Over the last six to nine months I've been pinged by a lot of friends and professional contacts asking for tips on implementing AI. Part of building my startup, NettWorth, in the open has been sharing my frustrations with the bottlenecks and roadblocks of building on and with AI. Some of what I wrote struck a chord, some of it hit a nerve, and the private conversations that followed turned into several small advisory gigs at companies that had got stuck somewhere in the middle of an AI initiative.
I've decided to start writing about them in the open, because the same patterns keep showing up. I'm hoping these posts generate conversations and ideally some disagreement, partly because the insights go back to the NettWorth team, and partly because if you're stuck in one of these patterns it helps to know you're not the only one.
A few months ago I was asked to respond to an industrial company's RFP. They wanted to use generative AI to solve a genuine production problem: how their equipment stretches and prints on packaging material under changing conditions.
The RFP was surprisingly detailed for an industrial company of their size. They had already decided the cloud platform (Microsoft Azure), the model family, the architecture, and even the shape of the "AI agent". From their perspective the hard work was done and they simply needed a supplier to wire it all together. Where it fell apart, in my view, was that they were asking a probabilistic system to solve a deterministic control problem.
They clearly understood the variables: material thickness, density, temperature, machine speed, roller tension, humidity, historical defect rates. There was obviously no shortage of data, since they build those machines themselves. The problem, put simply, was finding the mathematical relationship between inputs and outputs and controlling the process accordingly. That is exactly what deterministic models, optimisation algorithms and even classical machine learning have been doing for decades, not to mention that other hyped up buzzword of yesteryears, digital twins.
Large language models excel when the rules are fuzzy. Writing research, reasoning over incomplete information, summarising, some planning, extracting structure out of messy documents, that kind of stuff. They are remarkably good at problems where there isn't only one correct answer.
Industrial process control is almost (always?) the opposite. Every percentage point of prediction error has a real world cost. If the machine is going to adjust tension automatically, "usually right" isn't good enough. That doesn't rule an LLM out entirely. It can explain why the optimisation model made a recommendation, analyse maintenance reports, and help operators diagnose an unusual situation, sitting between the engineers and the underlying optimisation engine as the interface. It just shouldn't be the optimisation engine.
What struck me about the RFP wasn't that the engineers had misunderstood AI. They were clearly a competent team, and on the technology itself they were better informed than most. They had simply fallen into a trap I now see surprisingly often. Once AI becomes the strategic priority, every problem starts looking like an AI problem.
I don't think this is an engineering failure. Models have been good enough for a while now that I can say with some confidence AI projects don't fail because the model isn't smart enough. They fail earlier and further up. The board's "we have to do something in AI" becomes the CEO's "give me just enough for an AI narrative", which arrives on the CTO's desk as a mandate to deploy AI against problems it was never meant for. By the time it reaches the people who have to build the thing, the architecture has already been decided by a sentence in a strategy deck.
Sometimes the best AI architecture is one where the LLM is 10% of the solution and the rest is software engineering, data engineering, optimisation, integration and good old fashioned control theory. There are plenty of other operational areas where companies are piloting AI right now that will undoubtedly fail for the same reason, almost all of them the mathematically heavy efficiency use cases.
That's the first pattern I've started noticing.