The client-safe AI checklist: questions to answer before you run an AI agent for a client
Answer: Before an agent acts for a client, answer seven questions in writing: what it may do without asking, who approves the rest, what happens at the budget limit, who can stop it and how fast, who restarts it, what the client can check without you, and where data and logs live. If an answer is "we'd have to look", fix it before go-live.
1. What may the agent do without asking?
A good answer lists action types, not intentions: "draft replies: yes; send: only after approval; refunds: never."
2. Who approves the rest, and on what?
A good answer names a person at the client and what they see: recipient, amount, exact text and expiry. Silence counts as a no.
3. What happens at the budget limit?
A good answer: the call isn't made. Alerts aren't limits. OpenAI's spend alerts, for example, let traffic continue [1]. Know which limits your providers enforce and what they return: OpenAI's hard spend limits return HTTP 429 [1], and Anthropic returns HTTP 400 at a spend limit you set [2].
4. Who can stop it, and how fast?
A good answer: the client can stop it without calling you, and the stop is checked before every action, not at the next deploy.
5. Who restarts it?
A good answer: only you, after you know why it stopped. Keeping the restart with the operator means every restart has a reason on record.
6. What can the client check without you?
A good answer: a record of each approval and stop that is signed or anchored outside your control, so the client doesn't have to trust your logs. Execution histories get pruned. Self-hosted n8n, for example, deletes execution data after 336 hours by default [3].
7. Where do data and logs live, and for how long?
A good answer names each processor and its region (database, workflow host, model provider) and a retention period for logs. "In the cloud" isn't an answer.
FAQ
Is this checklist only for n8n?
No. The questions apply to Make, Zapier, GoHighLevel and custom code alike.
Who should keep the answers?
Both of you. Attach them to the service agreement, so the client knows what the agent may do and who can stop it.
How often should we revisit it?
Whenever the agent gets a new tool, a new model or a new action that reaches the client's customers.
Does answering these questions settle our legal obligations?
No. It makes operations checkable. Legal questions go to your adviser.
Sources
- OpenAI, Spend limits: https://developers.openai.com/api/docs/guides/spend-limits
- Anthropic, Rate limits: https://platform.claude.com/docs/en/api/rate-limits
- n8n, Executions environment variables: https://docs.n8n.io/deploy/host-n8n/configure-n8n/basic-configuration/use-environment-variables/executions
Disclosure: this page is written by the team building Keelstamp, a tool for client approvals, per-client caps and verifiable receipts. It isn't released yet, and everything above works without it.