In 30 seconds

Many AI projects fail before the AI is even involved. The workflow is unclear, the data is inconsistent, or nobody owns the source information. Starting with the model hides those problems instead of solving them.

Start with the work. Add AI only when you understand what the work actually needs.

Why this comes up

Someone asks, "Can we use AI for this?" The exciting part of that question is the AI, so the project begins with prompts, tools and a demonstration.

The demonstration usually works. That is the trap. A demo runs on a clean example that someone chose. Real work is full of missing fields, contradictory records and exceptions nobody wrote down.

AI does not remove uncertainty. It inherits it.

If two spreadsheets disagree about the current price, a person has to decide which is right. An AI does not fix that disagreement. It just picks one, often convincingly, and now the wrong answer looks trustworthy.

This is the core reason to start with the work. A confident wrong answer is harder to catch than an obvious blank.

A small example

A team wants an AI to answer staff questions from an internal handbook. The demo answers three questions perfectly.

In practice, the handbook has three versions in different folders, two of them out of date. The AI retrieves whichever passage matches the wording best, sometimes from an old version. The real problem was never the AI. It was that nobody owned the handbook.

Fixing the source made the assistant useful. No new model was required.

This is not anti-AI

Starting with the work is not a reason to avoid AI. It is how you find the places where AI genuinely helps, and avoid spending months on a demonstration that cannot survive a normal Tuesday.

Often the honest answer is that a small amount of structure, or a simple rule, solves most of the problem. Sometimes AI then adds real value on top. That order protects your time.

What this does not solve

Understanding the work does not guarantee a good outcome. You still have to design, test and operate the workflow. It only makes sure you are solving the right problem.

Good enough to move on when

  • You can state the operational problem in one sentence.
  • You know roughly how the work is done today.
  • You know where the important information is supposed to live.