Skip to main content
Audact
See the offer

Sector · Public sector

Public sector: which citizen interactions count as high-risk, and which do not.

Enquiries · 12-24 month procurement cycle

Not every public-sector deployment of AI is high-risk, and treating them all as if they were is as unhelpful as treating none of them that way. Annex III of the AI Act lists the cases, among them AI used to evaluate a person's eligibility for essential public assistance benefits and services. An information line about opening hours is not that. Where a deployment does fall inside Annex III, a public authority as deployer carries an extra duty that private deployers do not: the fundamental rights impact assessment in Article 27.

What the law requires

A person interacting with an AI system must be told so, unless it is obvious from the circumstances.
Regulation (EU) 2024/1689, Art. 50(1)
AI systems used by or on behalf of public authorities to evaluate eligibility for essential public assistance benefits and services are high-risk.
Regulation (EU) 2024/1689, Annex III, point 5(a)
Deployers of high-risk systems must use them in accordance with the instructions, assign human oversight to competent people, and keep the logs the system generates.
Regulation (EU) 2024/1689, Art. 26
A deployer that is a body governed by public law must carry out a fundamental rights impact assessment before putting an Annex III high-risk system into use.
Regulation (EU) 2024/1689, Art. 27(1)
Each Member State must have had at least one national AI regulatory sandbox operational by 2 August 2026.
Regulation (EU) 2024/1689, Art. 57(1)

What Audact does about it

The Article 50 disclosure spoken in the language of the caller. Germany: national policy layer active. Netherlands: in development. Belgium and Austria: planned.
In controlled testing
A receipt is visible in the product for governed calls: input hash, output hash, the link to the receipt before it, and Merkle inclusion. Three of those anyone can recompute without a key; the fourth, inclusion in the Merkle root, is checked against the published root with the inclusion proof from the call's Article 50 report.
Visible in product · Working today
Primary application hosting: Frankfurt. The subprocessor list and the transfer basis are in the Trust Center.
Working today
The national policy layer on top of Article 50.
Germany: active · Netherlands in progress
Early Customer Program intake.
Netherlands and Flanders (BE): intake open. Germany and Austria: intake is a conversation and not a start date.
Structured inputs for a fundamental rights impact assessment and for a sandbox application.
In development
Alignment with the EUDI Wallet for citizen-identity attestation.
Under research

Classification first

The first question is not which controls you need. It is which list you are on.

A phone line that tells a citizen when the counter is open, books an appointment and routes the rest to a person is an ordinary AI system with a transparency duty under Article 50. A system that helps decide whether someone qualifies for a benefit sits in Annex III point 5(a) and brings the whole high-risk regime with it. The distance between those two is enormous, and it is decided by what the system is used for rather than by how it is built.

That is worth settling before a procurement document is written, because the answer changes what has to be in it. Getting it wrong in the cautious direction is expensive; getting it wrong in the other direction is worse.

Two duties that only public bodies carry

The impact assessment, and the sandbox that is open to you.

Article 27(1) puts the fundamental rights impact assessment on deployers that are bodies governed by public law, before an Annex III high-risk system goes into use. It is a different instrument from the provider's conformity assessment and from a data protection impact assessment, and it asks about the people affected: who they are, what could go wrong for them, and what oversight exists when it does.

The other route is Article 57. Every Member State had to have at least one national AI regulatory sandbox operational by 2 August 2026, a supervised environment for testing with the competent authority involved. For a public authority it is often the cheapest way to find out what its own supervisor expects, and the exit report it produces travels with the deployment afterwards.

Questions

Is a citizen information line high-risk under the AI Act?
Not by itself. Annex III lists the high-risk cases, including AI used to evaluate eligibility for essential public assistance benefits and services. Answering questions, booking appointments and routing to a person is not that, although Article 50 still requires that the citizen is told they are speaking to an AI system. This page is general information, not legal advice.
Who has to do the fundamental rights impact assessment?
The deployer, where the deployer is a body governed by public law and the system is high-risk under Annex III, before putting it into use (Article 27(1)). It is separate from the provider's conformity assessment and from a data protection impact assessment. This page is general information, not legal advice.
Where is the data processed?
Primary application hosting is Frankfurt. The full subprocessor list, the processing locations and the transfer basis are set out in the Trust Center, which is the document to put in a procurement file rather than a sentence on a marketing page.

A record that survives the procurement file

What a public authority has to be able to account for later is what happened in a specific interaction. Audact records that per call and shows the receipt in the product; production export worked when it was retested on 26 August 2026 after the deploy (measured on browser calls rather than over a telephone line).

See a receipt

Related