The credibility of saying no
We build AI systems for a living. So when we tell a prospect that AI is the wrong tool for their problem, it costs us a sale and it is worth more than the sale. A shop that says yes to every AI request is not a partner, it is an order-taker, and you will pay for the difference when the over-engineered thing ships.
Here are five patterns where we have looked at the problem and recommended against putting a model in the loop.
1. The job is deterministic
If the task has a correct answer every time and clear rules to get there, you do not want a language model. You want a script. Routing a form, validating a field, moving a file based on its type - these are solved problems with deterministic tools that are cheaper, faster, and debuggable. Putting an LLM in the loop adds cost, latency, and a non-zero chance of a creative wrong answer to a problem that had none.
2. A five-line script would do it
Adjacent to the above. Sometimes the "AI automation" a team is excited about is genuinely a few lines of glue code between two APIs. The AI framing is exciting; the right build is boring and takes an afternoon. We will tell you it is an afternoon.
3. There is no data worth retrieving
A lot of "AI chatbot for our docs" ideas die on contact with the docs. If your knowledge base is thin, out of date, or does not actually contain the answers people ask for, a retrieval system will confidently surface nothing useful. Fix the corpus first. The AI is downstream of having something worth retrieving.
4. The cost of a wrong answer is too high and unguarded
AI systems can be made safe for high-stakes work, with guardrails in code, evals, and human review. But if a team wants to drop a model into a workflow where a wrong answer is expensive and they are not willing to build the guardrails, the right answer is not to ship it. An unguarded model in a high-stakes loop is a liability wearing an innovation costume.
5. The team has not defined the workflow
The most common one. A company wants "AI" before they have defined the actual process it is supposed to improve. No tool and no model fixes an undefined workflow; it just automates the confusion faster. The honest recommendation is to define the workflow first, often with a cheap off-the-shelf tool to learn, then revisit a build once you know what good looks like.
The point
None of this means AI is overhyped. We have shipped ten production products, several with real agents and retrieval doing real work. It means AI is a tool, and a senior shop knows the cases where the tool makes the system worse.
If you want a read on whether your idea is one of these five patterns or a genuine fit, book a 30-minute call. We will give you the honest answer, including "do not build this," if that is the answer.