The World Bank has graded the identity infrastructure governments already run against the questions an agentic transaction actually asks. Authenticating the person who approves an action rates strong. Establishing whose behalf the agent acts on rates partial. Proving that a specific action falls inside the agent’s grant is, in its own words, “to be built”.
What happened
At a webinar on 11 November 2026, “Digital Identity Systems for Agentic AI”, World Bank senior digital specialist Marie Eichholtzer walked through how far existing digital public infrastructure (DPI) gets a government that wants to let an agent act for a citizen or a company. The bank has published the deck. The assessment is not a product pitch: it is a gap analysis of national ID, business registries and representation rules against the questions a service counter — or an API — has to answer before it lets an agent do something.
| Question a service must answer | Current capability |
|---|---|
| Is the person approving this action who they claim to be? | Strong — national ID and authentication systems |
| On whose behalf is the agent acting? | Partial |
| May this person represent this company? | Weak — often administrative records alone |
| Is this specific action within the agent’s authority? | To be built |
The bank frames the missing layer around three legal concepts. Representation is the basis for acting for another person or entity. A mandate is a defined grant of authority. Delegation is passing that authority onward, where permitted. Every grant, Eichholtzer said, should record seven things: the principal, the grantor, the actor, the permitted acts and their limits, whether the action is individual or joint, its validity and status, and whether it can be delegated further.
There are three ways a service can check such a grant: hold the authorisation record itself, query an authoritative external source, or accept portable, digitally signed evidence. Finland’s Suomi.fi e-Authorizations is the example given for the second route — an online service asks, in real time, whether someone may act for another person or organisation. A verifiable credential is the third: the digital equivalent of a sealed letter of authority, checkable even without a live connection to the issuer.
A second strand of the presentation was aimed at builders. Drawing on the OWASP Top 10 for Agentic Applications 2026, published in December 2025, the bank paired three immediate risks with fixes: give agents their own time-limited credentials rather than a citizen’s login; accept only verifiable agents and keep a register of trusted ones; make consent screens plain and require human confirmation for high-risk acts. The principle underneath, in Eichholtzer’s phrasing: the AI never polices itself — systems outside the agent supervise and check each action.
The standards vehicle is the International Telecommunication Union’s Focus Group on Trust and Identity for Humans and Agentic AI (FG-TIDA), launched at the AI for Good summit on 9 July 2026 and parented by ITU-T Study Group 17, which writes the ITU’s security and trust standards including X.509. FG-TIDA is explicitly pre-standardisation: it does not issue binding standards, and its reports are handed to SG17 or other bodies. Its first in-person meeting runs 30 November to 1 December 2026 at UNESCO in Paris, hosted by the French Ministry of the Interior; it has asked governments for one-page public-service use cases. Further meetings are planned for Geneva in January 2027 and Dubai in May 2027.
Why it matters for KYA
KYC answers “who is this customer”. It has never had to answer “may this customer’s software do this, now, for this amount, to this account” — and that is the question every agentic service now has to answer in milliseconds, without a human on the other end. The World Bank’s scorecard is useful because it separates the parts that are already solved from the part that is not. Identity proofing is solved. Representation is half-solved. The authority check — the KYA step — is not.
That authority check has a data model, and the bank’s seven fields are a reasonable minimum: principal, grantor, actor, permitted acts, limits, individual or joint, validity and status, delegability. An organisation that cannot produce those fields for an agent it has deployed cannot answer a regulator, a counterparty or its own auditor. It also cannot revoke cleanly, because revocation needs a record of what was granted and to whom.
The three verification routes matter for architecture. Holding the record locally is fast but only as good as the local copy. Querying an authority is current but needs an interface and a fee model. A portable credential works offline and across borders but needs a trust framework for issuers. Most real deployments will use more than one, and the choice determines what an agent can prove when it is away from home.
The advice for governments — map existing systems, decide who issues and who revokes authority, pilot one transaction from grant to withdrawal — is also the right advice for an enterprise. Pick one agent workflow, write the grant down, run it end to end, then revoke it and check that the revocation actually took effect. That is a smaller project than a governance programme and it produces the evidence a governance programme needs.
There is a limit worth naming. A focus group produces reports, not requirements, and the ITU’s consensus process is slow by design. The bank’s own position is not to wait for it: most countries already have the registries and the representation rules, and the missing piece is a decision about who is allowed to say yes on someone’s behalf.
What to watch
- Whether FG-TIDA’s use cases turn into an SG17 work item with a draft recommendation attached, or stay a discussion forum. The group’s output list will show which.
- Whether the “register of trusted agents” the bank recommends becomes a government-run list, an industry one, or a per-service allowlist. Each has different failure modes.
- Whether DPI vendors build revocation as an API rather than an administrative process. A grant that takes a week to withdraw is not a grant an agent should hold.
- Whether the authority check gets a standard data model before the payment networks and platforms each ship their own.
Sources
- World Bank, “Digital Identity Systems for Agentic AI”, webinar deck, 11 November 2026 — thedocs.worldbank.org
- ITU-T, Focus Group on Trust and Identity for Humans and Agentic AI (FG-TIDA) — itu.int
- ITU-T, FG-TIDA Terms of Reference — itu.int
- ITU FG-TIDA community, kick-off meeting notice, 30 November–1 December 2026, Paris — tida.community
- Biometric Update, “World Bank, ITU: Digital ID cannot yet establish what AI agents are authorized to do”, 16 November 2026 — biometricupdate.com
- OWASP, “Top 10 for Agentic Applications 2026”, December 2025 — genai.owasp.org
Know Your Agent (KYA) explains agent identity, verification and accountability. This is an explainer, not legal or compliance advice — see our Sources & methodology. New to KYA? Start with What is Know Your Agent? and the glossary.