AI agents in government services

People deal with government for a short list of reasons: taxes, licences, permits, benefits and fines. Each involves identity checks, forms and waiting, and each has rules about who may act for someone else. An agent acting for a person could file, check status and book appointments with an agency’s agent, but only inside the delegation schemes governments already run, and with agencies’ own AI rules governing the other side.

Where agents meet today

Portals, forms and phone lines

Government services are delivered through online portals, paper and PDF forms, counters and phone lines. For US federal systems, NIST SP 800-63-4, the Digital Identity Guidelines finalized in July 2025, sets requirements for identity proofing, authenticators and federation, expressed as identity, authenticator and federation assurance levels. It superseded SP 800-63-3.

Identity on each side

An agency answering a person’s agent needs two kinds of proof. The first is that the person is who they claim to be, at the assurance level the service requires, which NIST SP 800-63-4 describes as identity and authenticator assurance levels. The second is that this software is acting for that person, within limits the person set. Portals answer the first with a sign-in. Outside formal representative schemes like the ones below, they have no way to express the second, so software acting for a person has to use that person’s own sign-in.

Delegation schemes that already exist

Tax agencies show how acting for someone else already works:

Scheme	What it grants
IRS Form 2848	Authorizes an eligible individual to represent a taxpayer before the IRS
IRS Form 8821	Authorizes a person to receive the taxpayer’s confidential tax information
CRA representative authorization, level 1 (individuals)	Access to information only
CRA representative authorization, level 2 (individuals)	Access to information and certain account changes

The IRS lets both forms be submitted online. The CRA offers three levels for business accounts, with level 3 able to authorize other representatives. These schemes are the closest thing government has to a scoped, revocable delegation for an agent: they state who may act, for whom, and how far.

What an agent exchange would add

Illustrative: a person’s agent renews a professional licence.

1. The agent asks the licensing body’s agent which documents and fees apply to this licence type.
2. The agency’s agent returns a checklist as structured data, with the renewal deadline.
3. The person’s agent submits the documents under the person’s own sign-in or an authorized representative’s access.
4. The agency’s agent returns a confirmation number and the expected processing time, and later a status update when a reviewer decides.

Every step exists today on a portal. The difference is that the person’s agent can do it without screen-scraping a form, and the agency receives complete, validated submissions.

Top use cases

Use case	Who talks to whom	Notes
Tax, licences, permits and benefits	Person’s agent and agency agent	Status checks, appointment booking, document requests
Parking, EV charging and tolls	Car or driver agent and municipal or toll authority agents	Parking permits, toll accounts, fines
Procurement from RFQ to PO	Supplier agent and agency buyer agent	Bid questions and order acknowledgments
When to keep a human	Either side and a person	Eligibility disputes, appeals, vulnerable applicants

Status checks are the natural first step. Application status, appointment availability and document checklists are read-only and change nothing on the person’s file.

Regulatory considerations

The rules below govern how agencies use AI and identity. Checked on 26 September 2026. This is not legal advice.

United States (federal)

• OMB M-25-21 (3 April 2025). Replaced M-24-10 as the policy for federal agencies’ use of AI. It directs agencies to apply minimum risk management practices to “high-impact AI”: AI whose output is a principal basis for decisions with a legal, material, binding or significant effect on rights or safety, including a person’s access to critical government resources or services. An agency agent whose output decides benefit eligibility would fit that definition. A companion memo, M-25-22, covers AI acquisition, and M-26-04 (11 December 2025) sets unbiased AI principles. The OMB memoranda index showed no replacement of M-25-21 when checked.
• Digital identity. NIST SP 800-63-4 sets the assurance levels agencies use for identity proofing and sign-in.
• Representation. The IRS representation and disclosure forms above define who may act or receive information for a taxpayer.

State and local agencies, which run motor vehicle, licensing and many benefit services, have their own rules.

Canada

• Directive on Automated Decision-Making (Treasury Board). Applies to automated decision systems in production that make, or make an assessment related to, an administrative decision about a client. Departments must publish an Algorithmic Impact Assessment before production, give notice through all service channels that a decision is automated or assisted, keep human involvement (for the two highest impact levels a human must make the final decision), and explain decisions to clients. The current version’s page was last modified on 24 June 2025.
• CRA representative authorization. Defines what a representative may see and change, by level.
• Ontario. The Strengthening Cyber Security and Building Trust in the Public Sector Act, 2024 received royal assent on 25 November 2024 and enacted the Enhancing Digital Security and Trust Act, 2024. Ontario says regulations under it on cyber security (O. Reg. 51/26) and on digital technology affecting people under 18 (O. Reg. 52/26) take effect from 1 July 2026, and points ministries and provincial agencies to its Responsible Use of AI Directive for AI requirements.

Other provinces and municipalities set their own rules for the services they deliver.

Where to start

For an agency:

1. Publish read-only answers first. Office hours, appointment availability, application status and document checklists.
2. Reuse the delegation you already run. Map an agent’s access to an existing representative level instead of inventing a new one, and log which authorization each request relied on.
3. Do the impact assessment before production. In Canada the directive requires it; in the US, check whether the use is high-impact under M-25-21.
4. Keep appeals with people. Eligibility denials and enforcement actions need a person with the full record.

For a person’s agent: work through the authorization the person has actually granted, keep a record of every submission and confirmation number, and stop at anything that needs the person’s signature. Delegated authority covers the pattern.