Building a Weekly Report on Agent Activity That Takes Ten Minutes to Read

You switched on AI agents at work and now nobody can say plainly what they did last week. Someone asks in a meeting, and the answer is a shrug or a number nobody trusts.

A weekly agent activity report fixes that if it stays small. Five parts: how much work the agents produced, how often a human said no, how much got edited before approval, a short list of exceptions worth a name, and one open question the manager has to answer. That fits on one page and reads in about ten minutes.

What the report has to answer

A manager reading this report is deciding three things. Is this working. Is anything going wrong quietly. Do I need to change something this week.

Everything else is padding. Charts that show a line going up, breakdowns by department nobody asked for, screenshots of interesting outputs: all of it makes the report longer and the decision slower. If a section does not move one of those three decisions, cut it.

The four numbers

Volume

Count of items the agents produced, split by type. Drafts, summaries, scheduled items, updated records. Raw counts, not percentages.

Volume on its own means very little. It matters because it is the denominator for everything below, and because a sudden drop usually means something broke rather than that people stopped needing help.

Rejection rate

How often a human looked at agent output and said no. This is the number people skip, and it is the one that carries the most information.

A very low rejection rate is not automatically good news. It can mean the work is genuinely fine. It can also mean approvers are clicking through without reading. Look at rejection rate next to how long approvals took, and a suspiciously fast approval with a suspiciously clean record deserves a question.

A high rejection rate is easier to read. Something upstream is wrong: the instructions, the records the agent is working from, or the choice of task.

Edits before approval

The quiet number. Approved is not the same as approved unchanged.

If most output goes through with a heavy edit, the agent is not saving the time you think it is. Someone is rewriting rather than reviewing. Track this as a rough band rather than a precise figure: approved as is, light edit, heavy edit. Three buckets is enough to see the shape.

Exceptions

Not a number. A short list, with names, of anything that went sideways. An agent that acted on a record that had already changed. A draft that went out with the wrong figure and was caught late. An approval that sat for four days because the only approver was on leave.

Cap it at five. If there are more than five, the report is not the problem. Exceptions come out of the agent's work log, which is worth reading every Monday rather than only after something goes wrong.

The one open question

Every week, end with a single question the manager has to decide on. Not a list of recommendations. One question.

"Do we keep the agent drafting client updates, given three of eleven were rewritten from scratch?" is a question that produces an answer. "We should consider reviewing our approach to client updates" produces nothing.

This is the part that makes people actually read the report, because it ends with something addressed to them.

Where the numbers come from

The report is only cheap to produce if the underlying system already keeps the record. If you are pulling volume from one tool, approvals from a chat thread and exceptions from memory, the report will be late, wrong and abandoned inside two months. The same underlying record is what an audit trail needs to hold if a regulator ever asks.

This is the practical case for keeping project work, people records and agent activity on one set of records, which is how SicherOne is built. A human approves agent output before it ships, and that approval is a record rather than a message someone half remembers. The weekly report then becomes a query rather than an investigation.

If your setup is spread across separate tools, write the report anyway, but keep it to what you can pull without a manual count. A report with three honest numbers beats one with six guessed ones.

Keeping it to ten minutes

Some rules that hold up:

• One page. If it does not fit, something is being explained that should be summarised.
• Same layout every week, so readers learn where to look.
• Numbers next to last week's numbers, so the direction is visible without a chart.
• No commentary on numbers that did not move.
• Names on exceptions. Anonymous exceptions do not get fixed.

Weekly is the wrong grain for some questions, which is what a quarterly audit you can run without a consultant is for. The point of the report is not to prove the agents are working. It is to make it obvious, early, when they are not.