to enhance AI performance. Specialized chips allow faster, larger, and more versatile models. Enterprises can now adopt AI solutions that require high processing power, enabling real-time applications and opportunities for scalability. For example, an e-commerce company could significantly improve customer service by implementing AI-driven chatbots that leverage advanced graphics processing units (GPUs) and tensor processing units (TPUs). Using distributed cloud computing, the company could ensure optimal performance during peak traffic periods. Integrating edge hardware, the company could deploy models that analyze photos of damaged products to more accurately process insurance claims.

Transparency is increasing

AI, like most transformative technologies, grows gradually, then arrives suddenly. Reid Hoffman, cofounder of LinkedIn and Inflection AI, partner at Greylock Partners, and author

AI is gradually becoming less risky, but it still lacks greater transparency and explainability. Both are critical for improving AI safety and reducing the potential for bias, which are imperative for widescale enterprise deployment. There is still a long way to go, but new models and iterations are rapidly improving. Stanford University’s Center for Research on Foundation Models (CRFM) reports significant advances in model performance. Its Transparency Index, which uses a scale of 1 to 100, shows that Anthropic’s transparency score increased by 15 points to 51 and Amazon’s more than tripled to 41 between October 2023 and May 2024.

Beyond LLMs, other forms of AI and machine learning (ML) are improving explainability, allowing the outputs of models that support consequential decisions (for example, credit risk assessment) to be traced back to the data that informed them. In this way, critical systems can be tested and monitored on a near-constant basis for bias and other everyday harms that arise from model drift and shifting data inputs, which happens even in systems that were well calibrated before deployment.

All of this is crucial for detecting errors and ensuring compliance with regulations and company policies. Companies have improved explainability practices and built necessary checks and balances, but they must be prepared to evolve continuously to keep up with growing model capabilities.

Achieving AI superagency in the workplace is not simply about mastering technology. It is every bit as much about supporting people, creating processes, and managing governance. The next chapters explore the nontechnological factors that will help shape the deployment of AI in the workplace.

Chapter 2 Employees are ready for AI; now leaders must step up

Employees will be the ones to make their organizations AI powerhouses. They are more ready to embrace AI in the workplace than business leaders imagine. They are more familiar with AI tools, they want more support and training, and they are more likely to believe AI will replace at least a third of their work in the near future. Now it’s imperative that leaders step up. They have more permission space than they realize, so it’s on them to be bold and capture the value of AI. Now.

People are using [AI] to create amazing things. If we could see what each of us can do 10 or 20 years in the future, it would astonish us today. Sam Altman, cofounder and CEO of OpenAI

Beyond the tipping point

In our survey, nearly all employees (94 percent) and C-suite leaders (99 percent) report having some level of familiarity with gen AI tools. Nevertheless, business leaders underestimate how extensively their employees are using gen AI. C-suite leaders estimate that only 4 percent of employees use gen AI for at least 30 percent of their daily work, when in fact that percentage is three times greater, as self-reported by employees (Exhibit 2). And while only a total of 20 percent of leaders believe employees will use gen AI for more than 30 percent of their daily tasks within a year, employees are twice as likely (47 percent) to believe they will (see sidebar “Who is using AI at work? Nearly everyone, even skeptical employees”).

The good news is that our survey suggests three ways companies can accelerate AI adoption and move toward AI maturity.

Image description begins: The exhibit shows the anticipated timeline for US employees’ and business leaders’ use of gen AI for more than 30 percent of their daily work tasks, presented as two stacked bar charts, one for C-suite respondents and one for employees. The segments are broken down into five categories representing different timeframes: Already using, less than a year, 1-5 years, over 5 years, and don’t anticipate it. A final category, not sure, is also included. A key finding highlighted in the chart is that employees are three times more likely to be using gen AI today than their leaders expect (4 percent of C-suite respondents estimate that employees are currently using gen AI for more than 30 percent of their daily tasks, while 13 percent of employees self-report they are currently doing so). For the C-suite, 16 percent expect employees to start using gen AI for more than 30 percent of their daily tasks within less than a year, 56 percent anticipate such adoption within 1-5 years, 11 percent expect it in over 5 years, and 10 percent don’t anticipate employees will ever use gen AI for 30 percent of their work tasks. 3 percent of C-suite respondents are unsure. 34 percent of employees expect to use gen AI for more than 30 percent of their work tasks in less than a year, 37 percent within 1-5 years, 5 percent in over 5 years, and 7 percent don’t anticipate ever using it in this way. 4 percent of employees are unsure. Source: McKinsey US CxO survey, Oct–Nov 2024; McKinsey US employee survey, Oct–Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Leaders can invest more in their employees

Who is using AI at work? Nearly everyone, even skeptical employees Our research looked at people who self-identify as “Zoomers,” “Bloomers,” “Gloomers,” and “Doomers” in their attitudes toward AI—a set of archetypes introduced in Superagency. We find that 39 percent of employees identify as Bloomers, who are AI optimists that want to collaborate with their companies to create responsible solutions. Meanwhile, 37 percent identify as Gloomers, who are more skeptical about AI and want extensive top-down AI regulations; 20 percent identify as Zoomers, who want AI to be quickly deployed with few guardrails; and just 4 percent identify as Doomers, who have a fundamentally negative view of AI (exhibit). Even those with a skeptical take on AI are familiar with it; 94 percent of Gloomers and 71 percent of Doomers say they have some familiarity with gen AI tools. Furthermore, approximately 80 percent of Gloomers and about half of Doomers say they are comfortable using gen AI at work. Image descriptions begins: The exhibit depicts US employee sentiment toward gen AI, categorized into four archetypes: Doomer, Gloomer, Bloomer, and Zoomer. Each archetype’s perspective is presented through a series of semicircular sunray charts showing the share of respondents within each group holding specific views. For example, two of the sunrays represent two separate sentiments, “has extensive familiarity with gen AI” and “has at least some familiarity with gen AI.” The Doomer archetype shows 16 percent with extensive familiarity and 71 percent with at least some familiarity. The Gloomer archetype demonstrates significantly higher percentages: 42 percent with extensive familiarity and 94 percent with at least some. The Bloomer archetype shows 55 percent with extensive familiarity and 96 percent with at least some, and the Zoomer archetype shows 67 percent with extensive familiarity and 96 percent with at least some. The exhibit further illustrates employees’ comfort levels with using gen AI results, belief in the net benefits of gen AI within the next five years, and plans to utilize gen AI more in their personal lives. In the Doomer archetype, 47 percent say they are comfortable using gen AI results, 54 percent believe in gen AI’s net benefit within the next five years, and 49 percent plan increased personal use of gen AI. The Gloomer archetype shows markedly higher percentages in these three areas: 79 percent, 82 percent, and 77 percent respectively. The Bloomer and Zoomer archetypes present even higher percentages across these three metrics; for instance, 91 percent of Zoomers are comfortable using gen AI results, 87 percent believe in gen AI’s net benefit within five years, and 85 percent plan to increase their personal use of gen AI. Finally, the exhibit the includes a separate section depicting the share of respondents within each archetype, indicating the size of each group with a series of donut charts. The Doomer group comprises 4 percent of employees, Gloomers are 37 percent, Bloomers are 39 percent, and Zoomers are 20 percent. Source: McKinsey US employee survey, Oct–Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

As noted at the beginning of this chapter, employees anticipate AI will have a dramatic impact on their work. Now they would like their companies to invest in the training that will help them succeed. Nearly half of employees in our survey say they want more formal training and believe it is the best way to boost AI adoption. They also would like access to AI tools in the form of betas or pilots, and they indicate that incentives such as financial rewards and recognition can improve uptake.

Yet employees are not getting the training and support they need. More than a fifth report that they have received minimal to no support (Exhibit 3). Outside the United States, employees also want more training (see sidebar “Global perspectives on training”).

