Most AI agent demos look impressive. The question operations leaders ask me is rarely whether an agent can do the work. It’s what happens when it gets something wrong, and who finds out.

My answer, almost every time, is an approval gate.

What an approval gate is

An approval gate is a point in a workflow where the agent stops, shows a person what it intends to do and why, and waits. The person approves, edits or rejects. Nothing after the gate runs until they do.

It sounds like it would slow everything down. In practice the agent has already done the slow part: reading the documents, checking the data, drafting the reply. The person spends seconds on a decision instead of twenty minutes on the whole task.

Step 1Request arrivesRefund request by email
Step 2Read and checkOrder, policy and history
Step 3Draft decisionRefund £84, with reasons
Step 4ApproveOne click, or edit
Step 5Refund issuedPayment system updated
Step 6Reply sentCustomer told the outcome
A refund workflow with one approval gate.
  • AI agent
  • People
  • Systems

Where to put them

Gates cost attention, so put them where a mistake would be expensive:

  • Money leaving the business: payments, refunds, credit notes.
  • Anything a customer or supplier will see: emails, quotes, contract changes.
  • Changes that are hard to undo: closing accounts, deleting records, bulk updates.
  • Decisions with a regulatory angle, such as KYC outcomes.

Everything else, such as reading, classifying, looking things up and drafting, can usually run without one.

What a good gate shows

A gate is only as good as the information on it. If the reviewer has to open three systems to check the agent’s work, they’ll either rubber-stamp it or stop trusting it. A good gate shows:

  • The proposed action, in plain words.
  • The evidence the agent used: the source document and the values it pulled out.
  • Anything the agent was unsure about, flagged rather than hidden.
  • Approve, edit and reject, each one click away.
  • A record of who decided, when, and what they changed.

Keep the clever part small

The most reliable agentic systems I build use AI for the parts that need it and plain code for the rest. Scoring, calculations and data movement are deterministic: the same input gives the same output, and they’re easy to test. The model reads unstructured input, drafts text and decides which step comes next. The smaller that surface, the fewer surprises reach the gate.

Widen the gates as trust builds

Start with gates everywhere it matters. Then track how often people approve without changes. When a type of decision is approved unchanged week after week, move it to spot checks, and later to fully automatic with a daily summary. That way an agent earns its autonomy instead of being handed it on day one.

Use the tool people already open

A gate doesn’t need new software. It can be a Teams card, an Airtable view, a Power Apps screen or an email with approve and reject buttons. Put it wherever the reviewer already spends their day. On a recent sourcing project, the gates live in Airtable because that’s where the team already tracked their suppliers, and adoption was never a question.

If you’re weighing up an agent for a process and risk is the sticking point, approval gates are usually the way through.