Guides · The approval gate

Use AI on customer-facing work without letting it send anything

In production, the machine prepares work behind a gate, a person approves it, and only then does anything leave the building.

The right objection

The most common objection to AI in operations is the right one: “I am not letting a machine talk to my customers.” You should not.

The machine reads the inbound work, prepares the response, and stops. A person reviews the draft, fixes it or approves it, and only then does anything go out. The machine does the reading, the retrieving, and the writing. The person does the judging and the sending.

In practice it looks like this. Work arrives overnight: emails, community messages, the day's incoming material. By morning there is a queue of drafted responses and prepared actions waiting for review. Someone approves, edits, or stops each one. Nothing reaches the audience unread. On the days something looks wrong, the gate is the reason it stayed inside.

Sort the work before you build the gate

“A person approves it” is too coarse to build from. Every step in a workflow lands in one of four categories, and you agree on them before anyone writes a rule.

Runs on its own. Reading the inbound work, running your checks, assigning an owner, moving a queue, and writing the workflow's own record of what it did. Nobody outside the company sees any of it, and a wrong one is cheap to undo, so it runs from the rules you wrote without a click.

Gets drafted. The reply, the request for missing information, the status note, the record change. The system writes them and holds them with the evidence attached, where a draft sitting in a queue has not done anything yet.

Written by a person. Scope, price, a legal or tax position, professional advice, and anything a customer could act on financially. The workflow gathers the facts and stops. The person who owns that language writes it.

Approved before it happens. Anything a customer, a vendor, or anyone outside the company sees, plus any change to a system of record that costs something to undo. One named approver per gate, with a backup.

A single workflow can be stricter than this. When it is, that stricter line belongs to that workflow, and the page for it says so instead of stating it as a rule for everything we build.

Four properties the gate needs

Everything customer-facing routes through it. A gate with a side door is theater. If the system can send under any condition, you do not have an approval gate, you have an approval suggestion.

Approval is efficient. The draft has to arrive with its context attached: what came in, what the machine looked up, and why it drafted what it drafted. A reviewer should be able to approve a sound draft from the evidence on the screen.

Rejection teaches the system or the spec. When you keep fixing the same mistake, the fix belongs in the rules, not in the daily edit. Track what you reject and why.

Exceptions stop instead of guessing. When the inbound work does not match the pattern, the correct behavior is to halt and hand a person the context. Guessing on an exception can send the wrong thing to the wrong customer with your name on it.

What the gate actually buys you

The system can check context and prepare a response before a person reviews it. The reviewer keeps the part that needs judgment: checking the evidence, correcting the draft, and deciding whether to send it.

The Lead Response & Qualification Engine is this pattern applied to inbound sales. If your version of the problem is a shared inbox or a case queue, the same gate applies.

Put inbound drafts behind an approval gate.

Describe how responses happen today. The written assessment identifies where the gate would sit.

Start the written assessment