Campaign reporting
Weekly reporting is the purest agent job in marketing: the numbers are in a system, the format never changes, and the only judgement is which movements are real. Automate the assembly, keep the interpretation honest by making it cite the numbers.
The job on one screen
| Runs when | Monday morning, for the previous week. |
|---|---|
| Reads | Spend, impressions, clicks, conversions, and the prior four week baselines. |
| Decides | Which changes are outside normal variation and worth writing about. |
| Produces | A short report: what moved, by how much, and the first thing to check. |
| Stops when | Report posted. It never changes budgets. |
| Tools it needs | Analytics API, ad platform APIs, chat webhook. |
| Autonomy to start at | Auto, once it has a logged track record. Start lower. |
Optional: load a real model
Run it
What it watches, and how far is too far
Tolerances are the whole design. Too tight and everyone mutes it, too loose and it misses the thing you built it for.
| Signal | Baseline | Tolerance | First thing to check |
|---|---|---|---|
| Cost per acquisition | 42 | 25% | Check whether a high-intent audience ran out of budget mid-week before blaming creative. |
| Click-through rate, per mille | 18 | 30% | Creative fatigue shows here first. Compare frequency against last week. |
| Conversion rate, per mille | 26 | 30% | Check the landing page and the form before touching the campaign. |
| Spend | 12000 | 20% | An automated bid strategy changed, or a campaign left a learning phase. |
Where this one goes wrong
Narrating noise
A 3% move in a small segment is not a story. Tolerances stop the agent writing fiction about randomness.
Correlation as cause
The model will happily explain why CPA rose. Restrict it to first things to check, never conclusions.
Reports nobody reads
If it is longer than a screen it gets skipped. Three movements maximum, ranked.
How you would know it is working
| Measure | Why that one |
|---|---|
| Actions taken per report | The only measure that matters. A report that changes nothing should not exist. |
| False alarms | Movements called out that turned out to be noise. Tune tolerances against this. |
| Time to report | The saving, usually a half day a week per channel. |
Earning more rope
| Assist | Draft the report for a marketer to check and send. |
| Approve | Post automatically to the team channel, a human adds interpretation. |
| Auto | Fully automatic weekly, with budget changes always human. Reporting is read-only, so this is safe. |
Read-only agents deserve more autonomy than write agents. This one can run unattended for years, because the worst case is a paragraph nobody agrees with.
Related: Anomaly explainer · Ad copy · Testing agents · all agent jobs · Agent Lab home
Related: Anomaly explainer · Ad copy · Testing agents · all agent jobs · Agent Lab home
Free from AI School - no signup, everything runs in your browser.