Report assembly
Recurring reports eat analyst weeks. The numbers should come from certified queries that a human wrote once, and the model should only write the sentences around them. Generating the SQL fresh each month is how a report quietly changes definition.
The job on one screen
| Runs when | A schedule, usually the first working day of a month or week. |
|---|---|
| Reads | Certified queries, the previous report, and the calendar of things that happened. |
| Decides | Which movements to write about, and how to say them without overclaiming. |
| Produces | A finished document with charts, numbers, and three paragraphs of commentary. |
| Stops when | Assembled and checked. A human sends it to the audience. |
| Tools it needs | Warehouse (read only), chart renderer, document API. |
| Autonomy to start at | Approve. It does the work and stops before the irreversible step. |
Optional: load a real model
Run it
The checks the draft has to pass
The model writes, then code checks. A draft that fails any check never reaches the send step, and the failure is shown rather than silently patched.
| Check | What it protects against |
|---|---|
| No causal assertion | Because and driven by are the words that turn a report into a rumour. Say what moved and what to check. |
| Numbers present | Commentary with no figures is opinion. Every claim carries the number it came from. |
| Both directions covered | A report that only mentions good movements gets stopped being trusted the first time someone checks. |
| Under 120 words | Nobody reads paragraph four of a monthly report, so the important sentence must be in paragraph one. |
Where this one goes wrong
Regenerated definitions
If the SQL is generated each month, your metric changes silently. Certified queries only.
Cheerful reporting
Models write optimistically by default. Force a both-directions check or the bad news gets softened out.
Chart lies
Truncated axes and rebased series flatter results. Fix the chart config once and never let it be generated.
How you would know it is working
| Measure | Why that one |
|---|---|
| Time to publish | Usually the point: days to hours. |
| Commentary edited before send | Falling edits mean the grounding rules are working. |
| Questions the report failed to answer | Collect them and they become next month's certified queries. |
Earning more rope
| Assist | Assemble numbers and charts, human writes the commentary. |
| Approve | Assemble everything with drafted commentary, human approves and sends. Right level. |
| Auto | Fully automatic for internal operational reports where the audience knows how it was made. |
Related: Ad hoc questions · Campaign reporting · Budget tracing · all agent jobs · Agent Lab home
Free from AI School - no signup, everything runs in your browser.