Image description begins: The first section of the exhibit is a horizontal bar chart showing the percentage of US employees who believe that specific company initiatives would increase their daily use of gen AI tools. Formal gen AI training from their organization scored highest at 48 percent, followed by seamless integration into existing workflows (45 percent), access to gen AI tools (41 percent), and incentives and rewards (40 percent). Lower percentages were observed for usage of gen AI being a requirement for a certification program (30 percent), explicit instructions from managers to use gen AI (30 percent), being involved in the development of the tools (29 percent), and OKRs/KPIs tied to gen AI usage (22 percent). The second section is a stacked pair of segmented bar charts illustrating the perceived level of support for gen AI capability building at their organizations, comparing current vs in three years. This chart shows the distribution of responses across four levels of support: not needed, none/minimal, moderate to significant, and fully supported. Currently, 6 percent of employees report that support for gen AI in their organizations is not needed, 22 percent report they receive none/minimal support, 44 percent report moderate to significant support, and 29 percent report they are fully supported. Looking ahead to three years in the future, these percentages are projected to shift considerably: gen AI support not needed drops to 4 percent, none/minimal support for gen AI usage decreases to 10 percent, moderate to significant support for gen AI usage increases to 56 percent, and fully supported increases to 31 percent. Source: McKinsey US employee survey, Oct–Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Sidebar Global perspectives on training To get a clearer picture of global AI adoption trends, we looked at trends across five countries: Australia, India, New Zealand, Singapore, and the United Kingdom. Broadly speaking, these employees and C-suite leaders—the “international” group in this report—have similar views of AI as their US peers. In some key areas, however, including the topic of training, their experiences differ. Many international employees are concerned about insufficient training, even though they report receiving far more support than US employees. Some 84 percent of international employees say they receive significant or full organizational support to learn AI skills, versus just over half of US employees. International employees also have more opportunities to participate in developing gen AI tools at work than their US counterparts, with differences of at least ten percentage points in activities such as providing feedback, beta testing, and requesting specific features (exhibit).

C-suite leaders can help millennials lead the way

Many millennials aged 35 to 44 are managers and team leaders in their companies. In our survey, they self-report having the most experience and enthusiasm about AI, making them natural champions of transformational change. Millennials are the most active generation of AI users. Some 62 percent of 35- to 44-year-old employees report high levels of expertise with AI, compared with 50 percent of 18- to 24-year-old Gen Zers and 22 percent of baby boomers over 65 (Exhibit 4). By tapping into that enthusiasm and expertise, leaders can help millennials play a crucial role in AI adoption.

Image description begins: The exhibit is a grid of proportional area charts displaying US employee sentiment toward gen AI by age group. Each row represents a different sentiment, from top to bottom: has extensive familiarity with gen AI, is comfortable using gen AI at work, provides feedback on gen AI tools, and wants to participate in the design of gen AI tools. The columns represent age groups: 18-24, 25-34, 35-44, 45-54, 55-64, and 65+. The data is presented as percentages of respondents who agreed with each sentiment within each age group. The chart reveals that the 35-44 age group exhibits the most positive sentiment across most categories. For example, 90 percent of this group reports being comfortable using gen AI at work, the highest percentage among all age groups for this metric. This group also shows the highest percentage (62 percent) reporting extensive familiarity with gen AI. In contrast, the 55-64 and 65+ age groups consistently show lower percentages across all four metrics, with only 26 percent and 22 percent of employees in these age groups reporting extensive familiarity with gen AI respectively. The 18-24, 25-34, and 45-54 age groups show intermediate levels of positive sentiment, generally lower than the 35-44 group but higher than the 55-64 and 65+ age groups. Source: McKinsey US employee survey, Oct–Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Since many millennials are managers, they can support their teams to become more adept AI users. This helps push their companies toward AI maturity. Two-thirds of managers say they field questions from their team about how to use AI tools at least once a week, and a similar percentage say they recommend AI tools to their teams to solve problems (Exhibit 5).

Image description begins: The exhibit examines US manager respondents and their experiences with gen AI tools. The exhibit is composed of two main sections. The top section of the exhibit examines the frequency of inquiries that managers field from their employees about using new gen AI tools at work. This is depicted as a horizontal bar chart showing percentages of respondents. 5 percent of managers report less than quarterly inquiries; 5 percent report quarterly inquiries; 12 percent report inquiries once a month; 15 percent report once a week; 28 percent report a few times a week; 9 percent report once a day; and 16 percent report multiple times a day. Finally, 10 percent of report not at all. The second section explores the use of gen AI tools to resolve team member challenges. This section uses two donut charts, each showing percentages of respondents. The first donut chart indicates that 68 percent of managers report recommending a gen AI tool to solve a team member’s challenge in the past month. The second donut chart shows that 86 percent of managers who recommended a gen AI tool report that the tool was successful in resolving the team member’s challenge. Source: McKinsey US employee survey, Oct–Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Since leaders have the permission space, they can be bolder

In many transformations, employees are not ready for change, but AI is different. Employee readiness and familiarity are high, which gives business leaders the permission space to act. Leaders can listen to employees describe how they are using AI today and how they envision their work being transformed. They also can provide employees with much-needed training and empower managers to move AI use cases from pilot to scale.

It’s critical that leaders meet this moment. It’s the only way to accelerate the probability that their companies will reach AI maturity. But they must move with alacrity, or they will fall behind.

Chapter 3 Delivering speed and safety

AI technology is advancing at record speed. ChatGPT was released about two years ago; OpenAI reports that usage now exceeds 300 million weekly users and that over 90 percent of Fortune 500 companies employ its technology. The internet did not reach this level of usage until the early 2000s, nearly a decade after its inception.

Soon after the first automobiles were on the road, there was the first car crash. But we didn’t ban cars—we adopted speed limits, safety standards, licensing requirements, drunk-driving laws, and other rules of the road. Bill Gates, cofounder of Microsoft

The majority of employees describe themselves as AI optimists; Zoomers and Bloomers make up 59 percent of the workplace. Even Gloomers, who are one of the two less-optimistic segments in our analysis, report high levels of gen AI familiarity, with over a quarter saying they plan to use AI more next year.

Business leaders need to embrace this speed and optimism to ensure that their companies don’t get left behind. Yet despite all the excitement and early experimentation, 47 percent of C-suite leaders say their organizations are developing and releasing gen AI tools too slowly, citing talent skill gaps as a key reason for the delay (Exhibit 6).

Image description begins: The exhibit shows US C-suite executive sentiment toward the pace of development and release of gen AI tools within their organizations, in the form of two segmented bar charts. The first bar chart presents the overall perception of the pace, where 47 percent of respondents find the pace to be too slow, while 45 percent feel it is about right, and a smaller 9 percent consider it too fast. The second bar chart delves into the top reasons behind the perceived slow pace of gen AI tool development and release in executives’ organizations, focusing on the responses from those who indicated that development was too slow. The most prominent reason cited is talent skill gaps, accounting for 46 percent of these responses. Resourcing constraints followed closely, with 38 percent of respondents identifying this as a key factor. Complex approval process and technical complexity each receive 8 percent of the responses. Source: McKinsey US CxO survey, Oct-Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Business leaders are trying to meet the need for speed by increasing investments in AI. Of the executives surveyed, 92 percent say they expect to boost spending on AI in the next three years, with 55 percent expecting investments to increase by at least 10 percent from current levels. But they can no longer just spend on AI without expecting results. As companies move on from the initial thrill of gen AI, business leaders face increasing pressure to generate ROI from their gen AI deployments.

We are at a turning point. The initial AI excitement may be waning, but the technology is accelerating. Bold and purposeful strategies are needed to set the stage for future success. Leaders are taking the first step: One quarter of those executives we surveyed have defined a gen AI road map, while just over half have a draft that is being refined (Exhibit 7). With technology changing this fast, all road maps and plans will evolve constantly. For leaders, the key is to make some clear choices about what valuable opportunities they choose to pursue first—and how they will work together with peers, teams, and partners to deliver that value.

Image description begins: The exhibit is comprised of two horizontal segmented bar charts. The first chart displays the share of US C-suite respondents who have a defined gen AI roadmap. 21 percent report not currently having a roadmap but one was in progress, 53 percent indicate having a roadmap that is still being refined, and 25 percent state that a comprehensive roadmap is already in place. The second bar chart illustrates the level to which US C-suite respondents have identified revenue-generating use cases for gen AI. 1 percent of respondents indicate they have not yet identified any such use cases, while 10 percent report they have minimally identified, 38 percent have partially identified, 39 percent have mostly identified, and 12 percent have fully identified such use cases. Source: McKinsey US CxO survey, Oct-Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends

The dilemma of speed versus safety

There’s a spanner in the works: Regulation and safety often continue to be seen as insurmountable challenges rather than opportunities. Leaders want to increase AI investments and accelerate development, but they wrestle with how to make AI safe in the workplace. Data security, hallucinations, biased outputs, and misuse (for example, creating harmful content or enabling fraud) are challenges that cannot be ignored. Employees are well aware of AI’s safety challenges. Their top concerns are cybersecurity, privacy, and accuracy (Exhibit 8). But what will it take for leaders to address these concerns while also moving ahead at light speed?

