AI agent for metrics and reporting
Most meetings start by finding the numbers, and the decision gets whatever time is left over. The agent takes the assembling: it pulls the figures from the systems that hold them, marks what moved and by how much, and hands you a draft where any number opens up to its source. Why it moved is your line.
What a metrics agent does
- Pulls the figures from wherever they live — exports, spreadsheets, dashboards — on a schedule rather than the night before.
- Compares the period against the one before it and marks what moved and by how much.
- Drafts the report in the format your meeting already uses: movement, variance, open questions.
- Attaches the source to every number, so it can be opened up the moment someone asks.
- Checks the same metric across systems and surfaces the ones that don't agree.
Processes that run themselves
A playbook isn't a button — it's a process with a condition that starts it. These three get set up most often.
The reporting period closes
The reporting week closes
- Collect the exports for the period as they stand at close
- Recalculate each metric and compare it against the previous period
- Lift out anything past the threshold you set into its own list
- Hand over the draft with the conclusions section left empty
A metric crosses a threshold
A figure crosses a limit you set
- Check the export first, in case the movement is a fault in the data
- Record the fact, the period and the size of the move
- Send it to you with the source attached and no theory attached
Two systems disagree
The same figure differs between two systems
- Put both values side by side, with their sources and periods
- Check the obvious causes: different filters, different period boundaries, a refresh that hadn't run
- Report the discrepancy as a discrepancy, without picking a winner
Where the agent stops
- It doesn't explain causes: the variance is flagged, the reason is yours — a guess at causation is worse than no guess.
- A number without a traceable source isn't published: where an export didn't reconcile, the report carries a gap rather than an estimate.
- Where two systems disagree it reports the discrepancy instead of deciding which one is right.
- A forecast is only ever presented as an assumption, with the conditions it rests on, and never as a fact.
FAQ
Will it tell me why a metric dropped?
No, and that's deliberate. It records the fact, the period and the size of the drop, checks the data isn't at fault, and stops there. The reason needs context it doesn't have: what you shipped, what changed for customers, what happened in the market. In the meeting you're the one who answers for that explanation, which makes an invented one worse than an empty line.
Where does it get the numbers, and do internal systems have to be connected?
The default is the exports and spreadsheets you already have access to, and that's enough to start with. Connecting internal systems is a separate conversation with your security people rather than a way around them.
What happens when two systems disagree?
You get both values side by side, with sources and periods, plus a check on the obvious causes: different filters, a different period boundary, a refresh that never ran. If the gap survives that, it goes into the report as a gap. Deciding which system is right isn't the agent's call.
Can it produce a forecast?
It can, but only with the assumptions written out: which data, at what rate held steady, from which period. The forecast sits in its own block and is never mixed in with the actuals.
Where this shows up
Walkthroughs where this role carries part of the work.