What an organisation actually needs before it needs AI

This is a placeholder draft. The structure and argument are here as a starting point — rewrite it in your own voice with your own examples before publishing. Set draft: false in the meta block above when it is ready.

---

The question arrives in roughly the same form every time. Someone on the executive has read about what a model could do, or been shown a demo at a conference, and wants to know what it would take to do the same thing here.

The honest answer is usually that the model is the last problem, not the first.

Three things that have to be true first

1. The same quantity has one number

Ask three departments for last month's figure and see whether you get three answers. If you do, the disagreement is not a reporting problem. It means there are three collection paths with different rules about what counts, when it counts, and who is allowed to adjust it afterwards. A model trained across all three learns the disagreement and repeats it back with more confidence.

This is dull to fix and it is the fix that makes everything after it possible.

2. The history is long enough and clean enough to learn from

Predicting an outcome needs examples of that outcome, recorded consistently, with enough surrounding context to tell one cause from another. Two years of records where a free-text notes field is doing most of the work is not the same as two years of data.

3. Someone will act on the output

A prediction that arrives after the decision window has closed, or that goes to someone with no authority to act on it, is not a prediction anyone can use. Work backwards from the decision and the person making it. If you cannot name both, the model has nowhere to land.

What to build instead, usually

• A pipeline that pulls from the source systems on a schedule, so nobody re-keys anything.
• Agreed definitions, written down, so the same question returns the same answer regardless of who asks it.
• Validation that flags records that cannot be right, rather than quietly averaging them away.
• One reporting layer that everybody argues in front of, instead of three that everybody argues about.

That work is unglamorous and it is what makes the interesting work possible later. It also tends to pay for itself before the interesting work starts.

When AI genuinely is the answer

There are real cases — forecasting where the history is good, extracting structure from decades of scanned paperwork, retrieval over documentation nobody can find. The test is whether the underlying data would satisfy an auditor. If it would not satisfy an auditor, it will not satisfy a model.