Image description begins: The exhibit shows the share of US employees with concerns regarding gen AI, through a series of proportional area charts, each representing a specific risk associated with gen AI. The size of each chart indicates the percentage of US employees who cite that risk as a concern. Cybersecurity risks are cited by 51 percent of respondents, inaccuracies by 50 percent, and concerns about personal privacy by 43 percent. Intellectual property infringement is a concern for 40 percent of respondents, followed by workforce displacement (35 percent), explainability (34 percent), and equity and fairness (30 percent). Less prominent but still significant concerns were regulatory compliance issues (28 percent), national security (24 percent), damage to organizational reputation (16 percent), environmental impact (15 percent), physical safety (14 percent), and political stability (13 percent). Source: McKinsey US employee survey, Oct–Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Employees trust business leaders to get it right

While employees acknowledge the risks and even the likelihood that AI may replace a considerable portion of their work, they place high trust in their own employers to deploy AI safely and ethically. Notably, 71 percent of employees trust their employers to act ethically as they develop AI. In fact, they trust their employers more than universities, large technology companies, and tech start-ups (Exhibit 9).

Image description begins: The exhibit depicts the share of US employees who highly trust different institutions to deploy gen AI tools responsibly, safely, and ethically. The data is presented as four separate unit charts, each representing a distinct institution: employer, universities, large tech companies, and start-ups. Each unit chart consists of a 10×10 matrix of squares. The number of light blue squares within each grid represents the percentage of employees who express high trust in each institution. The remaining squares are light gray. Employers receive the highest level of trust (71 percent), followed by universities (67 percent), large tech companies (61 percent), and start-ups (51 percent). Source: McKinsey US employee survey, Oct–Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

According to our research, this is in line with a broader trend in which employees show higher trust in their employers to do the right thing in general (73 percent) than in other institutions, including the government (45 percent). This trust should help leaders act with confidence as they tackle the speed-versus-safety dilemma. That confidence also applies outside the United States, even though employees in other regions may have more desire for regulation (see sidebar “Global perspectives on regulation”).

Sidebar Global perspectives on regulation A high percentage of international C-suite leaders we surveyed across five regions (Australia, India, New Zealand, Singapore, and the United Kingdom) are Gloomers, who favor greater regulatory oversight. Between 37 to 50 percent of international C-suite leaders self-identify as Gloomers, versus 31 percent in the United States. This may be because top-down regulation is more accepted in many countries outside the United States. Of the global C-suite leaders surveyed, half or more worry that ethical use and data privacy issues are holding back their employees from adopting gen AI. However, our research shows that attitudes about regulation are not inhibiting the economic expectations of business leaders outside the United States. More than half of the international executives (versus 41 percent of US executives) indicate they want their companies to be among the first adopters of AI, with those in India and Singapore being especially bullish (exhibit). The desire of international business leaders to be AI first movers can be explained by the revenue they expect from their AI deployments. Some 31 percent of international C-suite leaders say they expect AI to deliver a revenue uplift of more than 10 percent in the next three years, versus just 17 percent of US leaders. Indian executives are the most optimistic, with 55 percent expecting a revenue uplift of 10 percent or more over the next three years.

Risk management for gen AI

In Superagency, Hoffman argues that new risks naturally accompany new capabilities—meaning they should be managed but not necessarily eliminated. Leaders need to contend with external threats, such as infringement on intellectual property (IP), AI-enabled malware, and internal threats that arise from the AI adoption process. The first step in building fit-for-purpose risk management is to launch a comprehensive assessment to identify potential vulnerabilities in each of a company’s businesses. Leaders can then establish a robust governance structure, implement real-time monitoring and control mechanisms, and ensure continuous training and adherence to regulatory requirements.

One powerful control mechanism is respected third-party benchmarking that can increase AI safety and trust. Examples include Stanford CRFM’s Holistic Evaluation of Language Models (HELM) initiative—which offers comprehensive benchmarks to assess the fairness, accountability, transparency, and broader societal impact of a company’s AI systems—as well as MLCommons’s AILuminate tool kit on which researchers from Stanford collaborated. Other organizations such as the Data & Trust Alliance unite large companies to create cross-industry metadata standards that aim to bring more transparency to enterprise AI models.

While benchmarks have significant potential to build trust, our survey shows that only 39 percent of C-suite leaders use them to evaluate their AI systems. Furthermore, when leaders do use benchmarks, they opt to measure operational metrics (for example, scalability, reliability, robustness, and cost efficiency) and performance-related metrics (including accuracy, precision, F1 score, latency, and throughput). These benchmarking efforts tend to be less focused on ethical and compliance concerns: Only 17 percent of C-suite leaders who benchmark say it’s most important to measure fairness, bias, transparency, privacy, and regulatory issues (Exhibit 10).

Image description begins: The exhibit presents data on the utilization of benchmarks for gen AI tools among US C-suite executives. The exhibit is in two parts. The first part is a pie chart showing that 39 percent of respondents have benchmark standards for gen AI tools used by their employees. This indicates a significant minority of C-suite executives currently employ such standards. The second part of the exhibit is a horizontal bar chart displaying the benchmarks considered most important by the C-suite respondents. Performance-related benchmarks are deemed most important by 41 percent of respondents. Operational benchmarks follow closely behind, cited by 35 percent of participants. Ethical and compliance benchmarks are a lower priority, selected by 17 percent of the respondents, while other benchmarks account for only 7 percent of responses. This reveals a noteworthy disparity, suggesting C-suite leaders put a stronger emphasis on benchmarking the performance and operational aspects of AI rather than benchmarking ethical considerations. Source: McKinsey US CxO survey, Oct-Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

The focus on operational and performance metrics reflects the understandable desire to prioritize immediate technical and business outcomes. But ignoring ethical considerations can come back to haunt leaders. When employees don’t trust AI systems, they are less likely to accept them. Although benchmarks are not a panacea to eliminate all risk and can’t ensure that AI systems are fully efficient, ethical, and safe, they are a useful tool.

Even companies that excel at all three categories of AI readiness—technology, employees, and safety—are not necessarily scaling or delivering the value expected. Nevertheless, leaders can harness the power of big ambitions to transform their companies with AI. The next chapter examines how.

Chapter 4 Embracing bigger ambitions

Most organizations that have invested in AI are not getting the returns they had hoped. They are not winning the full economic potential of AI. About half of C-suite leaders at companies that have deployed AI describe their initiatives as still developing or expanding (Exhibit 11). They have had the time to move further. Our research shows that more than two-thirds of leaders launched their first gen AI use cases over a year ago.

This is a time when you should be getting benefits [from AI] and hope that your competitors are just playing around and experimenting. Erik Brynjolfsson, professor at Stanford University and director of the Digital Economy Lab at the Stanford Institute for Human-Centered Artificial Intelligence (HAI)

Image description begins: The exhibit is a horizontally stacked bar graph that shows the percentage of C-suite respondents who describe their gen AI rollouts by maturity stages. 8 percent of respondents report their organizations are in the nascent stage, characterized by minimal gen AI initiatives with no significant impact on employee workflows. A significantly larger portion, 39 percent, describe their organizations as being in the emerging stage, where gen AI pilot projects are starting to show value. The developing stage, where gen AI implementation is changing certain workflows and increasing efficiency, accounts for 31 percent of respondents. 22 percent of respondents place their organizations in the expanding stage, indicating that gen AI is scaled across departments, transforming workflows, and enhancing operations. Finally, only 1 percent of C-suite respondents describe their gen AI rollouts as mature, meaning that gen AI is fundamentally changing how work is done and driving substantial business outcomes. The exhibit highlights that the figures might not add up to 100 percent due to rounding. Source: McKinsey US CxO survey, Oct-Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Pilots fail to scale for many reasons. Common culprits are poorly designed or executed strategies, but a lack of bold ambitions can be just as crippling. This chapter looks at patterns governing today’s investments in AI across industries and suggests the potential awaiting those who can dream bigger.

