Find the report that repeats
Start with the report you rebuild most often. Write down the steps, the files, and the people who ask for it. If the same three people ask every week, the tool has a real user.
Skip the analysis you do once a quarter. The value of a small tool comes from repetition, and a quarterly job does not repeat enough to justify one.
Separate the data from the judgment
List the inputs, the transformations, and the judgment calls. Transformations are repeatable: join, filter, aggregate, format. Judgment is not: which outlier to exclude, how to frame a miss, what to recommend.
The tool should automate the transformations and leave the judgment visible. A dashboard that hides the judgment is worse than a spreadsheet that shows it.
Describe the checks, not only the output
Every recurring analysis has checks: totals reconcile, step changes are explained, missing data is flagged. Write those checks down before the first build, because they are the part that prevents the wrong number from being presented.
Ask for a visible check status on every refresh. A report that says "three checks passed, one series is stale" is more useful than a clean-looking chart.
Hand off with the sources attached
The handoff is where analyst work usually breaks. Keep the source files, the transformation steps, and the draft in the same place so the next person can rerun it rather than reconstruct it.
Write a two-paragraph handoff: what the tool does, and what still needs a human. The second paragraph is the one that prevents over-trust.
Where this breaks
It breaks when the source data is owned by another team and changes without notice, when the analysis depends on one person's judgment, and when the tool is expected to replace the analysis rather than speed it up.
Name the owner of each source and the person who reviews the output. Two names and a check schedule make the tool maintainable.
How to run it on Sigla
Keep the source files, the checks, and the report in one workspace. Describe the transformation steps and the checks, then let AI build the first version and run it on last week's data. Ask for changes by naming the step that took the most time.
For the general path, read build an internal tool without coding. If the recurring work is research, the AI deep research tool shows the evidence habit behind it.
FAQ
Where should the data live?
Where it already lives, with the tool reading from it. A tool that becomes a second copy of the data creates a reconciliation problem you did not have before.
Can the app do the analysis for me?
It can run the steps you describe and show the checks. The judgment calls still need you, and they should stay visible in the output.
What if the source changes format?
Add a check that detects the change and fails loudly. Silent format changes are the most common reason a recurring report goes wrong.
How do I get colleagues to use it?
Make the handoff explicit: what it does, what still needs a person, and who to ask when a check fails. A tool with a named reviewer gets used.
Next step
Pick the report you rebuilt last week, list its checks, and describe only the transformations before you ask for a screen.