How to scope a thirty-day AI discovery

The failure mode to design against

Discovery phases drift into stakeholder interviews and capability surveys because those activities are pleasant and low-risk. They also produce artefacts that nobody can act on. The version that works is narrower: a fixed window, one candidate workflow, and a small set of questions whose answers are facts rather than opinions.

Three facts worth the four weeks

Document reality. Take a representative sample of the documents the workflow depends on and read them. What share is scanned, what share has tables, how many languages, how old is the oldest, and how does the extracted text look? The output is a written assessment of parseability, not a count of files. Most discovery phases skip this and then discover it during delivery, at the cost of a delivery phase.

Permission reality. Who can see which documents, where that is enforced today, and whether the mapping into a retrieval index is feasible. The output is a named owner and a concrete plan for the entitlement mapping, because the technical version of this question is easy and the organisational version is where the time goes.

Workflow reality. Where the work happens now, what the system would replace, and what the person doing it would judge as success. The output is one workflow described precisely enough that a task matrix can be drawn — inputs, steps, exceptions, handoffs — rather than a list of candidate use cases.

How to run it so the facts arrive

• Fix the window and the scope. One workflow, one document family, one team. Discovery that expands is discovery that never closes.
• Spend the time reading, not surveying. Two days with the actual documents and the actual permission model beats a week of interviews, because interviews return beliefs and the documents return the truth.
• Keep a decision log with dates and named owners. Discovery's most common failure is producing conclusions that nobody was recorded as agreeing to.
• Size the delivery phase from what you found, not from the initial enthusiasm. If the parseability assessment says a third of the corpus is image-only, that is a line item, not a surprise.
• State the conditions under which you would recommend not proceeding. A discovery that cannot conclude no is a sales process with a longer name.

What you should be able to say at the end

At the end of the window you should be able to answer five questions in a sentence each: which workflow, which documents, who owns the data, what the first version will not do, and what would have to be true for the second version to be worth building. If any of those needs a slide to explain, it has not been settled, and the delivery phase will spend its first weeks settling it instead of building.

The deliverable is short on purpose. A one-page data flow, a parseability assessment, a permission mapping with an owner, a workflow description, and a decision record. Everything else is a distraction from the only thing discovery is for: making the delivery estimate honest.