A modern statistical office
The assessment says where an office stands and the cookbooks say what to do next. This page shows the destination: one fictional national statistical organization, the example organization of the cookbooks, drawn at three maturity levels. Choose a level to redraw the office, open a unit to see who works there and with what, follow a release day hour by hour, and switch the AI layer off to see what still stands.
Models run as stages of production, every output is verified or reviewed before it reaches a person outside the office, and the office publishes its own statement of model use with each release.
- Units
- 5
- AI components at this level
- 11
- Steps in a release day
- 10
- Cookbooks behind it
- 11
The five units of the office
Each unit opens to its people, its tools, and its AI components at the chosen level. Every component names the person who decides on its output.
Methodology and production floor
The nightly batch codes the quarter's open responses on the office's own server and resumes where it stopped. Coders see only the routed share. Editors receive value proposals that already passed every rule. The anomaly detector flags a jump and the analyst reads a drafted explanation with its evidence before deciding.
People and skills
- Survey methodologists
- Coding supervisor and coders
- Editors
- Subject-matter analysts
On their screens
- Model server inside the production network
- Editing system with rule export and proposal queue
- Anomaly detection pipeline
- Evaluation suites per task
AI components and who decides
- Occupation and industry coder codes responses above the threshold; routes the rest.
Decision: coders decide the routed share; a re-coded sample measures the rest. - Value proposer proposes a value with a reason, checked against every edit rule.
Decision: the editor accepts, changes, or rejects. - Anomaly explainer drafts the cause of a flagged value with cited evidence.
Decision: the analyst records a verdict.
Assessment dimensions this unit answers for
One release day of the labour force survey
One quarterly release passes through the five units. Step through it, or play it, and watch where a model works, where a check runs, and where a person decides.
The nightly batch coded the quarter's 250,000 open responses on the office's own server, resuming after an interruption at record 90,000. Seventy percent were accepted above the threshold of 0.80; the rest are queued for coders.
Running models in production →People, compute, and tools at the AI-native level
Illustrative ranges for three sizes of office. The skills column matters more than the counts, and the compute figures come from the sizing recipe of the small models cookbook.
| Role | Skills | Medium office |
|---|---|---|
| Methodologists with evaluation skills | Survey methods; designing labelled sets, thresholds, and measures | 3 to 5 |
| Data curators | Metadata standards (DDI, SDMX); the Metadata Editor; vocabulary work | 4 to 8 |
| Data engineers | Python, SQL, pipelines, version control, CI | 2 to 4 |
| Data scientists | Model serving and adaptation, evaluation suites, synthetic data | 3 to 6 |
| Web and API developers | OpenAPI or SDMX, schema.org, the MCP SDK | 3 to 5 |
| Systems and security staff | GPU servers, access control, monitoring, model-file verification | 3 to 5 |
| AI-use policy owner | Governance, the register, the statement of model use; usually part of an existing role | 1 |
| Coders and editors | Unchanged in number at first; their work shifts to the routed share and the re-coded sample | as today |
Counts are full-time equivalents and illustrative. The first three roles can be one person each in a small office; the skills matter more than the headcount, and the cookbooks are written as the training path.
How an office gets to this picture
- 1
Find the office in the picture
Set the level switch to where the office is today. Each unit's entry names the assessment dimensions it answers for; the readiness profile says which gaps are largest.
The assessment framework → - 2
Read what the next level adds
Every unit entry ends with what the next level adds. The operational companion lists, per dimension, the tools, recipes, and evidence that get there.
Operationalizing the assessment → - 3
Build it recipe by recipe
Each AI component on this page is a chapter of a cookbook, with scripts, a running example, and the check that gates it. The cookbooks are also the training path for the roles in the resources table.
The cookbooks →
The office on this page is a composite. No single organization runs every component shown, the figures are ranges for planning, and the people are roles. What is exact is the set of controls: every AI component has a decision point, a suite, a version, and a fallback, and the data layer stands without the AI layer.