TL;DR

  • Name who will use it, and how they notice a wrong answer, before you pick a model.
  • Write a use case you can test: the input, the output, and who checks it.
  • If better search or a simple rule already does the job, stop before you build.

Almost every time someone wants to build AI, the conversation starts with which model to pick. I usually try to move that conversation somewhere else: who will actually use this, and how will they notice if it goes wrong?

"We want AI to help support" sounds fine, but you cannot test it. "Find the right instruction and suggest an answer for the handler to approve" is something you can try in a week.

Gartner wrote on 26 January 2026 that at least 50% of generative AI projects were abandoned after the proof of concept. The top cause was poor use-case selection combined with lack of business value (Why 50% of GenAI Projects Fail).

And honestly, more projects stall because nobody can check the answer, or because the data cannot be used, than because of the model choice. Sometimes better search, or a simple rule, is the whole solution.

RAND's 2024 report, based on interviews with 65 experienced AI engineers and researchers, found the most common root cause was misunderstanding or miscommunicating the problem the project should solve. Leaders demanded machine learning where a few simple if-then rules would have worked. The recommendation is to focus on the problem, not the technology (The Root Causes of Failure for AI Projects).

When the work just feels slow

It is easy to mix up work that feels slow with work that is worth building something new for. The wait may come from nobody owning the question, from missing information, or from an approval queue. A language model does not fix that.

McKinsey's survey, The state of AI in 2026, finds that about 6% of respondents are high performers. Nearly three in four of them report having fundamentally redesigned workflows, against one in four among the rest. They reported it themselves, so the gap is a correlation. I would map how the work is done today first.

What goes in, and what should come out

When I look at a workflow, I want to know what is there when the work starts, and what should come back: a document, a draft, a classification, or a change in another system.

The data, and who may use it

Is someone responsible for the information, and is it current? May the person using the tool read it? Fetching a text does not mean it may be used. Sort that out before a larger model.

An example

Take a 40-person consultancy that answers questions about its own product. A handler gets a question about how to change a setting. The person asking names the product, but not which version it is. The instructions differ between versions.

The danger is not that the draft is badly written. The danger is that it sounds right but applies to the wrong version. An answer that sounds confident, pointing to the wrong instruction, is worse than asking for the right version first.

So the assistant should ask, not guess. A draft may only rely on an instruction that applies to the right version. If the version is missing, or two sources disagree, it should ask or hand the case to the handler. No setting is changed by itself. The handler reviews the draft and sends the reply.

Sometimes it is enough to make the right passage easier to find. If the handler can quickly open the right instruction, you may not need a generated draft at all.

A simple way to write those limits down can look like this:

Version missing: ask, do not guess
Two instructions disagree: hand to the handler
Draft: only with a source for the right version
Send the reply: the handler, not the system

This is a way to talk about what is allowed to happen, not a finished system.

Four questions before you build

  1. Who will use this, and how will they notice if it goes wrong?
  2. Is the source material there, is it current, and may it be used?
  3. Can someone who knows the work check the answer, and what happens if it is wrong?
  4. Have we tried the smallest version against how the work is done today, and against something simpler, such as better search or a rule?

If you cannot answer the questions, it is too soon to pick a model. If a simpler solution already does the job, you may not need to build AI at all. Sometimes the conclusion is that you should not build AI right now.

One workflow, taken through to a system a person can use, is an AI production sprint.

Email me at hello@aidoni.se if you want to talk about a workflow you are considering building for.