The point is that trying to address business challenges through innovation focused on a single technology will necessarily lead to only narrow, lower value use cases being addressed. LLMs for example are really suited to efficient information creation and Q&A processes. These can be high value in themselves, but are limited in scope. However, combining them with other technologies can open up the potential to create complete solutions to a whole range of issues. For example – using LLMs to create advertising copy which is optimized for customer segments using ML based test and learn approaches can scale advertising personalization in a way that LLMs on their own would not.

Second, as a result of this complexity, achieving the end goal requires a different set of skills to both build and run the systems. To successfully develop complex integrated solutions requires bringing together deep experts in each of the technologies required. Accessing such talent is hard, since ideally these experts will have gone through several programs that implemented similar solutions at scale. Then there is the orchestration challenge. For example, if you experience an issue, which technology is at fault? Is the designer of one element of the solution cutting corners that are causing issues in the other elements? For this reason, it’s imperative to look across the system to make the right architectural and design decisions, not just do so in each individual area. Integration is also an issue. The solution will inevitably have to interact with existing systems which will need to be adapted and tested with the broader system. And in many cases, the solution will require a foundation of robust, well-governed data, without which even the best designed technical solutions will lack the “intelligence” required to drive business value.

As if designing an integrated system is not challenging enough, running and maintaining it is harder. The business environment is fluid, so how are you going to push out updates as new product or market conditions arise? The fast-moving technology landscape also brings challenges. New algorithms emerge to replace today’s innovations, and hardware evolves as processing speeds continue to increase. Simply managing a live AI environment is extremely challenging, requiring AI data “cold starts”, distributed model management, network enablement… and what happens when a sensor or camera goes down, who will fix it? These are all real technical considerations that if you haven’t thought through in advance, will cause you to come unstuck later.

On top of this, there are the “people” considerations. If technology is automating routine tasks, what skills are your staff going to need now? They will need to focus on the more complex aspects of the business operations, and also understand how the technology interplays with those operations. Their jobs will shift towards directing the automation as much as executing the business processes at hand. Getting your team to accept and adapt to this evolution comes with real challenges.

Finally, when you are fully automating mission critical processes, there is no room for error, and eliminating the errors can require at least as much if not more effort than getting to a 95% correct solution. Recent news pages have been littered with tales of rogue AI. From the airline whose chatbot hallucinated a discount fare (*2), to the training company that was discriminating against older job applicants (*3), LLMs are improving and becoming more transparent, but you need to be certain your company doesn’t fall on the wrong side of the “edge case” gap or allow unidentified bugs in software to damage the brand.

In most innovation processes the “move fast and break things” mentality still prevails. For the next generation of business transformation that will be less relevant. Breaking things is OK if you’re designing the next consumer app. But it doesn’t fit so well if you are originating loans, diagnosing diseases or automating a physical activity. When technology transformation encroaches on the mission critical operations of the business there’s no room for failure. It is imperative that edge cases are handled well and continuously monitored. That requires a higher bar for accuracy, the ability to recognize if an AI application starts to drift, and a significant investment in “failover” processes that take over when the automation fails.

These types of engineering challenges used to be the preserve of highly automated manufacturing environments and defense laboratories. If companies want to transform in the “Era of AI”, they will need to embrace these realities too. Why? Because if they don’t, then their competitors will and there is a real risk that the first companies to achieve their automation goals will disrupt the economics of the industry. The pioneers may even become the new industry platform in the way that Google’s ad platform became the standard for the advertising industry or Amazon’s scalable IT architecture became the de-facto IT industry infrastructure. Getting left behind can have grave implications.