Running Work Through AI Agents: Start With One Role and a Human Check

A persona drafts. An agent does.

That difference is everything. A persona hands you text and waits. An agent can click, send, move money and change things in the world - which means its mistakes happen to other people, at your name. Brief the persona first; the same six-part brief is the foundation an agent stands on. The agent adds one more layer: limits on what it may touch.

The pilot: one job, drafted replies

A worked example, fictional: Nova's candle shop gets the same five questions every week. Her pilot agent drafts replies - nothing more. The setup, before anything runs:

1. Inputs. One inbox label only: "routine questions". Nothing else in the inbox is visible to the agent.

2. Approved facts. The same fact list her persona brief uses: hand-poured, ships in three working days, current scents only. The agent may not use anything else.

3. Prohibited actions. No sending. No deleting. No marking read. No replies to refund, complaint or payment messages - those get flagged for Nova, untouched.

4. Review owner. Nova reads every draft before anything sends. Not a sample. Every one, at first.

5. Stop rule. Nova's chosen rule for this pilot: if two drafts in a row invent a fact, the pilot stops that day - not "when she gets to it". Yours can be stricter or looser; what matters is that it is written before anything runs.

6. Rollback. Stopping means removing the agent's access in the tool's settings, because instructions are not permissions. That stops future work; it does not undo drafts already written - review those separately and delete any you do not want kept, in the tool where they live.

The permission boundary, laid out: BEFORE THE LINE - the agent may draft, sort and suggest. PAST THE LINE - sends, spends, refunds and deletes wait at the approval gate: your yes each time, or a written standing grant you set that names the action and its limits. Anything outside the grant stops and asks. Notice what the pilot never touches: sending, refunds, prices, ad spend. Those stay past the line for this pilot - whatever the tool offers.

Classify the action before you allow it

Green weeks do not widen the road

Four steady weeks of good drafts prove the agent can draft. They do not prove it can send, spend or be reviewed less. Widening what the agent may do is a new decision, made on purpose, one step at a time - never a reward the agent earns by not failing.

The standing rule: a pilot that never breaks is still a pilot. Steady drafts do not authorize sending. Clean runs do not authorize spending. And nothing authorizes reviewing less until the numbers - not the vibe - say the drafts need no fixes.

Sample feelings, measure facts

"It's saving me hours" is a feeling until you count. A sample - the three drafts you happened to read closely - is not a measurement. The log below counts what the pilot actually did, including the errors you caught and the minutes reviewing cost you. Review time is part of the price; an agent whose drafts you must rewrite is a cost, not a saving.

Fill it weekly for four weeks before you decide anything. That includes deciding to keep going.

How I run my own assistant

This is not theory for me. My own assistant, Instinct, works under limits like these: it drafts and prepares, and the actions that matter wait for my word. You can meet Instinct here - same discipline, real business.

Your next step

A pilot that works deserves to become a process: the same steps, the same limits, written so anyone - future you included - can run it. Write the repeatable process next.

Note. All shops, agents and outputs in this guide are fictional examples written for practice. No performance figures here are measured results - the measurement log is blank on purpose, for your numbers. What a specific tool can access or do lives in that tool's own settings and permission documentation; check them before connecting anything.