Back to Blog

AI Agent Approval Gates and QA: How We Keep Them Honest

October 3, 20266 min read

Status: EXPERIMENTAL. We use AI helpers on our own products. We have not set this up for a client yet. A small pilot is possible.

AI agents are fast. They are also capable of being wrong in a very confident voice. If you let them publish, send or spend without anyone looking, you will eventually pay for it. This post explains the two habits we rely on when we use AI helpers: approval gates and a second check. We tie both to two of our values, Truth and Duty.

We are Apne, a small team in Lahore. We use AI helpers on our own products and could discuss a similar structure for another business as a small pilot. We have no client results to quote, so this describes a method, not an outcome.

What is an approval gate?

An approval gate is a point where an agent's work stops and waits for a person to say yes. The agent can prepare, draft and propose. It cannot make the final move on its own.

In plain terms:

  • The agent drafts a post, a page or a message.
  • The draft goes into a queue, not out to the world.
  • A named person approves, edits or rejects it.
  • Only then does it go live, and what was approved and when is recorded.

Common guidance on human oversight of agents makes the same point: decide, action by action, which actions need a person, and require approval before anything public, irreversible, financial or customer-facing. Lower-risk internal steps can be reviewed after the fact once they have been tested.

Write the rules down

A gate is only useful if the rule behind it is written. Good rules are short and plain. For example:

  • Nothing is published, sent, spent or posted without a person's yes. Drafts are drafts until approved.
  • Agents never handle someone's login. A person signs in themselves.
  • Claims are checked against a list of what is really delivered. Anything not delivered yet is labelled as such.
  • When a rule is unclear, the agent asks instead of guessing.

Writing the rule is the Duty part. It means no one has to remember it, and nobody can say later that it was never agreed.

What is the second check?

The second check is a separate step whose job is to look for mistakes before a person sees the work. It is not done by the same agent that wrote the draft. A writer reviewing its own work tends to repeat its own errors.

It asks things like:

  • Does every claim match the source?
  • Are the numbers, names and spellings the same as in the source files?
  • Is anything in the draft on the banned list, for example results we cannot prove, testimonials or prices nobody has approved?
  • Does it follow the written rules for this kind of content?

A small kind of mistake it catches

Here is an example of the sort of slip this catches. A draft says delivery is free, while the source only gives a price that includes delivery. That is a tiny difference, but it is exactly what becomes a false promise. Checking the draft against its source catches it before anything goes out. Most of the value of a second check is catching small slips like this, consistently.

We are not saying agents are unusually bad. Every writer, human or not, needs a second pair of eyes.

Truth and Duty, in practice

  • Truth means we do not publish a claim we cannot back up, and we say plainly when something is not done yet.
  • Duty means the checking is a routine that happens every time, not something that depends on someone feeling careful that day.

Gates and checks are how those two words become a procedure.

How to set up gates for your own business

You do not need to start big. A workable first version:

1. List what the agent may do without asking

Internal, reversible, low-risk steps. For example, sorting notes or drafting an outline.

2. List what always needs a yes

Anything customer-facing, public, financial, or hard to undo.

3. Name the approver

A specific person, with a backup. "The team" is not an approver.

4. Add a second check

Give it the written checklist above, and let it flag problems rather than fix them silently.

5. Keep a short record

What was proposed, who approved, when. Review the record regularly and adjust the rules when something slips through.

What we have not done yet

  • We have not set up an agent team with gates for a paying client yet.
  • Gates and checks reduce mistakes; they do not remove them. We cannot promise zero errors and do not.
  • We do not have a number for how many mistakes the second check catches, so we quote none.
  • We have not built customer chatbots or WhatsApp automation, and this post does not describe them.

Keep reading

Call to action

If you want to talk about gates and a second check for one routine in your business, send a message on the contact page and tell us what the routine is.

Frequently asked questions

What are approval gates for AI agents?

They are points where an agent's work stops until a named person approves it. The agent can draft and propose, but publishing, sending or spending waits for a human yes.

Does a second check mean the AI is always right after it?

No. A second check catches many small errors before a person sees them, but it can miss things too. That is why final approval stays with a person.

Can a small business use approval gates without technical staff?

Yes, in principle. The core is written rules, a named approver and a short checklist. We can talk about setting that up as part of a small pilot.