AI Automation Client Report: What to Send Every Month So the Retainer Survives

If you charge a client a monthly fee for an AI employee, the first month ends with a quiet problem. The agent did its work in the background, the client saw none of it, and an invoice arrives for something they cannot point to. That is how a monthly fee gets cancelled in month three.

The fix is a one page report that goes out on the same day every month. Most advice on client reports is written for marketing agencies and starts with traffic charts. An AI employee is different. Its output is work done on a schedule, and the thing a client needs to see is whether that work happened, what it caught, and what is waiting on them.

This is the report I would send, section by section, and how to have the agent write most of it for you.

What should an AI automation client report include?

Five things, in this order. Keep the whole page short enough to read in two minutes on a phone.

1. What ran. The jobs the employee was supposed to do this month, and how many of those runs completed. Plain counts, no percentages dressed up as scores.

2. What it caught. The items that mattered: a lead that went cold, a broken page, a customer reply that needed a human. Name three at most.

3. What it stopped on. Every time the agent refused to act and waited for a person, listed with the reason. This section is the one clients trust most, and almost nobody includes it.

4. What needs a decision. Anything waiting on the client, with a date it has been waiting since.

5. What changed. Any edit you made to the employee's instructions, and why.

Notice what is missing. There are no impressions, no vanity totals, no hours saved estimate. Those belong in a report about marketing. An employee report answers a smaller question: did the person I hired do the job this month?

Why does the stop list matter more than the results?

A client cannot judge whether an AI employee did a good job on a hundred routine tasks. They can judge whether it knew when not to act. A list that says the agent held back a reply to an angry customer, waited for approval before touching the ad budget, and flagged an invoice that did not match the order tells the client the system has brakes.

I covered the design of those brakes in the post on an AI agent approval workflow. The report is where that work becomes visible. If your employee has no approval step, the stop list will be empty, and an empty stop list is a warning, not a good result.

Where does the data for the report come from?

From the run log, not from memory. Every scheduled job should append one line when it finishes: the date, the job name, whether it completed, anything it flagged, and anything it waited on. The report agent reads that log for the month and does the counting.

This matters because a language model asked to summarize a month from memory will invent a tidy story. A model asked to count lines in a file will count them. Give it the file and tell it to quote the lines it relies on.

Three rules keep the log honest:

• One line per run, written by the job itself at the end, never written afterward by a person.

• A failed run writes a line too. A missing line is the worst outcome, because it looks like nothing happened.

• Each line carries the client name, so one log never mixes two accounts.

If you are weighing how to know whether an agent is performing at all, the post on how to measure AI agent performance goes through the checks worth logging.

How do you have the agent write the report?

Give the report its own scheduled job, separate from the employee it reports on. A job that grades its own work will grade generously.

The instruction to the report agent can be short:

• Read the run log for the client for the previous calendar month.

• Count completed runs and failed runs per job.

• List the flagged items, three at most, choosing the ones with the largest cost if missed.

• List every stop, with the reason as written in the log.

• List items waiting on the client, oldest first, with the age in days.

• Write it in plain sentences under five headings. Quote the log line behind every claim.

• If anything is missing from the log, say so at the top of the page instead of filling the gap.

The last line is the important one. A report that admits a gap is worth more than a clean report that hides one.

The agent drafts, and you read it before it goes out. That is the same rule as every other client facing message: the agent stages it and a person sends it. For a monthly report that is about ten minutes of reading, which is the part of the job you are being paid for.

What does a failed month look like in the report?

Put it first. If a job failed for eleven days because a login expired, the top line says so, with the dates, the cause and the fix.

Clients forgive a failure that is reported early and fixed. They do not forgive one they find themselves. The report is the cheapest place to be the first to tell them, and the account usually survives it.

The same goes for the cost line, if your arrangement includes one. If the employee runs on the client's own agent plan, say what share of it the work used. If it runs on yours, say that too. Surprises on a cost line do more damage than the cost itself.

How long should the report be, and when should it go out?

One page. About 300 to 400 words. Send it on a fixed day, the same day every month, ahead of the invoice rather than with it. When the report arrives before the bill, the bill reads as payment for something you just showed them.

If a client wants more, add a call, not more pages. A fifteen minute call after the report catches the things a document cannot, such as a change in their business that should change the employee's instructions. Those changes belong in section five, and they are also the reason a monthly arrangement is worth more than a one time setup.

For how the fee itself is shaped, see how to sell AI employees to clients, and for what to collect before you install anything, the post on AI agent client onboarding. If you want a version of the same page for your own business, the AI weekly business report covers that.

Should the report include a risk section?

A short one, yes, once a quarter. Write down what the employee can touch, what it cannot, and what would happen if an instruction were wrong. The structure in the NIST AI Risk Management Framework is a reasonable reference for the vocabulary, even for a one person service: map what the system does, measure how it behaved, and manage what you found. You do not need the framework's full process. You need the habit of writing the risks down where the client can read them.

A template you can start from

Copy this structure into the report agent's instructions and adjust the headings to your client's words:

• Month and client name

• One sentence: did the employee do its job this month, yes or no, and what is the exception

• Ran: jobs and counts

• Caught: up to three items

• Stopped on: each stop and why

• Waiting on you: each item and its age

• Changed: each instruction edit and why

• Next month: one thing you will do

That is the whole page. Everything else is decoration.

Where to start

Hire a first employee on your own work and write its run log for a month before you sell anything. The report is easy to produce once the log exists, and nearly impossible to fake without one. The eight AI Employees are open source in the AI Employees repository, and the Claude Code documentation explains how scheduled work runs in a desktop agent app.

If you want the broader toolkit, the Agent Ops Club has a free plan that includes the walkthrough for hiring your first employee and a free scan you can run on a prospect's website. The club is also hosting a live release of two new premium AI Employees on Saturday, November 14 at 10 AM Eastern. You can read about it and register on the event page.

FAQ

How often should I send an AI employee report?

Monthly is enough for most clients. Send a short weekly note only in the first month, while the client is still deciding whether to trust the employee.

Can the AI employee write its own report?

Not the version that goes to the client. Use a separate job that reads the run log, and read the draft yourself. An agent grading its own work leaves out the failures.

What if the client never reads it?

Then the stop list and the waiting list are doing their job in the background, because those are the sections that prompt a reply. If they still do not reply, a fifteen minute call is the next step.

What should I do if the log is incomplete?

Say so at the top of the report. Fix the job that failed to write its line, and note the repair under what changed.

Do I need a dashboard?

Not at the start. A one page document sent on a fixed day beats a dashboard nobody opens. Add one later if a client asks for it.