Methodology To create our report, we surveyed 3,613 employees (managers and independent contributors) and 238 C-level executives in October and November 2024. Of these, 81 percent came from the United States, and the rest came from five other countries: Australia, India, New Zealand, Singapore, and the United Kingdom. The employees spanned many roles, including business development, finance, marketing, product management, sales, and technology. All the survey findings discussed in the report, aside from two sidebars presenting international nuances, pertain solely to US workplaces. The findings are organized in this way because the responses from US employees and C-suite executives provide statistically significant conclusions about the US workplace. Analyzing global findings separately allows a comparison of differences between US responses and those from other regions. Due to rounding, percentages may not always sum to 100 percent. Three-quarters of survey respondents in the United States work at organizations generating at least $100 million in annual revenue, and half work at companies with annual revenues exceeding $1 billion. All US C-suite leader respondents work at organizations with annual revenues of at least $1 billion. Looking at workforce size, 20 percent of US respondents work at companies with fewer than 10,000 employees, 49 percent work at companies with between 10,000 and 50,000, and 31 percent work at companies with more than 50,000. The analysis extended far beyond surveys. The researchers also conducted interviews with dozens of C-level executives and industry experts to understand their perspectives on AI’s transformative potential and the steps they are taking to lead their organizations through this transition. The report was further enriched by discussions with experts at Stanford HAI, the Digital Economy Lab at HAI, and McKinsey’s leading AI experts. Our survey and research primarily focus on gen AI; however, it is important to note that participants in the survey may not have consistently differentiated between gen AI and other forms of AI. Additionally, we developed a comprehensive database featuring more than 250 potential AI use cases, building on the 63 gen AI use cases identified by McKinsey’s Digital Practice. This database also incorporates proprietary McKinsey research on personal productivity as well as industry reports, along with secondary research from the US government’s Federal AI Use Case Inventories, NASA, press articles, and public interviews with technology leaders.

AI investments vary by industry

Different industries have different AI investment patterns. Within the top 25 percent of spenders, companies in healthcare, technology, media and telecom, advanced industries, and agriculture are ahead of the pack (Exhibit 12). Companies in financial services, energy and materials, consumer goods and retail, hardware engineering and construction, and travel, transport, and logistics are spending less. The consumer industry—despite boasting the second-highest potential for value realization from AI—seems least willing to invest, with only 7 percent of respondents qualifying in the top quartile, based on self-reported percentage of revenue spend on gen AI. That hesitation may be explained by the industry’s low average net margins in mass-market categories and thus higher confidence thresholds for adopting costly organization-wide technology upgrades.

Image description begins: The scatterplot exhibit depicts how companies’ gen AI spend does not match the economic potential in their industries. The exhibit illustrates that several industries with a high economic potential from gen AI are not yet spending significantly on the technology. It shows the relationship between the industry share of overall survey respondents and the industry share of top-quartile gen AI spending. On both axes, the top value is set to 35 percent. The size of each circle represents the economic potential from gen AI in billions of dollars for each industry. Light blue circles represent industries where the share in the top quartile of gen AI spending is higher than their overall survey share. These include healthcare, which has a large circle indicating significant economic potential from gen AI, and technology, also with a substantial circle suggesting large economic potential. Media and telecom and advanced Industries are also shown in light blue to illustrate strong economic potential from gen AI, but with smaller circles indicating less economic potential than healthcare and technology. Agriculture is represented by a small light blue circle. Dark grey circles represent industries where the share in the top quartile of gen AI spending is lower than their overall survey share. These include financial services, which has a large circle showing high economic potential. Energy and materials, consumer goods and retail, hardware engineering and construction, and travel, transportation, and logistics are represented by overlapping circles of varying sizes, implying a range of economic potential. No industry scores higher than 20 percent on the share of overall survey respondents. Some industries such as media and telecom, advanced Industries, and agriculture account for around 5 percent or less of overall survey respondents. Industries that scored high percentages on the share of top-quartile gen AI spending include healthcare and technology. Source: The economic potential of gen AI: The next productivity frontier, McKinsey. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

In some industries, employees are cautious

Employees in the public sector, as well as the aerospace and defense and semiconductor industries, are largely skeptical about the development of AI’s future. In the public sector and aerospace and defense, only 20 percent of employees anticipate that AI will have a significant impact on their daily tasks in the next year, versus roughly two-thirds in media and entertainment (65 percent) and telecom, at 67 percent (Exhibit 13). What’s more, our survey shows that just 31 percent of social sector employees trust that their employers will develop AI safely. That’s the least confidence in any industry; the cross-industry average is 71 percent.

Employees’ relative caution about AI in these sectors likely reflects near-term challenges posed by external constraints such as rigorous regulatory oversight, outdated IT systems, and lengthy approval processes.

There’s a lot of headroom in some functions

Our research finds that the functional areas where AI presents the greatest economic potential are also those where employee outlook is lukewarm. Employees in sales and marketing, software engineering, customer service, and R&D contribute roughly three-quarters of AI’s total economic potential, but the self-reported optimism of employees in these functions is middling (Exhibit 14). It may be the case that these functions have piloted AI projects, leading employees to be more realistic about AI’s benefits and limitations. Or perhaps the economic potential has made them worry that AI could replace their jobs. Whatever the reasons, leaders in these functions might consider investing more in employee support and elevating the change champions who can improve that sentiment.

Image description begins: The exhibit is made up of a scatter plot and a separate bar chart visualizing the relationship between the potential economic value from gen AI and the share of employees with a positive outlook on gen AI, categorized by business function. The exhibit illustrates that the functions with the employees most optimistic about gen AI are not the functions with the greatest potential economic value from gen AI. The scatter plot displays business functions: sales and marketing, software engineering, customer service, R&D, legal, risk, and compliance, operations, HR, strategy, supply chain, finance, procurement, and IT. Each function is represented by a data point, with its horizontal position indicating the percentage of employees expressing a positive outlook on gen AI, and its vertical position representing the potential economic value of gen AI in those functions, in trillions of dollars. Sales and marketing shows the highest potential economic value and around 50 percent of employees with a positive outlook. Software engineering is the function with second-highest economic potential from gen AI, with again about 50 percent of employees in that function reporting being optimistic about gen AI. Customer service and R&D also show about 50 percent of employees with a positive outlook on gen AI, but a much lower potential economic value. Several functions, such as operations, HR, Strategy, and IT, cluster together with low potential economic value and similarly middling employee optimism. Employes in IT, finance, and procurement are the most optimistic about gen AI, with about 70 percent of employees reporting positive sentiment, but these functions represent low economic potential from gen AI. The adjacent bar chart breaks down the share of the total potential economic value contributed by each function. Sales and marketing accounts for 28 percent of the total potential economic value from gen AI, followed by software engineering at 25 percent. Customer service contributes 11 percent, while R&D contributes 9 percent. The remaining 27 percent is attributed to other functions. Source: McKinsey US CxO survey, Oct-Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Gen AI has not delivered enterprise-wide ROI, but that can change

Across all industries, surveyed C-level executives report limited returns on enterprise-wide AI investments. Only 19 percent say revenues have increased more than 5 percent, with another 39 percent seeing a moderate increase of 1 to 5 percent, and 36 percent reporting no change (Exhibit 15). And only 23 percent see AI delivering any favorable change in costs.

Image description begins: In this exhibit, a pair of segmented bar charts displays US CxOs’ perceptions of whether gen AI has delivered significant return on investment for their enterprises, with details on whether they believe gen AI has impacted their revenues and costs. The data is divided into two bar graphs, one for revenues and one for costs, each presented as a stacked bar chart showing the percentage of respondents who report various levels of change. In the revenues section, 39 percent of respondents report that gen AI has delivered a revenue increase of 1–5 percent, 12 percent report an increase of 6–10 percent, and 7 percent report a revenue increase of more than 10 percent. A significant 36 percent report no change in revenue, while a small percentage (2 percent) report a decrease. An additional 3 percent are not tracking revenue related to gen AI, and 2 percent indicate they do not know. The costs section presents a similar breakdown. A substantial 31 percent of respondents report that gen AI has resulted in no change in their organizations’ costs, followed by 29 percent who report an increase of 1–10 percent. Furthermore, 17 percent report a cost decrease of 1–10 percent, while 6 percent report a decrease of 11–19 percent. A smaller percentage of 10 percent indicate a cost increase of 11–19 percent, while 4 percent report a cost increase of 20 percent or more. Similar to the revenue section, 2 percent of respondents are not tracking cost changes related to gen AI, and 3 percent indicate they do not know. The exhibit highlights that the figures might not add up to 100 percent due to rounding. Source: McKinsey US CxO survey, Oct-Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Despite this, company leaders are optimistic about the value they can capture in the coming years. A full 87 percent of executives expect revenue growth from gen AI within the next three years, and about half say it could boost revenues by more than 5 percent in that time frame (Exhibit 16). That suggests quite a lot could change for the better over the next few years.

