Every consultancy sells a discovery phase. The difference between the useful ones and the expensive PowerPoint ones is simple: what do you physically hold at the end, and could you act on it without hiring the firm that wrote it? A discovery sprint should be a product with a spec, not a warm-up act. Here is what two weeks should produce.
Week one: evidence over opinions
- Workflow map — the actual sequence of who does what, built from watching and asking, not from the org chart's theory of it. The gap between the two is usually where the opportunity lives.
- Data reality check — which sources exist, who owns them, how clean they are, and the specific gaps between what your target use cases need and what your data can support today.
- System inventory — the integration surface: which APIs exist, which systems are facade-able, which are genuinely closed.
Week two: decisions over inventories
- Ranked opportunity portfolio — every automation candidate scored on value, feasibility, and risk, with the scoring visible so you can argue with it. A shortlist you can defend to a board beats a long list that flatters everyone.
- Architecture sketch for the top candidate — concrete enough to estimate honestly: which model class, which integration points, where the human gates sit.
- Proposal with real numbers — timeline, team shape, and cost for the first build, plus the explicit decision gates where you can stop without sunk-cost hostage-taking.
The test of an honest discovery: it names at least one thing you should NOT build yet, and says why. A discovery that finds only green lights wasn't discovering — it was selling.
What it deliberately is not
Two weeks won't produce a trained model, a finished integration, or certainty about problems that need production data to reveal. Anyone promising those in discovery is either scoping a different thing or borrowing against week six. What it does produce is the ability to make your next AI decision on evidence — and ownership of every artifact, whether you build with the firm that wrote them or not.