Using a model is optional. Being accountable is not.
The useful question is not "can a model do this." It is what happens if the draft is wrong and we are too busy to notice.
Aside. A model will take the meeting for you in the same sense that a parrot will take the meeting for you. Both can produce sentences. Neither can be deposed.
The five no's
Skip the model — or use it only as a private scratchpad you will rewrite — when any one of these is true.
1. The cost of a quiet error is high
Quiet means the error looks finished. Prices, medical or legal advice, safety steps, access control, "I deleted that," "I paid that," "you are approved."
If a wrong send costs more than an hour of human time (money, safety, legal, a relationship you cannot buy back), the model does not send. A human may still use a model to outline, then writes the facts from the system of record.
2. There is no source of truth in the room
No ticket, no SOP, no spreadsheet, no person who knows. The model will still answer. That answer is a story.
Do the gathering first. Or write NEED and stop.
3. It is a one-off that takes longer to specify than to do
If writing a good packet takes twenty minutes and doing the task takes eight, do the task. Prompting is not free. It is work you are prepaying for reuse.
4. The job is the relationship
Apologies after you caused harm. Firing. Telling a customer you lost their data. Negotiating with a person who is already angry at a bot. You may outline talking points. You send human words.
5. You cannot name a reviewer
If nobody will read it, do not generate it for an external audience. Unreviewed external content is not "moving fast." It is unsigned mail.
The yellows (use the model, keep a gate)
These are allowed with the operator packet and the customer test:
- Rewrite for length or reading level, source sitting next to you
- Extract fields from a document you provided
- First draft of a repeating internal note
- Compare two texts you pasted ("what did they change")
- Generate options you will pick among, not a decision
Yellow becomes red if you remove the gate "because it has been fine so far."
Copy-paste: go / no-go
I will describe a task. Answer only with GO, YELLOW, or NO, then one sentence, then the missing gate if YELLOW.
>
NO if: high cost of a quiet error, no source of truth, faster to do by hand, the job is the relationship, or no reviewer.
YELLOW if: mechanical draft from a source I will provide, with a named human send.
GO if: private brainstorm or throwaway list with no facts that will be reused as truth.
>
TASK:
[describe the task, who sees the output, and what a wrong answer would do]
Do not treat the model's GO as permission. It is a mirror for your own rule.
Worked calls
- "Draft a social caption from this approved blog paragraph." YELLOW. Source exists. Someone posts.
- "Tell me if we can fire this person." NO. Employment is not a next-token problem.
- "Summarize this contract and tell me we are safe to sign." NO. Summary YELLOW for your lawyer; the safety call is human.
- "Make a list of angles for a retrospective." GO, if it stays in the room.
- "Reply to this chargeback using our policy snippet." YELLOW. Policy in the prompt. Human sends.
- "What's the customer's private address from memory." NO. If it is not in the ticket you just opened, you do not guess.
Failure modes
- Competence theater: you use the model so it looks like the team is "AI native." Usage is not value.
- Outline laundering: you generate a confident outline, then fill facts to match the outline. The shape was fiction; the facts get bent.
- Cheap task, expensive tail: the draft was free; the correction across twelve customers was not.
- Shadow sending: someone pastes customer mail into a consumer chat, likes the draft, and sends it from their phone. That is still a send. It needs the same rule.
Decision rule you can put on a wall
No source, no send. No reviewer, no external. No packet, no repeating job.
If you want a fourth line: If I would not let a first-week hire do it unsupervised, I will not let a model do it unsupervised.
Download
Go / no-go checklist: /downloads/when-not-to-use-ai-checklist.md
You are done when
You have turned down at least one tempting use this week on purpose, and you can say which no it hit.