Image description begins: In this exhibit, a segmented bar graph shows the extent to which C-suite executives perceive gen AI will affect their organizations’ revenues over the next three years, and the percentage of respondents who anticipate gen AI will result in different levels of revenue change. The bar graph shows that 36 percent of respondents anticipate that gen AI will deliver a 1-5 percent increase in revenue, 34 percent anticipate a 6-10 percent increase in revenue, and 17 percent anticipate a greater than 10 percent increase in revenue. In contrast, 10 percent of respondents anticipate that gen AI will deliver no change in revenue. A total of 51 percent of respondents anticipate that gen AI will deliver a revenue increase of over 5 percent. No respondents anticipate that gen AI will deliver a decrease in revenue, while 3 percent are not currently tracking revenue changes related to gen AI. Source: McKinsey US CxO survey, Oct-Nov 2024. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

Big ambitions can help solve big problems

To drive revenue growth and improve ROI, business leaders may need to commit to transformative AI possibilities. As the hype around AI subsides and the focus shifts to value, there is a heightened attention on practical applications that can create competitive moats.

[It] is critical to have a genuinely inspiring vision of the future [with AI] and not just a plan to fight fires. Dario Amodei, cofounder and CEO of Anthropic

To assess how far along companies are in this shift, we examined three categories of AI applications: personal use, business use, and societal use (see sidebar “AI’s potential to enhance our personal lives”). We mapped over 250 applications from our work and publicly shared examples to understand the spectrum of impact levels, from localized use cases to transformations with more universal impact. Our conclusion? Given that most companies are early in their AI journeys, most AI applications are localized use cases still in the pilot stages (Exhibit 17).

Image description begins: This text-based exhibit illustrates how gen AI use cases can be categorized by their impact levels, ranging from localized impact to universal impact. The text table presents three sections of gen AI deployments: use cases, domains, and transformations. Within each section, examples of gen AI applications are shown, with colored dots indicating whether the type of primary impact for each example is personal, business, or societal. The use cases section focuses on gen AI deployments that provide productivity boosts through automation of specific tasks or jobs. Examples include conducting smarter searches for everyday information (personal), planning events (personal), assessing candidate recruiting performance (business), accelerating contract generation (business), processing customer information faster (business), and identifying high-value consumers for tailored sales actions (business). These gen AI examples are all positioned on the more localized end of the impact spectrum. The domains section shows gen AI applications that reshape multiple roles across an area of operations. These are all classified as having a primary business impact and include developing and executing data-based campaigns, conducting synthetic customer research, conducting real-time supply chain monitoring, and accelerating coding processes. These use cases fall in the middle of the spectrum between more localized and more universal impact levels. Finally, the transformations section highlights use cases that fundamentally reshape industries, fields, and lives. These are positioned at the more universal end of the impact spectrum and include accelerating discovery and manufacturing in material science (business), predicting natural disasters and supporting crisis management strategy (societal), and accelerating drug development by reducing cost and time (societal). This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

In many cases, that’s perfectly appropriate. But creating AI applications that can revolutionize industries and create transformative value requires something more. Robotics in manufacturing, predictive AI in renewable energy, drug development in life sciences, and personalized AI tutors in education—these are the kinds of transformative efforts that can drive the greatest returns. These weren’t created from a reactive mindset. They are the result of inspirational leadership, a unique concept of the future, and a commitment to transformational impact. This is the kind of courage needed to develop AI applications that can revolutionize industries.

It is in [the] collaboration between people and algorithms that incredible scientific progress lies over the next few decades. Demis Hassabis, cofounder and CEO of Google DeepMind

To truly harness the potential of AI, companies must challenge themselves to envision and implement more breakthrough initiatives. Success in the era of AI hinges not just on technology deployment or employee willingness but also on visionary leadership. The ingredients are here. The technology is already highly capable and rapidly advancing, and employees are more ready than leaders think. Leaders have more permission space than they realize to deploy AI quickly in the workplace. To do so, leaders need to stretch their ambitions toward systematic change, laying the foundation for real competitive differentiation. If they want to be more ambitious about AI, companies must increase the proportion of transformational initiatives in their portfolios. The next chapter examines the headwinds that leaders must overcome—and how they can do so.

Sidebar AI’s potential to enhance our personal lives Outside of the business context, individuals are increasingly using AI in their personal lives. In previous research, we analyzed the potential impact of AI across 77 personal activities and across age, gender, and working status in the United States. While individuals have limited desire to automate certain personal activities, including leisure, sleeping, and fitness, the data shows significant opportunity for AI combined with other technologies to help with chores or labor-intensive tasks. Already in 2024, our research identified about an hour of such daily activities with the technical potential to be automated. By 2030, expansion of use cases and continued improvements in AI safety could increase automation potential up to three hours per day. When people use AI-enabled tools—say, an autonomous vehicle for transportation or an interactive personal finance bot—they can repurpose time for personal fulfillment activities or being productive in other ways. Using human-centric design and tapping into gen AI’s potential for “emotional intelligence” are unlocking new personal AI applications that go beyond basic efficiencies. Individuals are beginning to use conversational and reasoning AI models for counseling, coaching, and creative expression. For example, people are using conversational AI for advice and emotional support or to bring their artistic visions to life with only verbal cues. Further, to the notion that AI superagency will advance society, AI has potential to become a democratizing force, making experiences that were previously expensive or exclusive—such as animation generation, career coaching, or tax advice—available to much wider audiences.

Chapter 5 Technology is not the barrier to scale

There is no question: AI offers a rare and phenomenal opportunity. Almost 90 percent of leaders anticipate that deploying AI will drive revenue growth in the next three years. But securing that growth entails corporate transformation, and businesses have a poor track record in this area. Nearly 70 percent of transformations fail.

As we build this next generation of AI, we made a conscious design choice to put human agency both at a premium and at the center of the product. For the first time, we have the access to AI that is as empowering as it is powerful. Satya Nadella, chairman and CEO of Microsoft

To make their companies part of the minority that succeed, C-level executives must turn the mirror on themselves. They need to embrace the vital role their leadership plays. C-suite leaders participating in our survey are more than twice as likely to say employee readiness is a barrier to adoption as they are to blame their own role. But as previously noted, employees indicate that they are quite ready.

This chapter looks at how leaders can take the reins, recognizing and owning the fact that the AI opportunity requires more than technology implementation. It demands a strategic transformation. There is no denying that companies face a set of AI headwinds. To tackle these challenges, leadership teams will need to commit to rewiring their enterprises.

The operational headwinds that slow execution

Business adoption of AI faces several operational headwinds. Our interviews and research surfaced five that are most challenging: aligning leadership, addressing cost uncertainty, workforce planning, managing supply chain dependencies, and meeting the demand for explainability.

Leadership alignment is a challenging but critical first step

Securing consensus from senior leaders on a strategy-led gen AI road map is no simple task. The key to meeting this challenge is first recognizing that leadership alignment cannot be oversimplified or assumed. The process requires ongoing engagement from senior leaders across business domains, each of which may have distinct objectives and risk appetites. Together, leaders must clearly define where value lies, how AI will drive this value, and how risk will be mitigated. They must collectively establish metrics for performance evaluation and investment recalibration. To facilitate alignment, they may want to appoint a gen AI value and risk leader or institute an enterprise-wide leadership and orchestration function. These actions can enhance collaboration among business, technology, and risk teams. Although challenging, aligning leadership is a crucial step to ensure that AI projects are not disparate, avoid liability, and deliver transformative business outcomes.

Cost uncertainty makes it difficult for enterprises to predict ROI

Many companies are still determining if they can “take” AI solutions off the shelf from tech vendors or if they need to “shape” and customize them, which can be more costly but brings the potential for greater differentiation from competitors. Additionally, while leaders can budget for AI pilots, the full cost of building and managing AI applications at scale remains uncertain. Planning for a limited pilot is very different from assessing the costs of a mature solution that helps most employees multiple times a day. These factors lead to tough tradeoffs. But to move at the pace of AI, technology leaders must prioritize accelerated decision-making.

