# Reference note: how public administrations govern AI (2026)

Last checked: **18 November 2026**.

This note supports the [public-service track](../tracks/en/public-sector.md).
The playbook is international: no single country's rules are the default.
The note compares how several jurisdictions govern AI in their own
administrations, then gives a checklist to build the same profile for yours.
It describes rules and institutions, not measured effects. For time and quality
evidence, see the [task-time evidence register](../docs/task-time-evidence.md)
and [AI in public administration and planning](public-sector-planning-ai-cases.md).

## Instruments that recur across jurisdictions

Most administrations combine some of the same five instruments. Their legal
force differs, so record which one applies before relying on it.

| Instrument | What it does | Examples below |
| --- | --- | --- |
| Strategy or policy | Sets goals, roles and principles for the administration | Swiss federal AI strategy; US OMB memorandum |
| Inventory or public register | Lists AI or algorithmic systems in use | UK ATRS; Dutch algorithm register; US use-case inventory |
| Impact assessment before use | Scores risk and sets matching controls | Canada AIA; EU fundamental rights impact assessment |
| Audit or oversight guidance | Tells auditors and managers what to check | Swiss Federal Audit Office guide |
| Binding law | Creates enforceable duties and rights | EU AI Act; data protection law in most jurisdictions |

## Jurisdiction profiles

These are snapshots with dates. Rules change, and some pages quoted below
already differ from each other. Check the primary source before a decision.

### European Union

The AI Act gives public bodies specific duties as **deployers** of high-risk
systems. Bodies governed by public law, and private entities providing public
services, must carry out a fundamental rights impact assessment before first
use of most Annex III high-risk systems (Article 27). A public-authority
deployer must also register its use of such a system in the EU database before
putting it into service (Article 49(3)). After the 2026 Omnibus, the Annex III
high-risk rules apply from **2 December 2027**. Transparency duties under
Article 50 apply since 2 August 2026. See the
[Switzerland and EU legal page](../docs/legal-switzerland-eu.md) for dates
and the [fundamental rights impact assessment template](../templates/fundamental-rights-impact-assessment.md).

### Netherlands

The national [algorithm register](https://algoritmes.overheid.nl/en/dashboard)
lists algorithms used by public bodies, including municipalities. Its own
dashboard showed 1,450 published descriptions in June 2026. Publication is not a
general legal duty, and coverage of high-impact systems is still debated in
parliament. The register is a useful model of what a public description can contain.

### United Kingdom

The [Algorithmic Transparency Recording Standard](https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub)
(ATRS) is mandatory for central government departments under a
[scope and exemptions policy](https://www.gov.uk/government/publications/algorithmic-transparency-recording-standard-mandatory-scope-and-exemptions-policy)
of December 2024. Other public bodies are encouraged to use it. Records have
two tiers: a plain summary for the public and technical detail for specialists.
UK public-sector evaluations of office copilots are in the task-time register.

### Canada

The federal [Directive on Automated Decision-Making](https://www.tbs-sct.canada.ca/pol/doc-eng.aspx?id=32592)
applies to systems that fully or partly automate an administrative decision.
A mandatory [Algorithmic Impact Assessment](https://www.canada.ca/en/government/system/digital-government/digital-government-innovations/responsible-use-ai/automated-decision-making/algorithmic-impact-assessment.html)
sets an impact level, and the level determines required controls such as
review, notice and human involvement. Results are published before production.
The latest amendment took effect in June 2025, with a one-year transition for
existing systems. Canada shows how an impact score can drive controls.

### United States (federal)

OMB memorandum [M-25-21](https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf)
(April 2025; the White House copy may no longer be online) keeps agency Chief AI Officers, an annual AI use-case inventory
and minimum risk practices for "high-impact" AI: pre-deployment testing,
impact assessment and monitoring. Its stated emphasis is faster adoption.
Federal policy has kept moving since, so check the current text and any later
executive order before relying on a requirement.

### Switzerland

Switzerland has no general AI law. The revised data protection act applies
directly to AI processing, as the
[FDPIC states](https://www.edoeb.admin.ch/en/ai-and-data-protection).

- **Federal administration**: an AI strategy presented in March 2025. On
  [18 September 2026](https://www.admin.ch/fr/newnsb/uj3obgFHSuVW) (French) the
  Federal Council was informed of a reprioritised plan: six of eight measures
  continue, with a 2026 to 2028 roadmap. On
  [2 October 2026](https://www.admin.ch/fr/newnsb/NJ04UgpFFUoA) (French) it
  adopted a report on automation and AI for administrative efficiency, starting
  with an internal overview of the AI systems the Confederation develops, buys
  or uses. There is no public register comparable to the UK or Dutch ones.
- **Audit**: the Swiss Federal Audit Office published a
  [guide to auditing AI in the federal administration](https://www.efk.admin.ch/en/guide-to-the-audit-of-artificial-intelligence-in-the-federal-administration/).
- **Future binding rules**: the Federal Council of
  [28 October 2026](https://www.admin.ch/fr/newnsb/s2TQiUUjrA5D) (French) plans a
  consultation in spring 2027 on measures to implement the Council of Europe AI
  Convention (transparency, data protection, non-discrimination, supervision,
  risk assessment).
- **Cantons**: practice varies by canton. For example, the canton of Zurich
  approved the [SynerKI project](https://www.zh.ch/bin/zhweb/publish/regierungsratsbeschluss-unterlagen./2026/749/RRB-2026-0749.pdf)
  (German) on 1 July 2026: a shared AI platform for the cantonal
  administration, including for protected data, with one-off spending of
  CHF 2.76 million and CHF 3.87 million a year (figures as
  [reported by Netzwoche](https://www.netzwoche.ch/news/2026-07-10/kanton-zuerich-treibt-ki-einsatz-in-der-verwaltung-voran)).
  This is an investment decision, not evidence of effect.

## What transfers, and what does not

- **Transfers**: the five instruments above, as a checklist. A public register
  record and an impact assessment are useful even where no law requires them.
- **Does not transfer**: legal force, deadlines and thresholds. A requirement
  from one jurisdiction is a good practice elsewhere, not a duty.
- **Not evidence of effect**: strategies, budgets, inventories and register
  counts describe activity. They do not show time saved or better decisions.

## Build the profile for your jurisdiction

Answer each question with a source and a date, and keep the answers with gate
P1 of the public-service track.

1. Which national and local rules govern AI use by your body, and which are binding?
2. Is there a public register or inventory, and must your system appear in it?
3. Is an impact assessment required before use, and who approves it?
4. Which data protection rules apply to automated decisions, notice and human review?
5. Which audit body or ombudsperson can review the system?
6. Which procurement rules or model clauses apply to AI suppliers?
7. Which dates are announced but not yet in force?
