The AI agent approval checklist. Who is accountable, and what happens when it goes wrong.
Eighteen checks in five groups. Each is a rule we use on our own company, so you can copy it into your own setup today. Read it all here, no signup.
What should be approved before an AI agent acts?
Before an AI agent touches a client, money, email or a public post, a named person should approve it, the agent should hold only the access its job needs, spend should have a cap, and every change should leave a record. This is the 18-point checklist ShooflyAI runs its own company on, free to use today.
What always needs a human
-
Anything a client or prospect will see waits for approval.
Forms, decks, proposals, pages and portal updates carry your name, and an agent draft is not a commitment.
-
Anything that states a price, fee or total gets a higher bar.
A wrong number sent once is hard to take back, so only an owner can approve it, never an agent and never a timeout.
-
Changes to a deal in your CRM that move money or ownership wait for an owner.
Amount, won or lost, and who owns the deal drive your forecast. Nothing is ever deleted by an agent.
-
No agent posts in a public forum on its own.
A draft goes through a check and a person, with disclosure, one reply per thread, from a labelled account.
-
No agent sends email.
It writes the text and a person sends it. We made this a rule because one slipped once.
Who approves what
-
Every approval has a named approver, not "the team".
When everyone can approve, no one owns the result.
-
Put the decision where the person already is.
A choice that can be made in chat gets a button, so approval takes one tap and the answer comes back with the person's own note.
-
A typed "go" counts only when it points at the exact request.
We accept it only if it is the approver's latest message, under 15 minutes old, a plain yes and used once. Money and client-facing sends still need the button.
-
Approving from a phone needs a passcode.
A lost or borrowed phone should not be able to approve a change.
Spend and access limits
-
Give each agent its own scoped key.
It can reach only the data and tools its job needs, and a refusal is reported to a person, not worked around.
-
Agents write to business systems through one governed path.
No agent holds a raw token to your CRM or other core system, so every write is checked and recorded the same way.
-
Cap spend per person and per month.
When a cap is hit, the agent drops to a cheaper model instead of running up the bill.
-
Agents publish drafts only.
Live sites are changed by an approved card on a set schedule, so a rushed batch cannot go out at once.
The audit trail
-
Every agent change is a card with a receipt.
The receipt shows who asked, what changed, who approved and when.
-
An agent says "done" only with proof from this turn.
After any write it reads the record back and shares the link. If the read-back differs, it says it could not confirm.
-
Keep an undo for every kind of change that can have one.
The fastest fix for a wrong change is to reverse it, then talk about why.
When something goes wrong
-
A failure gets one plain line with the real error, then it stops.
No silent retries and no cheerful updates over a broken step.
-
Every situation has a named person to go to.
Decide in advance who gets a failed job, a blocked deliverable or a pricing question. After two unanswered asks, escalate once and stop nudging.
-
Passwords and keys never go in chat, documents or commands.
They go into a masked entry in a protected store, and the agent only ever sees a reference.
Get the editable copy and our agent policy template.
The checklist above is yours to read and use. Leave your email and we will send the editable version plus the agent policy template we use.
FAQ
What is an AI agent approval checklist?
It is a short list of checks that decide when an AI agent can act on its own and when a named person must approve first. This one covers five areas: what needs a human, who approves what, spend and access limits, the audit trail, and what happens when something goes wrong.
What should an AI agent never do without a human?
Anything a client will see, anything that states a price, changes to a deal that move money or ownership, sending email, and posting in public. Those are the five we hold back from agents in our own company.
Who should approve what an AI agent does?
A named owner for each kind of action, with a higher bar for pricing and client-facing work. Spread approvals across people only when each one is accountable for what they approve, and require a passcode for approvals from a phone.
What if an AI agent makes a mistake?
Reverse it if the change has an undo, read the record to see who asked and who approved, and fix the rule that let it through. Plan for this before launch: a named person for each type of failure, and one plain line about what broke.
Want a person to run this with you?
ShooflyAI is the AI operations partner for owner-led $5M to $75M companies. We run the AI your company uses, rented or custom, under one human-approved operating model, and we prove the business result.