Workforce planning is more difficult than ever

There is still a world of uncertainty to manage. Employers do not know how many AI experts they will need with what type of skills, whether that talent bench even exists, how quickly they can source people, and how they can remain an attractive employer for in-demand hires after they come aboard. On the other hand, they do not know how fast AI may depress demand for other skills and thus require workforce rebalancing and retraining.

Supply chain dependencies can wreak havoc

Fragile supply chains can expose enterprises to disruptions and technical, regulatory, and legal challenges. The AI supply chain is global, with significant R&D concentrated in China, Europe, and North America and with semiconductor and hardware manufacturing concentrated in East Asia and the United States. Today’s geopolitics are complex. Furthermore, models and applications are increasingly created in open-source forums spanning many countries.

Demand for greater explainability is a central challenge

Safe AI deployment is increasingly a must-have. Yet most LLMs are often black boxes that do not reveal why or how they came to a certain response, nor what data was used to make it. If AI models cannot provide clear justifications for their responses, recommendations, decisions, or actions—showing the specific factors that led to a credit card application denial, for example—they will not be trusted for critical tasks.

These AI-specific headwinds are formidable but addressable. Companies are pushing ahead. For example, they might use dynamic cost planning or look at procuring NVIDIA clusters to secure the infrastructure they expect to need. Chief HR officers (CHROs) are developing training programs to upskill their current workforces and support some employees in job transitions. But lasting success will take more than that.

To capture AI value, leaders must rewire their companies

McKinsey’s Rewired framework includes six foundational elements to guide sustained digital transformation: road map, talent, operating model, technology, data, and scaling (Exhibit 18). When companies implement this playbook successfully, they cultivate a culture of autonomy, leverage modern cloud practices, and assemble multidisciplinary agile teams.

Image description begins: The text-based exhibit is a framework outlining six key success factors for tech-business transformations. It’s structured as a grid with a title, “Framework for the coordinated execution of value creation,” and is further divided into six rectangular sections, each representing a different success factor. An overarching section, business-led digital road map, focuses on the importance of an organization aligning senior leadership to create a clear vision, value proposition, and roadmap for transformation, thereby improving the customer experience and competitiveness. The next four sections are categories representing the core operational aspects of a business-tech transformation. Talent emphasizes the importance of having employees with the right skills and capabilities to execute and innovate. Operating model highlights the need for organizations to relentlessly focus on value creation by integrating business, technology, and operations. Technology stresses the importance of using technology effectively through adoption of the right platforms, solutions, and practices to drive innovation. Finally, data and AI focuses on why it’s essential to provide people in the organization with easy access to high-quality data and to leverage AI insights to enhance customer experiences and business operations. The final section, underpinning the four category sections, is activation and scaling. This section underscores the need for organizations to maximize value capture by ensuring the activation and enterprise scaling of digital solutions, while carefully managing the transformation’s progress and mitigating risks. This image description was completed with the assistance of Writer, a gen AI tool. Image description ends.

While these six elements are universally applicable, AI has introduced a few important wrinkles for leaders to address:

Adaptability. AI technology is advancing so rapidly that organizations must adopt new best practices quickly to stay ahead of the competition. Best practices may come in the form of new technologies, talent, business models, or products. For example, a modular approach helps future-proof tech stacks. As natural language becomes a medium for integration, AI systems are becoming more compatible, allowing businesses to swap, upgrade, and integrate models and tools with less friction. This modularity allows enterprises to avoid vendor lock-in and put new AI advancements to use quickly without constantly reinventing their tech stacks.

Federated governance models. Managing data and models can give teams autonomy to develop new AI tools while centrally controlling risk. Leaders can directly oversee high-risk or high-visibility issues, such as setting policies and processes to monitor models and outputs for fairness, safety, and explainability. But they can set direction and delegate other monitoring to business units, including measuring performance-based criteria such as accuracy, speed, and scalability.

Budget agility. Given technological advances across models, as well as the opportunity to curate an optimal mix of LLMs, small language models (SLMs), and agents, business leaders should keep their budgets flexible. This helps enterprises optimize their AI deployments simultaneously for costs and performance.

AI benchmarks. These tools can serve as powerful means to quantitatively assess, compare, and improve the performance of different AI models, algorithms, and systems. If technologists come together to adopt standardized public benchmarks—and if more C-level executives start employing benchmarks, including ethical ones—model transparency and accountability will improve and AI adoption will increase, even among more skeptical employees.

AI-specific skill gaps. Notably, 46 percent of leaders identify skill gaps in their workforces as a significant barrier to AI adoption. Leaders will need to attract and hire top-level talent, including AI/ML engineers, data scientists, and AI integration specialists. They will also need to commit to creating an environment that is attractive to technologists. For example, this can mean providing them with plenty of time to experiment, offering access to cutting-edge tools, creating opportunities to engage in open-source communities, and promoting a collaborative engineering culture. Upskilling existing employees is just as critical: Research from McKinsey’s People and Organizational Performance Practice underscores the importance of tailoring training to specific roles, such as offering technical team members bootcamps on library creation while offering prompt engineering classes to specific functional teams.

Human centricity. To guarantee both fairness and impartiality, it is important that business leaders incorporate diverse perspectives early and often in the AI development process and maintain transparent communication with their teams. As it stands, less than half of C-suite leaders (48 percent) say they would involve nontechnical employees in the early development stages of AI tools, specifically ideation and requirement gathering. Agile pods and human-centric development practices such as design-thinking and reinforcement learning from human feedback (RLHF) will help leaders and developers create AI solutions that all people want to use. In agile pods, technical team members sit alongside employees from business functions such as HR, sales, and product, and from support functions such as legal and compliance. Further, leaders can empathize with employees’ uneasiness about AI’s impacts on potential job losses by being honest about new skill requirements and head count changes. Forums where employees can provide input on AI applications, voice concerns, and share ideas are valuable for maintaining a transparent, human-first culture.

Conclusion Meeting the AI future

The pace at which AI has advanced over the last two years is stunning. Some react to that pace by seeing AI as a challenge to humanity. But what if we take the advice of Reid Hoffman and imagine what could possibly go right with AI? Leaders might realize that all the pieces are in place for AI superagency in the workplace.

Learn from yesterday, live for today, hope for tomorrow. Albert Einstein, theoretical physicist

They might notice that their employees are already using AI and want to use it even more. They may find that millennial managers are powerful change champions ready to encourage their peers. Instead of focusing on the 92 million jobs expected to be displaced by 2030, leaders could plan for the projected 170 million new ones and the new skills those will require.

This is the moment for leaders to set bold AI commitments and to meet employee needs with on-the-job training and human-centric development. As leaders and employees work together to reimagine their businesses from the bottom up, AI can evolve from a productivity enhancer into a transformative superpower—an effective partner that increases human agency. Leaders who can replace fear of uncertainty with imagination of possibility will discover new applications for AI, not only as a tool to optimize existing workflows but also as a catalyst to solve bigger business and human challenges. Early stages of AI experimentation focused on proving technical feasibility through narrow use cases, such as automating routine tasks. Now the horizon has shifted: AI is poised to unlock unprecedented innovation and drive systemic change that delivers real value.

