How to Give an AI Agent More Responsibility, Safely

On a call in January I said the line I use more than any other about AI output. It is an eighty percent solution. It is a junior employee, and that is how you have to treat it.

I used the same line here two weeks ago, in a post about deciding where the human signs before you build an agent, and left it there as if it explained itself. It does not. Once the review gate exists, you still have to decide what to do with the junior employee sitting behind it.

The answer is the same three things you would do with a person. Give it the standard. Check its work. Widen what it may do alone as it earns it. The common failure is doing only the middle one, forever, until the reviewer stops reading.

Give it the standard, in writing

A capable new hire does not know on day one what good looks like at your company. Nobody expects them to. You sit them down with the last few pieces of work that went out, the price sheet, and the handful of things everyone knows and nobody wrote down.

An agent needs the same pack and usually gets less. It gets a prompt someone wrote in an afternoon and an instruction to sound professional.

The standard for an agent is three things, written before it runs.

Examples of finished work you would sign. Five real ones teach tone and structure better than any paragraph describing them.

The rules that live in people's heads. The customer who is always addressed by first name. The product that is never discounted. The report that goes out on the Tuesday even when Monday is a holiday. Every company runs on dozens of these and an agent cannot guess a single one.

What done means. Which fields must be filled, what the output must cite, what format it lands in, and where it goes next.

Writing this down is tedious, and it is most of the work. It is also what decides whether the eighty percent is eighty or sixty. When an agent gets something wrong, my first question is whether the standard said otherwise. Often it did not, and the fix is a sentence in the standard rather than a change to the model.

Check its work so the checking teaches you something

Reviewing output and learning from it are different things. A reviewer who quietly fixes a draft and sends it has protected the customer and taught the system nothing. Next week the same person makes the same fix again.

So the review needs one more step. Every correction gets a label. Wrong fact. Wrong tone. Missed a rule. Missing data. Right answer, wrong format. A column in a sheet is enough.

After a month the labels tell you what to do next. If most corrections are missed rules, the standard is incomplete, and you write the rules down. If most are missing data, the agent cannot see a system it needs, and that is an integration job, not a prompt job. If the corrections are rare and scattered, you are close to the point where the reviewer can step back.

The labels also tell you when a review has gone stale. A reviewer approving a hundred outputs a day without changing any of them is doing one of two things. Either the agent has earned more room, or the reviewer has stopped reading. The log tells you which.

Widen its authority on the record, not on a feeling

Nobody gives a new hire signing authority in their first week. Nobody keeps a good one on supervised drafts for three years either. People get promoted in steps, and each step is earned by a record somebody can point to.

An agent should move the same way, one category of work at a time, never all at once.

1. Draft only. Everything it produces is read before anything happens.
2. Draft and propose. It prepares the action, a person approves it with one click, and the approval is logged.
3. Act, with a sample checked. Routine cases go out unattended and a person reads a fixed share of them every week.
4. Act, with exceptions routed. It handles the category and sends anything unusual to a named person.

The rule for moving up a step is written before the first output, not decided in a good mood in month three. Something like: fifty consecutive outputs in this category approved without a change, and nothing labeled wrong fact in the last thirty days. Set the numbers to the risk. A meeting confirmation can move up quickly. A figure that goes to a customer or a lender moves slowly, or never.

The rule runs in both directions. When an unattended category produces a bad output, it drops back a step until the record rebuilds. You would do the same with a person, and nobody would call it unfair.

Where the eighty percent sits

One of the principles we work by is Automate the 80%. People hear it as automate most of everything. It is narrower than that. The machine does the eighty percent of a task that is routine, a person does the twenty that needs judgment, and the design job is drawing that line in the right place for each category of work.

The line moves. That is what the steps are for. A category that needed a person in January may not need one by summer, because the standard improved and the record says so. Another category may never leave draft only, and that is a perfectly good answer.

What I will not accept is the line moving by drift. An agent getting something wrong in its first week does not worry me. Everyone expects that, and the review catches it. The failure that worries me is the agent that was quietly promoted because everyone got busy, with no record of when it happened or why.

Here is the exercise that catches it. Pick one agent or automation you already run. For each kind of output it produces, write down which of the four steps it is on today. Next to each, write the result that would move it up one step and the event that would move it back down. Any category where you cannot fill in both is running on a feeling. Start the correction log this week, with the labels, and read it again in thirty days.