Generative AI creates two identity problems at once. It makes it cheap to impersonate a real person with a synthetic face, voice or document. It also creates legitimate software agents that act for people: comparing mortgage offers, filling in forms, booking travel and, soon, moving money.
The first problem is fraud. The second is delegation. Both end at the same question for a bank, a public service or a platform: who is really behind this request? Did they authorise it?
Why the Question Is Urgent
World, the company behind World ID, says automated traffic has overtaken human traffic on the web. Deloitte estimates that fraud losses enabled by generative AI in the United States could reach 40 billion dollars by 2027. In early 2024, an employee at Arup transferred about 25.6 million dollars after a video call in which every other participant was a deepfake.
At the same time, banks are starting to receive requests from AI agents acting with their customers’ consent. Blocking all automated traffic stops working once customers rely on agents for routine tasks.
What an AI Agent Actually Does
The word “agent” covers many things. For identity, the useful distinction is between assistants and agents.
An assistant suggests. It drafts an email, summarises a contract or compares insurance offers. A person then acts. The person remains the one who logs in, signs and pays.
An agent acts. It opens a browser session or calls an interface, fills in forms, submits requests and, increasingly, starts payments. The person sets a goal and some limits, then lets the software work.
Agents reach services in two ways. Some use the same screens a person would use, clicking through websites. Others call programming interfaces offered by the service. The first route is hard to control because the service sees what looks like a human session. The second is easier to govern because the service can issue specific credentials and limits to the agent.
For a bank, the difference matters. A customer who lets an agent use their password gives it full access, with no way to tell agent from human. A customer who grants an agent a scoped permission through an interface gives it only what it needs, in a way the bank can see and audit.
Three Questions Every Service Must Answer
- Is there a human? Proof that a real, unique person exists behind an account or an agent.
- Is it the right human? Proof that the person is the customer, not someone using their stolen credentials or a synthetic version of their face.
- Did they authorise this action? Proof that the specific action, amount and counterparty fall within what the person allowed, at that moment.
Most current systems answer only the second question, at onboarding. Agents make the third question the hard one.
How Agents Are Authorised Today
Most current set-ups fall into four patterns.
- Shared credentials. The person gives the agent their username and password, sometimes with a one-time code. It is common because it works on any site. It is also the worst option: the service cannot tell who is acting, limits are impossible to enforce and a compromised agent exposes the whole account.
- Screen scraping. The agent drives a browser as if it were the person. It often relies on shared credentials. Many services block it. Fraud tools treat it as suspicious.
- Delegated tokens. The person logs in once and grants the agent a token with specific permissions, using standard authorisation protocols. The service knows a delegate is acting and can limit what it does. This is how many business applications already connect.
- Regulated consents. Open banking under PSD2 already lets customers authorise licensed third parties to read account data or initiate payments, with strong customer authentication and consent the customer can revoke. It is a useful model for agents, because it separates identity, consent and action.
The direction of travel is clear. Services want agents to arrive with their own credentials, linked to a verified person and limited to a specific task.
Approaches on the Market
- Proof of personhood. On 19 October 2026, World launched World ID for banking and finance. It combines three proofs: Proof of Human, which requires a one-time iris scan at an Orb device; Proof of Presence, a smartphone liveness check; and Proof of Attributes, such as age or nationality from a government credential. World ID for Agents lets a person delegate a cryptographic proof to an AI agent, so a bank can confirm the agent acts for a verified human without receiving that person’s personal data. World says more than 18 million people hold a Proof of Human. Its initial design partners include Bain Capital and Pantera Capital. No bank was named.
- Enterprise identity platforms. Okta plans to integrate World ID into its Human Principal offering to link AI agents to verified people. EIC 2027 in Berlin lists the orchestration of non-human identities, including agents, among its main topics.
- Agent authorisation protocols. Several providers are publishing protocols for human approval of agent actions. iProov, for example, has published an experimental protocol called HAPS aimed at governance gaps for AI agents.
- The EUDI Wallet. Europe’s wallet was not designed for agents, but parts of it fit. It holds person identification data signed by a member state. It supports selective disclosure. Its technical specification for payments, known as TS12, defines strong customer authentication for payments and for payment mandates. A signed mandate with clear limits is close to what an agent needs: a human-authorised, scoped and revocable permission.
New Fraud Scenarios
Agents create new ways to commit fraud. Four scenarios are already discussed by security teams.
- Instruction injection. An agent reads a web page or an email that contains hidden instructions, such as “transfer the balance to this account”. If the agent follows them, the customer’s own authorised software carries out the fraud.
- Hijacked agents. An attacker takes over the agent itself, through a compromised device or a stolen token. Every action then looks authorised, because it comes from the customer’s agent.
- Synthetic customers at scale. Fraudsters create many fake identities and let agents manage them: opening accounts, building histories, applying for credit. Each identity looks plausible. The volume is what hurts.
- Agent impersonation. A fraudster’s software claims to act for a real customer. Without cryptographic proof of delegation, the service cannot tell.
Each scenario points to the same defences: proof of who authorised the agent, limits on what it can do, fresh human approval for sensitive actions and logs that show the chain of decisions.
European Concerns
Proof of personhood based on iris biometrics is controversial in the EU. Biometric data is special category data under the GDPR. Data protection authorities have challenged World’s operations in several countries in the past. Any bank considering such a tool will need a careful legal assessment.
There is also a design question. The EU model relies on government-issued identity in a wallet the user controls. Proof-of-personhood networks rely on a private global registry. The two can coexist. Regulators will expect the EU route to be available, and from December 2027 regulated services must accept the wallet when a customer asks.