Glossary The following terms in this report are defined specifically for its context. Adoption and deployment: Deployment typically refers to the extent to which an organization rolls out a technology product (whether developed in-house or purchased off the shelf), and adoption reflects how extensively these products are used to generate measurable business value. Given the emerging nature of AI, many companies are simultaneously deploying and adopting, iterating as they go. Therefore, this report often uses adoption and deployment interchangeably to refer to the overall uptake of AI tools. Agentic AI: Systems with autonomy and goal-directed behavior capable of making independent decisions, planning, and adapting to achieve specific objectives without direct, ongoing human input. Application programming interface (API): Intermediary software components that allow two applications to talk to each other; a structured way for AI systems to programmatically access (usually external) models, data sets, or other pieces of software. Artificial Intelligence (AI): The ability of software to perform tasks that traditionally require human intelligence, mirroring some cognitive functions usually associated with human minds. Deep learning: A subset of machine learning that uses deep neural networks, which are layers of connected “neurons” whose connections have parameters or weights that can be trained. Deep learning is especially effective at learning from unstructured data such as images, text, and audio. Digital workforce: A collaborative ecosystem where humans and automated agents work together, leveraging digital platforms, AI, and cloud computing to enhance productivity, efficiency, and scalability across various industries. Employee: A worker in a corporate setting, either a manager or independent contributor. Examples of the type of employees represented in this report include people working in product management, marketing, technology, business development, sales, and finance. Foundation models: Deep learning models trained on vast quantities of unstructured, unlabeled data that can be used for a wide range of tasks out of the box or adapted to specific tasks through fine-tuning. Examples of these models are DALL-E 2, GPT-4, PaLM, and Stable Diffusion. Generative AI (gen AI): AI that is typically built using foundation models and has capabilities that earlier forms Iacked, such as the ability to generate content. Foundation models can also be used for nongenerative purposes (for example, classifying user sentiment as negative or positive based on call transcripts). Graphics processing units (GPUs): Computer chips originally developed for producing computer graphics, such as for video games, that are also useful for deep learning applications. In contrast, traditional machine learning usually runs on central processing units (CPUs), normally referred to as a computer’s “processor.” Hallucination: A scenario where an AI system generates outputs that lack grounding in reality or a provided context. For instance, an AI chatbot may fabricate information or present a false narrative. Large language models (LLMs): A class of foundation models that can process massive amounts of unstructured text and learn the relationships between words or portions of words, known as tokens. This enables LLMs to generate natural-language text, performing tasks such as summarization or knowledge extraction. GPT-4 (which underlies ChatGPT) and the Llama family of models from Meta are examples of LLMs. Modality: A high-level data category such as numbers, text, images, video, and audio. Multimodal capabilities: The ability of an AI system to process and generate various types of data (text, images, audio, video) simultaneously, enabling complex tasks and rich outputs. Productivity (from labor): The ratio of GDP to total hours worked in the economy. Labor productivity growth generally comes from increases in the amount of capital available to each worker, the education and experience of the workforce, and improvements in technology. Prompt engineering: The process of designing, refining, and optimizing input prompts to guide a gen AI model toward producing desired and accurate outputs. Reasoning AI: AI systems that perform logical thinking, step-by-step planning, problem solving, and decision making using structured or unstructured data, going beyond pattern recognition to draw conclusions and solve complex problems. Superagency: A state where individuals, empowered by AI, amplify their creativity, productivity, and positive impact. Even those not directly engaging with AI can benefit from its broader effects on knowledge, efficiency, and innovation. Unstructured data: Data that lack a consistent format or structure (for example, text, images, video, and audio files) and typically require more advanced techniques to extract insights.

To meet this more ambitious era, leaders and employees must ask themselves big questions. How should leaders define their strategic priorities and steer their companies effectively amid disruption? How can employees ensure they are ready for the AI transition coming to their workplaces? Questions like the following ones will shape a company’s AI future:

For business leaders:

Is your strategy ambitious enough? Do you want to transform your whole business? How can you reimagine traditional cost centers as value-driven functions? How do you gain a competitive advantage by investing in AI?

Do you want to transform your whole business? How can you reimagine traditional cost centers as value-driven functions? How do you gain a competitive advantage by investing in AI? What does successful AI adoption look like for your organization? What success indicators will you use to evaluate whether your investments are yielding desired ROI?

What skills define an AI-native workforce? How can you create opportunities for employees to develop these skills on the job?

For employees:

What does achieving AI mastery mean for you? Does it extend to confidently using AI for personal productivity tasks such as research, planning, and brainstorming?

How do you plan to expand your understanding of AI? Which news sources, podcasts, and video channels can you follow to remain informed about the rapid evolution of AI?

How can you rethink your own work? Some of the most innovative ideas often emerge from within teams, rather than being handed down from leadership. How would you redesign your work to drive bottom-up innovation?

These questions have no easy answers, but a consensus is emerging on how to best address them. For example, some companies deploy both bottom-up and top-down approaches to drive AI adoption. Bottom-up actions help employees experiment with AI tools through initiatives such as hackathons and learning sessions. Top-down techniques bring executives together to radically rethink how AI could improve major processes such as fraud management, customer experience, and product testing.

These kinds of actions are critical as companies seek to move from AI pilots to AI maturity. Today only 1 percent of business leaders report that their companies have reached maturity. Over the next three years, as investments in the technology grow, leaders must drive that percentage way up. They should make the most of their employees’ readiness to increase the pace of AI implementation while ensuring trust, safety, and transparency. The goal is simple: capture the enormous potential of gen AI to drive innovation and create real business value.

Retailers share images of prolific shoplifters

Major retailers are sharing the images and details of prolific shoplifters.

M&S, Morrisons, Boots, Primark and Greggs are among the stores using a new database.

The system, known as Auror, was created by a New Zealand software company.

Auror told Sky News it provides “crime reporting software” to retailers which helps them to “collaborate with law enforcement”.

It also helps them to connect what they might have thought were individual events with repeat offenders.

Auror said the top 10% of offenders are responsible for more than 65% of total loss and harm in UK stores.

It added that there is a “comparatively small group of prolific offenders linked to organised crime networks that cause a disproportionate amount of harm”.

It said one in seven “events” involved “threats, aggression, verbal abuse and other serious behaviours in 2024”.

One in 10 events involved “violence, weapons, assault, arson, hate crimes, harassment and aggression”.

It added that repeat offenders are up to four times more likely to be violent.

Hundreds of places are getting an increased police presence until the end of September.

More than 500 towns across England and Wales have signed up to the Home Office’s safer streets summer initiative.

Yvette Cooper, the home secretary, said residents and businesses have a “right to feel safe in their towns”.

She said she had asked police forces and councils to “work together to deliver a summer blitz on town centre crime”.

Cooper said she wanted to send a “clear message to those people who bring misery to our towns that their crimes will no longer go unpunished”.

With 77% of Americans feeling stressed about the future of our nation, the personal-finance website WalletHub today released its report on 2025‘s Most & Least Stressed Cities in America, as well as expert commentary.

In order to determine where Americans cope best with their stress, WalletHub compared more than 180 cities across 39 key metrics. The data set ranges from average weekly work hours to the unemployment rate to divorce and suicide rates.

Stress Levels in Baltimore (1=Most Stressed; 91=Avg.):

Overall Rank: 3 rd

89 th – Avg. Weekly Hours Worked

– Avg. Weekly Hours Worked 12 th – Traffic Congestion

– Traffic Congestion 32 nd – Poverty Rate

– Poverty Rate 9 th – Divorce Rate

– Divorce Rate 68 th – Job Security

– Job Security 32 nd – % of Adults with Inadequate Sleep

– % of Adults with Inadequate Sleep 83rd – Unemployment Rate

For the full report, please visit:

https://wallethub.com/edu/ most-least-stressed-cities/ 22759

Key takeaways and WalletHub commentary are included below in text and video format.

“Some stress is out of our control, due to issues with family, friends or employers. However, where you live can play a big role in how stressed you are. Cities with high crime rates, weak economies, less effective public health and congested transportation systems naturally lead to elevated stress levels for residents. When moving, it’s important to consider how a certain city may impact your mental health – not just your financial opportunities.”

–

“Detroit is the most stressed city, due in part to the fact that it has the lowest median household income in the country and the highest unemployment rate, at 11.4%. In addition, Detroit has the highest poverty rate in the country, residents are physically active at low rates, too, and the city has the sixth-highest obesity rate. On top of that, Detroit has one of the highest violent crime rates in the country.”

– Chip Lupo, WalletHub Analyst

Expert Commentary

How can employers reduce work-related stress?

“I think flexibility in the workplace is one important aspect for reducing work-related stress. Personal issues often require flexibility in remote options or time off. As long as the work gets done at some other point, the employers could allow more flexibility to help employees. In addition, not placing too much burden on a single employee would also be good, as well as providing the appropriate resources for employees to succeed. For example, if they need to learn how to use a new piece of software, there should be appropriate training and transition time. For good employees, micromanaging also creates unnecessary work-related stress. Finally, an appropriate pay structure can alleviate some stress if it aligns with the company’s overall mission, regular increases/bonuses are attainable by employees, and is distributed fairly.”

