The First Workflow

The default move is to pick the highest-impact workflow. Usually the one a senior leader is excited about. Usually customer-facing, usually visible, usually the one with the most political weight on it.

This is wrong. Not because the workflow is wrong — but because the first workflow is also the riskiest one to fail at. You don't yet have eval discipline. You don't yet have a clean data pipeline. Your team hasn't run a production AI workflow before. The methodology has to be learned on this one, and failure has costs.

Picking the highest-impact workflow first means you're betting the biggest poker chips on the hand where you're still learning the rules.

The right first workflow is the one that scores high on both:

• Pain — somebody on your team is doing this work today, by hand, and visibly suffering. They'll celebrate when it's automated. They'll co-build the eval set willingly.
• Leverage — the work has a clear correct answer (so evals are possible), happens often (so the dataset is real, not synthetic), and lives inside systems the team already uses (so we're not also doing a migration).

“Highest-impact” usually scores high on impact but low on leverage — the data is messy, the “right answer” is subjective, and three teams have to coordinate. You want high pain, high leverage, moderate impact for the first workflow. Trust your team to scale from there.

Concrete signs you've found a real first workflow:

• Someone can describe what “done correctly” means in one sentence.
• You can find 50–200 examples of the work already done by humans, with the “right answer” visible.
• The work happens on data that's already in a system (not stored only in someone's head, or in PDFs scattered across SharePoint).
• The cost of being wrong is recoverable — a human can review and correct, the world doesn't end.
• The cost of being right is meaningful — hours saved, tickets closed, deals moved, complaints prevented.

If you can answer yes to four of five, that's your first workflow. If you can't, keep looking. The wrong workflow can burn an entire quarter and the credibility of the whole AI program.

We score each candidate on a 1–5 scale across five dimensions: pain visibility, data readiness, integration cost, eval feasibility, and time-to-value. Total possible: 25.

Workflows scoring 20+ are first-workflow candidates. 15–19 are second-build candidates (try them once you have eval infrastructure in place). Below 15 are usually conversations to revisit after a year.

Most engagements surface 3–5 workflows in the “20+” range. We pick the one with the cleanest data pipeline among those — not the one with the biggest business splash. The splash comes in workflow #2 or #3, after the team has internalized the methodology.

The discovery sprint output isn't a single workflow — it's a sequenced roadmap. The first build proves the methodology. The second build uses everything you learned in the first one (eval framework, integration patterns, drift monitoring) at half the time.

Track the rest of the candidate list. As your team's eval discipline matures, the “impossible” workflows become reachable. The biggest-impact workflow on your initial list is usually workflow #3 or #4, not workflow #1.

This is why operators who pick the boring first workflow tend to ship five real systems in the time tool-shoppers ship none.