The daily exception report: what your AI agents should tell you each morning

Once AI agents are doing real work, you need one message a day that tells you what needs you — and nothing else. Most reports fail in one of two ways: a long list of “done, done, done” that you stop reading by week two, or no report at all, so problems surface when a customer rings. The fix is an exception report: short, consistent, and organised around decisions rather than activity.

This is the review loop that makes delegation sustainable. Approvals handle individual actions in real time; the daily report catches everything that fell between the cracks. Both are part of the approach in our guide to delegating work to AI agents.

What an AI daily report for business owners should contain

Order matters. Put the things that only you can do at the top, so that if you read three lines over breakfast, you have read the important three.

1. Decisions waiting on you, oldest first, each with the one-line question and a link back to the item.
2. Stuck work: tasks the agents could not finish, with the reason (missing data, access error, conflicting rules) and what they need.
3. Rule breaks and anomalies: anything that didn’t match expectations — a payment that doesn’t reconcile, a lease billed differently to the signed amount, a supplier invoice twice the usual size.
4. Tomorrow’s risks: deadlines in the next 48 hours that are not on track, appointments without confirmation, bonds or renewals coming due.
5. Counts: one line of numbers for routine work completed, so you can see volume without reading it.

The “anomalies” section is often where the value is. Tony Lorenzo’s agent at Elite Partners Real Estate caught a commercial lease being billed about $1,049 a month below the signed lease — exactly the kind of thing that should be the first line of a morning report, not buried in a reconciliation log. Individual results vary.

What to leave out

• Lists of routine successes. “Sent 12 reminders” is a count. Listing all twelve teaches you to skim.
• Narrative. No “Today was a productive day!” The report is a tool, not a diary.
• Things you already approved. You know about them. Include them only if something happened afterwards, such as a bounce or a reply.
• Raw logs. The full record should exist and be searchable, but it doesn’t belong in the report.
• Duplicate items. If an exception is still open from yesterday, show it once with its age (“3 days”), not three times.

A copyable exception report template

Ask your agent to use the same headings every day, in the same order, and to omit any heading with nothing in it. Consistency is what makes a report fast to read.

Morning or end of day?

Most owner-operators prefer the morning, because it sets up the day and includes anything that happened overnight (bank feeds, after-hours enquiries). End-of-day suits teams with a handover to another shift or time zone. Pick one; two daily reports is one too many.

Keeping task status honest

Reports are only as good as the statuses inside them. “In progress” is the most dishonest status in business, for people and agents alike: it can mean actively being worked on or untouched for a fortnight. Use a small set of statuses that each imply a next action:

Ban “in progress” without a date. If the agent can’t say when something will be done or what it is waiting for, it is stuck, and should say so. This is also why definitions of done matter: without one, “done” is a matter of opinion. See how to write instructions for AI agents.

When agents get stuck: the exception queue

Stuck items need a home that isn’t your memory. The exception queue is simply the running list behind the “Stuck” and “Needs you” sections: each item has an owner, an age and the thing that would unblock it. The report shows the queue; the queue is the source of truth.

Two rules keep it healthy. First, every item has a named human owner, even if the agent is doing the work, so nothing sits ownerless. Second, anything older than a threshold you set (three working days is a common choice) moves to the top with its age in bold. If the same type of exception keeps appearing, it is a missing rule or a broken connection — fix the cause, not the instance.

For approval items specifically, the queue is where unanswered approval requests land, so they don’t silently expire. Our approval gates article covers how to stop those piling up in the first place.

The weekly review that keeps the system improving

Once a week, spend about half an hour with the agent on four questions. You can do it as a chat thread or a voice note while driving:

1. Which exceptions repeated this week, and what rule or fix would stop them?
2. Which drafts did I edit most, and what should the brief say instead?
3. What is waiting on someone else for more than a week, and should we chase or close it?
4. Is any task ready to move up or down an autonomy level?

Ask the agent to prepare a one-screen summary before the review — the week’s counts, the repeated exceptions and its own suggestions — so the review is about decisions, not data gathering. For follow-ups specifically, the CRM side of this is covered in the CRM follow-up queue.

Replacing status meetings for small teams

For a team of three to ten, a good exception report can replace most of a daily stand-up. Everyone reads the same message in the team thread; the meeting, if you keep one, covers only the “Needs you” items and anything contentious. When several agents and people share a request, the orchestrator posts progress into the same thread so the report and the work live together — see how multi-agent orchestration works.

Don’t replace conversations that are really about people — workload, morale, a hard client — with a report. Those still need a human conversation.

In Cortex, agents report back in the chat thread where you gave the instruction, so the daily report arrives in Telegram, WhatsApp, Slack or Discord rather than in another dashboard. The Sessions page shows how reports and hand-offs appear in the thread.