Write an approval rule an AI can follow

An approval rule is one page that tells an AI and your team who decides what. It names one task, defines a good draft, lists what the AI may never do, names the person who approves and sets in advance the evidence needed to remove the approval step. Without it, approval is either removed on a hunch or never removed at all.

Starting in draft mode has evidence behind it. In a study of 5,172 customer-support agents, an AI assistant that suggested replies, while each agent stayed responsible for the conversation, raised issues resolved per hour by 15% on average. Less experienced agents gained the most, and the most experienced saw small declines in quality.¹ The rule keeps a person responsible in the same way until your own log shows a task can run alone. AI earns its place when it drafts and a person approves sets out the wider case.

The rule takes an hour to write, and all of it comes before the first draft

Each of the five steps settles in advance a question that would otherwise be decided by opinion once the drafts arrive.

Step 1: Name one task, narrowly (10 minutes)

Pick a task your team does many times a week, where a draft would save time and an error is easy to catch. “Reply to client emails” is too broad. “Reply to client emails that ask for the status of an active project” is the right size: one kind of request, one kind of answer and one place the answer comes from.

Step 2: Define a good draft as a checklist (20 minutes)

Write four to six checks a person would apply to a draft. For a status reply: names the right project; states the current stage as recorded in the project system; gives the next milestone and its date; uses the client’s name and the company’s tone; promises nothing new. A checklist turns “looks fine” into something you can count, and it doubles as the AI’s instructions.

The second check only works if the stage is recorded in one place. If two tools both hold project status, settle which one is the record first. Run a keep, connect or replace review shows how.

Step 3: List what the AI may never do (10 minutes)

Some actions stay with a person however accurate the drafts become:

• Commit to a price, a discount or a date that is not already in the system.
• Approve a refund, a credit or a payment.
• Send anything legal, contractual or related to safety.
• Reply to a client the company has marked as sensitive.

To decide whether an action belongs on the list, ask one question: would an error cost more to fix than the AI saves in a month? If so, it goes on the list.

Step 4: Name the approver, a backup and a time limit (5 minutes)

Name the role that approves and a backup. Set a time limit: if a draft waits longer than two working hours, for example, it goes to the backup. An approval queue that nobody owns becomes the new place where work waits.

Step 5: Set the threshold before you start (15 minutes)

Decide what evidence will move the task to automatic, and write it down before the first draft. A workable threshold has three parts:

• A minimum number of drafts reviewed, such as 200.
• A minimum share approved without edits over a recent period, chosen by the owner, such as 95% in each of the last two weeks.
• No errors of the kinds on the “never” list.

Agreeing the threshold in advance keeps the decision from turning on whoever had the best or worst week.

Four weeks of logged drafts decide whether approval can go

The log turns impressions of the AI into counts the owner can act on.

Step 6: Run in draft mode for four weeks and log every draft

Log each draft as approved as is, edited or rewritten, with a few words on why. Review the log once a week. Edits that repeat are rules the AI has not been given yet, so add them to its instructions.

Step 7: Move to automatic, and keep checking a sample

When the threshold is met, let that task run automatically and have a person review a regular sample, such as one reply in ten in the first month and one in twenty after that. If an error from the “never” list appears, the task returns to draft mode the same day.

The sample matters because reviewers relax. A review of the research on automation found that people who supervise it make errors of omission and commission when the aid is imperfect, that experts are affected as well as novices, and that training or instructions do not prevent it. The lapses show up when other tasks compete for the reviewer’s attention.² An account manager approving drafts between client calls is in exactly that position.

Illustration: status replies met the threshold after four weeks and 260 drafts

An agency receives about 65 client emails a week asking where a project stands. Its rule fits on one page.

Part	What the rule says
Task	Replies to clients asking the status of an active project
Good draft	Right project; stage from the system; next milestone and date; company’s tone; nothing new promised
Never	Dates not in the system, prices, scope changes, clients marked sensitive
Approver	Account manager on the project; backup: operations lead
Time limit	2 working hours, then the draft goes to the backup
Threshold	200 drafts; 95% or more unedited in each of the last 2 weeks; no “never” errors
After	Automatic; 1 reply in 10 checked in month 1, then 1 in 20

Source: Verticulate illustration

In the first week, 29% of drafts need edits before they go out. Most of the edits rewrite how a delay is described, which the instructions never covered. Once the company’s way of describing a delay becomes an instruction, the share falls.

0%10%20%30%40%

29%

14%

5%

3%

Week 1Week 2Week 3Week 4

Drafts needing edits fell from 29% to 3% in four weeks

Week	Edited or rewritten
Week 1	29
Week 2	14
Week 3	5
Week 4	3

Source: Verticulate illustration

Note: The threshold allows 5% or fewer drafts edited in each of the last two weeks. It was met at the end of week 4, with no errors from the “never” list.

By the end of week 4, the task meets every part of the threshold: 260 drafts reviewed, 95% or more approved without edits in each of the last two weeks (5% and 3% edited), and no errors from the “never” list. Status replies move to automatic, with one in ten checked in the first month. Replies about scope or budget were never part of the task, and they still go to a person.

Four mistakes remove approval too early or keep it forever

Each one leaves the decision to opinion instead of the log.

Mistake 1: Starting with a task that is too broad

“Client emails” mixes status questions, scope changes and complaints. Drafts for each fail in different ways, the log cannot show which kind is ready, and nothing ever meets the threshold.

Mistake 2: Setting the threshold after the results are in

A threshold written after four weeks of drafts gets fitted to the results. Write it before the first draft, and change it only before the next trial.

Mistake 3: Treating approval as a formality

An approver who is busy with other work, looking at drafts that are usually right, is in the setting where people miss automation’s errors.² Keep the checklist in front of the approver and the log complete. A draft edited without a note is evidence lost.

Mistake 4: Extending automatic to the task next door

A threshold met for status replies says nothing about replies on scope or budget. Each task earns automatic on its own log, under its own rule.