Skip to content
Bernhard GötzendorferAI automation that runs on its own · Vienna

StatAgent

A statistic anyone can recompute, and that stops when data is missing.

For developers · Figures from practice data. Open source, MIT. · github.com (opens another website)

StatAgent: 24 months computed, 22 published, 2 held back

What you do with it

You look at the chart across all the months. One of the months is marked in yellow rather than shown as a sudden drop: its delivery was missing employment records, so StatAgent held it back and wrote a query. In the next month's delivery, almost all of the missing records are back. The slump was never real, and it was never published.

Who it's for
For public bodies and statistics offices that publish figures from large monthly data deliveries and want anyone to be able to check the sums.
Where it stands
Built at a competition on practice data, not in use at any public body. I have switched the online demo off; the code and a demo video are public.

Step by step

  1. Check the delivery

    Every month a large table of all employment records arrives. The program reads it in and checks it against 24 fixed rules.

  2. Count at the reference date

    If the delivery adds up, the program takes the count as of the reference date and writes statistics and a quality report. No language model does the maths.

  3. Stop rather than guess

    If something is missing, StatAgent holds the month back and writes a query instead of publishing a wrong figure.

  4. Sign off a new rule

    When a rule changes, an agent proposes the new one and the program runs both. The new rule only takes effect once a named person signs it off.

How it works

No language model
in the monthly run. When a rule changes, an agent only drafts the proposal.
The code
checks every delivery against 24 rules, takes the count as of the reference date and stops when a delivery doesn't add up.

Why it exists

One monthly delivery in the practice data was missing one in five employment records. Published as it was, that would have looked like a slump in the labour market.

The proof

  • 22 of 24months
  • 2held back

Built at the KI Challenge 2026, 5 and 6 November 2026.

The slides

  • Slide “A wrong figure costs trust”: in October 2025, 78,949 employment records are missing from the monthly delivery, one in five. StatAgent does not publish, stops and writes a query.
  • Slide “We recompute the organiser’s report”: every figure in the report matches our own computation, no deviation. For the other months, 24 checking rules keep watch.
  • Slide “Errors found, nothing deleted”: birth dates in a 13th month, people aged well over a hundred, duplicate rows. Each error is corrected by a rule, or StatAgent stops and asks.
  • Slide on the core: every figure has evidence. Only a person may approve. The program computes, the AI only suggests. 24 monthly deliveries.
Jury slides from the KI Challenge 2026, in German. Figures from the practice data

Further reading

Part of what you get

This product is an example of:

Other things I have built