An automation can draft a reply, write up a quote or fill in an invoice in a few seconds. What it cannot do is care whether the number on that quote is right. That is still somebody’s job, and the faster the drafting gets, the more that job matters.

Most owners know this. The harder question is how much checking each kind of work needs, because checking everything defeats the point, and checking nothing is how a wrong price reaches a customer.

Even the AI companies say checking is the slow part

On October 15, Anthropic made a paid code review tool generally available to its business customers. It is built for software teams, so it is not something a service shop would use. But its announcement says something useful for everyone: AI keeps speeding up how fast work gets produced, “and the review bottleneck hasn’t eased.”

The tool also shows a sensible design. Admins choose Quick, Standard or Deep for each project, so routine work gets a light pass and the most critical work gets a deep one. Anthropic’s own documentation says deeper reviews take longer and cost more. And the review bill has no flat price: it is charged by usage, with a monthly spending limit the organization sets.

You can borrow all three ideas for your own office: match the checking to the risk, expect careful checking to cost time, and set a limit.

Three levels of checking

Sort what your automations produce into three levels.

Level 1, a glance. Internal work that cannot hurt anyone if it is slightly wrong: sorting an email into a folder, tagging a request, a note for your own files. Someone glances at a sample once a week.

Level 2, a proper check. Anything a customer will read or act on: replies, appointment confirmations, reminders. A named person reads each one before it goes out, at least until you trust the setup. Then they check a sample.

Level 3, a person signs off. Anything involving money, contracts or deleting things: quotes above a set amount, invoices, refunds, price changes, anything that cannot be undone. Nothing in this level goes out until a named person approves it.

The point is not to slow everything down. It is to spend your attention where a mistake would cost you.

A made-up example

Here is a made-up example. A flooring company has an automation that reads new quote requests, drafts a reply and puts together a price from the price list.

Its owner sorts the work into levels. Sorting the requests into “residential,” “commercial” and “not a real job” is Level 1: someone glances at a sample on Fridays. The drafted reply is Level 2: the office manager reads each one before it is sent. Any quote over $2,000 is Level 3: the owner approves it personally.

In the first month, the office manager catches two replies that quoted the wrong room size. Neither reached a customer. By the third month, the glance check on sorting has found nothing, and the owner moves the Level 2 check to a sample of every fifth reply, and keeps Level 3 exactly as it was.

Keep the check from becoming a rubber stamp

A check that always says “yes” is not a check. Two ways to keep it honest:

  • Count how often a person changes something. If nobody ever edits or rejects anything, either the setup is excellent or nobody is really reading. Look at a few approved items yourself to find out which.
  • Keep the sign-off specific. “Approve this quote for $2,450” is a real decision. “Approve all” is not.

Give each level a named person, not “the office,” and make sure they have the time to do it.

Questions to ask whoever builds your automations

  1. Where does each step get checked, and by whom? Ask to see it, not just hear about it.
  2. Can the check be turned off by accident? Approval steps should not disappear when the setup changes.
  3. What gets recorded? Who approved what, and when.
  4. What does checking cost in time and money? If part of the checking uses an AI tool, ask how it is billed and whether there is a monthly limit.
  5. Which items can never go out without a person? Ask for the list in writing.

Your one step this week: take one automation that touches customers or money and sort what it produces into the three levels. Write the name of the person who checks each level next to it. If you cannot name someone for Level 3, do not leave it running unattended.

If you want help setting this up, here is what we build.

Sources

Kush AI Automation builds in accounts you own, and nothing that moves money happens without a person approving it first. Bring the automation you are least sure about and book a free call.