A ten-question self-check for finding the workflow most worth proving in thirty days. The same survey a craftsman makes before the first cut — applied to the work you'd hand an agent.
段取り (dan·do·ri) is the Japanese word for the preparation a craftsman does before the work begins — laying out the tools, measuring the stock, and sequencing every cut so that when the blade finally moves, it moves once and moves right. It is not the work. It is everything that decides whether the work will go well.
段取り八分、仕事二分
Preparation is eight-tenths of the work; the work itself, only two.
Most AI projects are not lost in the building. They are lost in the choosing — in picking the wrong first piece of work, then asking good engineering to save a workflow that was never a good candidate. Choose one that is bounded, measurable, and owned, and the first thirty days produce a plain, checkable result on real work. Choose one that is sprawling, subjective, or ownerless, and no amount of engineering produces a verdict anyone trusts. This guide is the preparation before the preparation: how to choose well.
Every workflow worth proving passes four checks. Three of the four carry a single question that can't be made up for elsewhere — we call those gates.
Frequency and named cost. Preparation pays back across repetition — work you run many times a day rewards a prepared agent many times a day. Q1–Q2. No gate — rare work can still be worth proving; it simply pays back more slowly.
Rules you can write, a clear start and finish, a shape that holds still. Q3–Q5. Gate at Q4: a clear finish.
A number for "done right," systems you can reach to check it, and a way to catch a bad result before it hurts. Q6–Q8. Gate at Q6: a numeric success line.
Someone to route exceptions to, and one person who can settle what "correct" means. Q9–Q10. Gate at Q10: one accountable owner.
A missing edge, an absent measure, or no clear owner doesn't slow the work down — it stops it. If any gate is a hard No, no amount of strength elsewhere makes the workflow ready. Fix the gate before you count the rest.
A menu of work that tends to make good early candidates — and a few that don't, yet. Read down the last column: the high-value work at the bottom isn't off-limits, it's earned. You don't hand an apprentice the finish cut.
| Workflow | Recurs | Time to proof | If it's wrong |
|---|---|---|---|
| Meeting & call summaries | Daily | Fast | Low |
| Knowledge & document search | Daily | Fast | Low |
| Customer support drafting | Daily | Fast | Low |
| Scheduling & coordination | Daily | Fast | Low |
| Request routing & triage | Daily | Fast | Medium |
| Proposal & quote drafting | Weekly | Fast | Low |
| Data cleanup & reconciliation | Daily | Medium | Medium |
| Invoice & document processing | Daily | Medium | Medium |
| Contract & policy summaries | Weekly | Medium | Medium |
| Financial approvals & payments | Daily | Longer | High |
| Automated production operations | Varies | Longer | High |
The lesson of the last column: prove the method where a mistake costs a correction, not a crisis. The disciplined execution you show on forgiving work is exactly what earns you the right to bring an agent to the work where a mistake costs more.
Pick one workflow and answer for it. Mark No (0), Partly (1), or Yes (2). Your running score sits at the top of the page.
一 · Repetition
A workflow you run many times a day rewards a prepared agent many times a day; one you touch twice a year barely repays the setup.
Backlog, overtime, rework, delayed cash, missed commitments. If you can name what it costs while it stays as it is, the improvement will be felt when it lands.
二 · Boundary
If a capable new hire could learn the task from a written procedure, an agent can be prepared for it. If the decision lives only in one veteran's shifting judgment, it needs to be understood first.
"Process a returns request" has edges. "Improve customer experience" has none — it's a direction, not a task. You should be able to trace it from trigger to finished state in a single breath.
If it's about to be redesigned or replaced, preparing an agent for today's version spends the preparation on something that's leaving. Prove the method on stable work.
三 · Measure
A proof needs a verdict, and a verdict needs a line drawn before the work starts: cycle time under a target, error rate below a threshold, items cleared per day. Without a measure, "better" is an opinion.
A prepared agent works inside the systems you already run — so those systems need a door: an interface, an export, an API, a screen a governed agent can operate.
A good first workflow is forgiving. A wrong answer is visible and correctable, not silent and irreversible. Prove the method where a mistake costs a correction, not a crisis.
四 · Ownership
Every real workflow has cases the rules don't cover. A prepared workflow has an answer for "when unsure, route to ___" before it meets its first hard case.
When two people disagree about the right output, someone has to decide. Without that person, the success criteria drift and the proof has no judge. One accountable owner is worth more than a room full of stakeholders.
It's ready to prove. Bring it to a thirty-day proof, agree the success criteria in writing, and let a plain result on real work speak for itself.
The fix is almost always one of two moves: narrow the scope (Q4) until the edges are clean, or define the measure (Q6) until "done right" is a number. Make one of those moves, then score it again.
Take a smaller, more bounded cut of the same work — or a different workflow entirely — and run the check again. The strongest single lever is almost always narrowing scope until the other answers turn.
We prepare the workflow you chose the way a craftsman prepares a job: the rules written down, the systems mapped, the success criteria agreed in writing before a single piece of work is done. Then we build one governed agent, run it on your real work, and measure it against the line you set.
If the proof does not meet the written success criteria, you owe no build fee. You keep the prepared workflow documentation, rules, success criteria, and integration design. Production operation begins only if you choose to continue.
Because you keep the agent, the workflow map, the rules, the approval model, the corrections, and the reusable operating logic, the second workflow starts from a foundation the first one built — and the method compounds from there.