Pick one workflow. Draw the gate.
Al Sene, Founder & CTO
- ai-transformation
- running-the-initiative
- gated-agents
- playbook
Pick one workflow. Draw the gate.
The first agent project usually fails before the model does. It fails at the whiteboard: eight candidate workflows, a demo on the calendar, and no shared picture of what "done" looks like — or of what a tired reviewer is supposed to check at 2 a.m.
You don't need five agents. You need one workflow you can operate, with a human gate that still works when the person who designed it is asleep.
Those are the first two plays in Phase 2 of our AI Transformation Playbook, Gated Agents: pick one workflow, then draw the gate. Single agents, defined steps, human approval on every action that matters. That is the production pattern that scales trust, and for some outputs it is also the lasting ceiling.
Gartner's public article on its 2026 Hype Cycle for Agentic AI (April 15, 2026) says most deployments remain narrowly scoped, and that "fully autonomous agents are not ready for the majority of enterprise use cases." Phase 2 is how you run on that fact on purpose.
Three tests for the one workflow
The playbook's test is short. High repetition. A clear definition of done. Failure that is recoverable. Prefer internal or low-blast work for the first one.
High repetition. The work happens often enough that the team will actually run the gate and see patterns. A workflow that shows up twice a quarter produces a demo, not an operating habit. Rank by how often a complete instance occurs, not by how impressive the slide looks.
A clear definition of done. Two people who do this job should agree, without a meeting, whether a given output is finished. "Draft the coding for this vendor invoice" can be done. "Write a good strategy memo" can't — not as a first agent workflow. If you can't write "correct / wrong / must never ship" for it, you don't have a definition of done yet. That writing is also how you start a golden set, before the first prompt change.
Failure you can recover from. A bad draft can be caught, corrected, or rolled back before it costs you something you can't undo. If the next step is a payment, a deletion, a legal notice, or anything that reaches a customer, a child, or a family, it is the wrong first workflow. Some of those surfaces stay human-gated permanently, no matter how good the model gets.
Start inside the building, or keep the blast radius small. Anthropic's Enterprise AI Transformation Guide says to prioritize projects "where failure creates minimal business disruption," and to "avoid selecting customer-facing applications or mission-critical processes for initial pilots regardless of their potential value." If you have to start near a customer surface anyway, keep the gate thick: more checks, narrower tools, and no outbound path without a named human.
If you finished Phase 1, you already have a candidate list of three to five workflows, ranked by repetition times clarity of done. Run the three tests and pick one. Adding a second workflow before the first gate is real is how a pilot turns into a pile of demos.
Draw the gate as a 2 a.m. checklist
Play 2 asks two things. Which steps may the agent auto-draft? Which steps need a human to approve, edit, reject, or escalate? Write the answers as a checklist an operator can run at 2 a.m. without calling you.
A reviewer who can't name the checks isn't reviewing. They're clicking. We already wrote about where that leads.
For each step, name:
- Auto-draft. What the agent may prepare for a human to see: which fields, sources, and format.
- Approve. The output can go forward as-is. Say what "forward" means — queued for payment, merged, sent, published — and who is allowed to do that. Approving a draft is not the same as sending it.
- Edit. The human changes it, then it can go forward. Edits are a signal, and they should be visible later.
- Reject. It doesn't go forward. It goes back, or it is dropped, with a reason from a short list.
- Escalate. It is outside this reviewer's authority or the happy path. Name who receives it and what they decide.
Then write what the reviewer is actually looking at: facts against the source, policy, the "must never ship" cases, and whether this is even the right next step. Keep the list short enough to run when they are tired.
Also write what the agent may never do. The playbook's default is strict on purpose: no cold outbound messages, no irreversible writes, no payments, no account deletion, no broad data export. Expand a tool only when a named person owns that expansion and can tell you how to roll it back.
Hand the checklist to someone who didn't write it and have them run a real example without you in the room. If they have to guess, it isn't drawn yet. Cross-train against the page so you don't end up with the playbook's one hero operator.
A written checklist is still a habit until something makes it hard to skip. Ada, our VP of Engineering, wrote about moving that layer into config, not a promise. Draw the gate first so you know what the settings are supposed to enforce.
A sketch: coding vendor invoices
This is an illustration, not a case study. An invoice arrives as a PDF. The agent drafts the vendor, invoice number, date, amounts, tax, a suggested ledger code and cost center from prior invoices with that vendor, and whether it matches an open purchase order. A person in accounts payable reviews that draft before anything is queued for payment. Payment itself stays on the path you already trust.
- Auto-draft: the extracted fields, the suggested coding, the PO match, and flags for no matching PO, a repeated invoice number, an amount over a threshold you name, or a vendor name that doesn't quite match the record.
- Never: email the vendor, approve or release payment, create a new vendor, change bank details, or write anything irreversible in the finance system.
- The reviewer checks: the vendor on the draft is the vendor on the invoice; amounts and tax match the document; the ledger code and cost center fit this spend; a claimed PO match is real; no escalate condition is hiding in the flags.
- Approve means the coding may enter the existing payment path. Edit means fix the coding or the notes, then approve. Reject — unreadable, duplicate, or the wrong legal entity — goes back to whoever submitted it. Escalate a new vendor, an amount over the named threshold, a missing PO where policy requires one, a request to change payment details, or anything else off the written policy, to a named person who can decide.
The reviewer isn't rebuilding the invoice or inventing policy at 2 a.m. They're running the list. If a case isn't on the list, they escalate.
The next Phase 2 habits are the golden set and watching the approve button. Widen volume, not autonomy.
What to do this week
Write down three to five candidate workflows, or pull the list you already made in Phase 1. Score each one: does it repeat, can two practitioners define done, and can you recover if the agent is wrong? Pick one. Park the rest.
Then write the gate on one page: what auto-drafts, what the agent may never do, what approve / edit / reject / escalate mean, what "forward" is, and what the reviewer checks. Give that page to a colleague who didn't write it and have them run one real example. Fix whatever they had to ask you.
That is a drawn gate. It is also how Phase 2 actually starts.
Get the playbook
The full Phase 2 chapter — the seven plays, the exit artifacts, and the failure modes, including the gate checklist and the one-hero-operator repair — is in the 231 Labs AI Transformation Playbook. It's free at 231labs.com/playbook.