The most valuable sentence a technical partner can say is often "you do not need AI for this." It rarely gets said, because saying it means turning down work. We think it is exactly the sentence worth paying for --- so it is worth being precise about when it applies.
The demo is not the problem
A convincing demo proves a model can produce a plausible output once, under conditions you chose. It says little about whether the same approach holds up against real inputs, real edge cases, and a business that needs the answer to be right most of the time. Mistaking one for the other is the most expensive error we see --- and it usually starts with choosing AI because it is impressive, not because it is the best tool for the job.
Signs you may be reaching for the wrong tool
AI is often a poor fit when one of these holds:
- The problem is deterministic. If a rule, a query or a formula already produces the correct answer, a probabilistic model tends to add cost and uncertainty without buying much back.
- You cannot tolerate wrong answers, and cannot check them. When a mistake is expensive and there is no cheap way to verify the output, you are usually buying risk rather than leverage.
- You do not have the data or the feedback loop. A model with no useful data to ground it, and no feedback loop to flag drift, is often a liability dressed as an asset.
What to do instead
Sometimes the honest recommendation is a dull one: a well-placed rule, a cleaner data pipeline, a better interface, or fixing the process that creates the problem in the first place.
We have watched a company reach for an LLM to sort incoming email into a dozen fixed categories --- a task a handful of keyword rules and a small classifier had handled for years, at a fraction of the latency and cost, and with no prompt to babysit. The LLM was more impressive in the demo and worse in production: slower, pricier, and occasionally creative in ways a fixed taxonomy cannot allow.
When AI genuinely is the right tool --- and often it is --- the business case survives this kind of scrutiny easily. Good engineering starts by choosing the simplest thing that solves the problem, and choosing it because it pays, not because it is new. That is the whole point of a discovery workshop: to find where AI earns its place, and to say so plainly where it does not.