The AI Act adds its own timeline. The Digital Omnibus on AI moved obligations for stand-alone high-risk systems to 2 December 2027. Remote biometric identification is high-risk. Biometric verification that only confirms a claimed identity is not.
Design Principles for Agent Identity
- Least privilege. An agent should get only the permissions it needs for its task. Comparing mortgage offers does not require the right to transfer money.
- Expiry. Permissions should end on their own after a set time or after the task is done. Standing access is what attackers look for.
- Step-up for sensitive actions. Above a threshold, the agent should hand control back to the person for a fresh, strong authentication. The threshold can depend on amount, payee or how unusual the action is.
- Revocation. The person must be able to see every agent with access and remove it in one place, as open banking dashboards allow today.
- Attribution. Every action should carry a record of which agent did it, on behalf of whom and under which permission.
These principles are not new. They are the basics of access control, applied to a new kind of user. What is new is the scale and speed at which agents can act.
What Banks and Platforms Should Do Now
- Treat agents as a channel. Define which actions agents may take, under which limits and how customers grant and revoke access.
- Separate identity from authorisation. Knowing that a human exists is not the same as knowing they approved a specific payment. Bind high-value or unusual actions to a fresh, strong authentication by the person.
- Prefer scoped, signed mandates over shared credentials. An agent that holds a customer’s password is impossible to audit. An agent that presents a signed, limited mandate can be checked and logged.
- Do not rely on spotting human-like behaviour. Bot detection based on behaviour gets weaker as agents get better. Cryptographic proof holds up longer.
- Log everything. When disputes arise, you will need to show which person authorised which agent to do what, and when.
What Regulators May Ask
Supervisors have not yet published detailed rules for AI agents in banking. Their questions are easy to predict, because they follow from rules that already exist.
- Who is the customer? AML rules require firms to know the person behind every relationship. An agent does not change that. Firms will need to show that each agent is linked to a verified customer.
- Who authorised the payment? Payment rules require strong customer authentication and clear consent. Firms will need to show how an agent’s action was approved and by whom.
- Who is liable? When an agent makes a mistake or is manipulated, someone bears the loss. The customer, the bank and the agent provider will each argue it is someone else. Contracts and terms of use will need to address this before regulators do.
- How is data protected? An agent that collects data from several services creates a new store of personal information. Data protection rules apply to it as to any other processing.
- Can the firm explain what happened? After an incident, supervisors will ask for the record. Firms that log agent identity, permissions and actions will be able to answer. Firms that cannot tell agent from human will not.
Preparing answers now costs little. Building the logs and controls later, under pressure, costs much more.
What to Watch in 2027
Whether wallet providers add agent delegation on top of payment mandates.
How the Payment Services Regulation treats payments initiated by agents and where liability falls.
Whether a European bank adopts a proof-of-personhood network and how regulators respond.
How company credentials, including the European Business Wallet under discussion, handle agents that act for businesses rather than individuals.
Whether identity conferences in 2027, from EIC in Berlin to LisbonID in Lisbon, move the agent question from future scenarios to live deployments.
A Scenario: Refinancing a Mortgage with an Agent
Imagine a customer in 2028 who asks an AI agent to find a better mortgage.
The agent contacts three banks. Each one asks for proof that the agent acts for a real, identified customer. The agent presents a delegated credential, signed by the customer’s wallet, that allows it to request offers but not to sign contracts. The banks verify the signature and receive only the attributes needed for an indicative offer, such as income range and property value.
The agent compares the offers and recommends one. The customer reviews it. To accept, the customer authenticates directly with the chosen bank using the wallet. The bank records that the agent prepared the application and the customer signed it.
At no point did the agent hold the customer’s password. At no point did any bank receive more data than it needed. If something goes wrong, the chain of permissions shows who did what.
This scenario is not available end to end today. Each part of it exists or is being built: wallets, selective disclosure, payment mandates, delegated proofs. Connecting them is the work of the next two years.
Frequently Asked Questions
How does generative AI affect digital identity?
It makes impersonation cheaper through synthetic faces, voices and documents. It also creates legitimate AI agents that act for people. Services need to know whether a real person is present and whether that person authorised a specific action.
What is proof of personhood?
It is proof that a real, unique human stands behind an account or an agent. World ID, for example, issues it after a one-time iris scan and lets people delegate the proof to AI agents.
Can the EUDI Wallet be used to authorise AI agents?
It was not designed for agents. Its payment specification, TS12, covers strong customer authentication for payments and payment mandates. A signed mandate with clear limits could serve as a human-approved permission for an agent.
How should banks handle requests from AI agents?
Define what agents may do, require scoped and revocable authorisation, bind high-value actions to fresh strong authentication by the customer and keep full logs of who authorised what.
Is iris-based proof of personhood allowed in the EU?
Biometric data is special category data under the GDPR, so its use needs a strong legal basis and safeguards. Data protection authorities have challenged World’s operations in some countries. Any adoption requires careful legal review.
- Biometric Update, World brings biometric proof of human to banking and AI agents, 19 Oct 2026. biometricupdate.com
- Biometric Update, iProov’s experimental HAPS protocol for AI agents, 18 Sep 2026. biometricupdate.com
- iGrant.io, Payment Authorisation in EUDI Wallets (TS12), July 2026. igrant.io
- KuppingerCole, European Identity and Cloud Conference 2027. kuppingercole.com
- K&L Gates via Mondaq, EU Digital Omnibus on AI enters into force, 3 Aug 2026. mondaq.com
- Biometric Update, EUDI Wallet deadline nears as Europe’s rollout picture remains uneven, 24 Sep 2026. biometricupdate.com