Start With the Workflow, Not the Model: How to Scope an AI Project

The most common way an AI project goes wrong is also the most understandable. Someone sees a capable model, imagines what it could do, and the project is scoped around the technology. A few months later there is an impressive demo and very little change in how the business actually runs.

The fix is simple to describe and harder to do: start with a piece of work, not with a model.

Look for work, not for use cases

"Use case" is a slippery phrase. It invites brainstorming about what AI could do in theory. A workflow is more concrete: a request arrives, people handle it in a series of steps using particular systems and documents, and something is produced at the end.

Good candidates for a first AI deployment usually share a few traits:

• They happen often. Dozens or hundreds of times a week, not twice a quarter.
• They follow recognisable rules. An experienced person could explain how they decide most cases.
• They involve reading and re-typing. Information arrives in emails, PDFs or messages and gets copied into another system.
• Mistakes are catchable. A person can review the output before anything irreversible happens.

Quotation preparation, invoice processing, order entry, collections follow-up and routine customer queries tend to fit this pattern. Strategic decisions and one-off judgement calls usually do not.

Talk to the people who do the work

Process documents describe how work is supposed to happen. The people doing it know how it actually happens: the spreadsheet everyone relies on, the customer who always sends orders in a strange format, the approval that is technically required but usually skipped.

An hour spent watching someone handle ten real cases is worth more than a week of workshops. It shows where time really goes and which exceptions matter.

Score before you build

Once you have a list of candidate workflows, compare them on the same few questions:

1. How much time does it take today, and how often does it happen?
2. How consistent are the inputs and the rules?
3. Which systems would the AI need to read from and write to?
4. Where must a person stay in control?
5. How would we know it is working?

The workflow that scores well on volume and consistency, with manageable integrations and a clear review point, is usually the right first deployment — even if it is not the most exciting one.

Decide what should stay with people

Scoping is as much about what AI should not do. Some steps should remain human because they carry commercial risk, need relationships, or simply work well today. Writing that down early prevents a project from drifting into automating things nobody wanted automated.

The output of good scoping

By the end you should be able to state, in a paragraph, what the AI will do, which systems it touches, where people approve, and how success will be measured. If you cannot write that paragraph, you are not ready to build — and that is a useful thing to know before spending the budget.