Rachel Wu, Ph.D. – Associate Professor, University of California – Riverside

“The two major drivers of employee work-related stress are demand – control imbalances, and effort – reward imbalances. In demand-control imbalances, an employee’s level of control (decision making, independence and autonomy) does not match the demands of their job. This occurs when an employee is not allowed to make decisions about how to do their job or have no control over the pace and demands placed upon them to do their job. It can also occur when a supervisor micromanages an employee – imposing themselves into that person’s work processes and responsibilities. In effort-reward imbalances, an employee does not receive sufficient reward for the amount of effort they put into their work. An employee who goes the extra mile, works overtime, or makes sacrifices to get things done should receive some reward for their labor. Employees who do quality work and put forth demonstrable effort in their work should receive appropriate and commensurate rewards. Rewards can be in the form of increased pay, bonuses, perks, recognition, awards, promotions, etc. When effort is not commensurate with reward, employees become significantly stressed. When a person has both demand-control, and effort-reward imbalances then that person is fertile for high levels of stress, along with increased risk for cardiovascular, musculature, and psychological ill-health, including some forms of cancer. Employers who structure jobs in a way that allow employees to have control over their work and provide mechanisms for rewards for effort given can greatly reduce the amount of stress workers experience.”

Michael Peterson, Ed.D. – Professor; Director, Social Marketing and Health Communication Lab; HBS Internship Director, University of Delaware

What tips do you have for a person who finds managing finances to be stressful?

“If the source of the stress is that your income is not keeping pace with your expenses, then make a budget. See what you can cut from your expenses. If this seems too overwhelming, get help doing this. There are many free tools on the internet to help you tackle your finances. If the problem is bigger than financial tools to manage, there are also free financial counseling in most areas. The most important issue is to make a plan that will move you into a better financial situation and stick to it.”

Joanne H. Gavin, Ph.D. – Professor, Marist College

“If managing finances seems stressful, reframe the process to make it less threatening. One way to do that is to tell yourself it’s going to be an adventure that will help you get what you want in life. Think first about what matters to you so you have a visual of what you want your life and finances to look like. You can even make a Pinterest board so you can visualize the life you want. Often, we want less chaos and more control in our lives, so follow pages that inspire you like ‘Becoming Minimalist’ on social media. Visualizing the life you want gives you the motivation to look at your finances and come up with a plan to get you to the life you are envisioning.”

Suzie Duff, PhD, LMHC – Professor; Department Co-Chair, Human Services, Palm Beach State College

What tips do you have for a person who wants to relax on a budget?

“The beauty of relaxing is that most options for relaxation operate on a spectrum of FREE (for instance, the library might have a lending program for some hobbies, like a seed bank for those who want to start gardening, or free services, such as a free DVD/movie services); to AFFORDABLE (perhaps you can buy used or second-hand materials or take a staycation instead); to quite COSTLY (the ‘dream vacation’ or ‘Cadillac’ of experiences). There will be times when you will splurge but, other times, you can be very cost efficient so you can be judicious about how much you want to invest each time. To make your choices, balance and budget. When doing so, keep in mind that a lot of times we emphasize saving money to do recreational activities but also do not forget about earning extra income. A trip to the craft store, for instance, is a great reward after earning the needed amount by filling out paid surveys online. Or, maybe your hobby might even be something that could earn money, such as furniture restoration. I also want to encourage researching free events in your local area (for instance, Facebook has an Events page) because there are a lot of great organizations hosting community events where you can make social connections (e.g., festivals, book clubs, talks, volunteer days). Healthy social connections can improve well-being and are valuable for stress-coping.”

Alisia (Giac-Thao) Tran – Associate Professor, Arizona State University

“Sit down and make a list of what makes you feel relaxed and at peace. Maybe it’s being in nature, or being on the water, or reading, or spending time with friends and family. All it takes now is creativity. How can I do these things that bring me peace for little money? If you love being in nature, maybe check out local state or regional parks close by. If you live near a beach – that’s a great option because usually they are free! If you love reading, create a book exchange with friends or go to the library. If you love having company over but feel like entertaining is expensive, have a Sunday afternoon ice cream get-together instead of a fancy dinner party. Being smart with money involves being intentional. For example, I live in Florida and love to paddle board. Investing in a board is expensive at first but provides endless free hours on the water. That’s a great way to relax on a budget.”

Suzie Duff, PhD, LMHC – Professor; Department Co-Chair, Human Services, Palm Beach State College

Q: Among the reasons given for tariffs is the need to protect U.S. jobs and reshore or inshore manufacturing. Is there evidence that this will happen?

The central claim is that America can be revitalized and indeed the American middle class can be revitalized; workers without college education can be helped, and left-behind places can be restored if we stimulate manufacturing production in the United States. But at the moment it’s important to realize and recognize that only just over 8% of Americans work in manufacturing and even if we were to entirely close the trade deficit that we have in manufactured goods, it’s likely that manufacturing employment would increase by somewhere between one or two percentage points. So instead of around 8% of Americans working in manufacturing, 10% of Americans would work in manufacturing.

Manufacturing is simply too small to have a significant impact on the American labor force and both Presidents Biden and Trump have been obsessed with restoring American manufacturing. In my own view, the claims that they make that somehow this is going to have a significant impact on the availability of jobs for the middle class is completely unrealistic.

There are some manufactured products that are very important. We need semiconductors—they’re important for artificial intelligence, they’re important for national security. We need to decarbonize, and so electric vehicles and solar panels are important. So there are certain kinds of products which can help us meet national goals. But it is unrealistic to see manufacturing as a policy that is going to have a significant impact on the major problems of less educated Americans.

Both Biden and Trump are in a sense appealing to nostalgia for a world that no longer exists, in which manufacturing is a major driver of access to the middle class. They’re about 30 to 40 years out of date because, as a result of both automation and the way we spend our money today, manufacturing has shrunk and is a relatively small part of our economy.

Q: Will tariffs help raise revenues, as the administration has claimed?

It’s problematic, because the higher the tariffs that you impose, at some point the less revenue you’re actually going to receive. The kind of estimates we’re seeing from the administration are that they will raise $600 billion. I think that’s an extremely optimistic view because as you make products more expensive, consumers will pay less or will be prepared to spend less on those imported products. In addition, one of the purposes of the tariffs is to get foreigners to come and invest in the United States. Well, if they do, they’ll no longer be paying the tariff. So ironically, the long run achievement of goals like bringing a lot of investment into the United States to replace the imports is going to undermine the goal of raising revenue, and that’s why it’s very difficult to know exactly how much is going to be raised.

But it’s important to point out that people, as they get richer, spend less and less on goods and more on services, and that means that tariffs have a regressive incidence because they take much more out of the pockets of poor Americans than they do of rich Americans. So to the degree that we now raise revenue using tariffs and use the money we save or the money we raise to reduce the taxes patented after the previous Trump tax cuts, this is an extremely regressive move for American households and the estimates are that the typical household is going to spend an additional $2,000 to $4,000, depending on which economist you believe.

There’s also an exaggeration of the employment impact that you’re going to get from tariffs. Let’s take the example of a tariff on steel. You might create more jobs in the steel industry, but you will also raise input costs for the users of steel, and this in turn affects somewhere between 60 and 80 jobs for every one you save in the steel industry itself. So in the aggregate, the tariffs can be counterproductive, especially if they’re put on inputs which are used in producing other products.

Q: Is the United States’ large trade deficit sustainable?

I think firstly there’s an obsession with goods that isn’t the right measure. What we ought to be looking at is not only our trade in goods, but also our trade in services, and we have a significant surplus in our trade in services. Therefore, when you aggregate the two together, you get a much smaller percentage and a smaller number relative to our GDP.

The second point is that we’ve been running deficits for 30 or 40 years, and what it means is that the United States is borrowing much more from the rest of the world than we lend, and therefore our net position has been declining over time. But remarkably, Americans earn more from, or earn just about as much from, their total investments abroad as foreigners earn in the United States. So if you look historically, we have felt no additional pressure about sustainability of our position. As long as we borrow the money and use it productively to increase investment in the United States, it is eminently sustainable, as with any investment.

Q: How would U.S. exports be impacted?

One of the effects of the tariffs is going to be over the medium term to strengthen the American dollar because Americans will need less foreign exchange in order to import, and when the dollar gets stronger, this affects all American exporters, whether they are exporting goods or whether they are exporting services.

A second point is that foreigners are not going to take these tariffs lying down. They are going to retaliate. Much of their retaliation can take the form of higher tariffs on American exports of goods, but in addition, foreigners are talking about levying taxes on the sales of American services and indeed some of the information technology company services that are being sold abroad. So there are going to be an adverse impact on exporters virtually any way you look—there are going to be higher input costs, they are going to have to sell into markets which are closing to them because of foreign retaliation, and the currency is going to get stronger and so their products are going to get more expensive.