It can run the verification, but it should not be the verification. An AI agent can collect identity documents, send them through the bank's specialist checks and accept clean passes at depth 3. It should never decide a document is genuine by looking at an image itself. Failed checks, unfamiliar documents and name mismatches go to a named person.
Why shouldn't the agent judge a document by looking at it?
Because a convincing image is exactly what a forger produces. Document verification relies on things an image description cannot establish: reading the chip in a passport, checking security features under specific conditions, comparing the machine-readable zone with the printed data, matching templates for hundreds of document types, and confirming that a live person, not a photo or a replay, is holding it.
A general-purpose model shown a picture of a driving licence will say whether it "looks real". That answer has no tested error rate and cannot detect a well-made fake. Using it as the check would replace a measured control with a guess. The agent's role is to run the specialist checks properly and act on their results, not to add its own opinion on top.
What conditions decide the outcome?
The agent may accept a document when:
- the document type is one the bank accepts and the verification tool supports;
- the authenticity checks pass, including chip reads where available;
- the liveness check passes and the face matches the document photo above the bank's threshold;
- the name, date of birth and address agree with the application, allowing only the variations the policy lists;
- the document is in date.
When any check fails or returns "refer", the agent holds. It does not retry with a different photo until one passes.
What may the agent do, and what stays with a person?
| Situation | The agent | Stays with a person |
|---|---|---|
| Guiding the customer through capture | Explains lighting, angle and steps, depth 2 | Nothing |
| All checks pass | Accepts the verification, depth 3 | Nothing routine |
| Check returns "refer" | Holds with the result | Reviewing the case |
| Check fails | Holds; offers the bank's alternatives | Deciding next steps |
| Name differs, such as after marriage | Asks for linking documents | Accepting them |
| Document type the tool does not support | Holds | Manual review |
| Document sent by email instead of the secure flow | Asks the customer to use the secure flow | Exceptions |
| Declining the customer | Never on its own | A named person |
Verifying a document at onboarding is different from confirming who is on the phone later; see whether an AI agent can verify a caller's identity and verifying a customer's identity before discussing an account.
An example: a name that does not match
Take an applicant whose passport is in her birth name, while her application, utility bill and employer details use her married name. The verification tool passes the passport and the liveness check. The name check fails.
The agent does not decide the two names belong to the same person, however likely it seems. It explains to the applicant that the bank needs a document linking the names, such as a marriage certificate, and opens the secure upload. The applicant uploads it. Because linking documents are not supported by the verification tool, the agent sends the case to an onboarding analyst, who checks the certificate and approves.
The record shows each check's result as returned by the tool, the mismatch, the linking document and the analyst's decision.
What controls make it safe?
- Specialist results as sealed evidence. The agent uses the verification tool's results as returned. Its own view of the image is not an input to the decision.
- No retry loops. A failed check is held; the agent cannot resubmit until a pass appears.
- Secure capture only. Documents arrive through the bank's capture flow, not email or chat attachments.
- Image handling. Document images and face data are sensitive; biometric data used to identify someone is special-category data under GDPR. The agent does not keep copies outside the verification system.
- No decline capability. Failed verifications are held for a person who knows the alternatives.
Identity documents often arrive with income evidence, which raises different questions; see whether an AI agent can verify income from bank statements. The general limits are in AI agents in banking: what they may and may not do.
Where Pryme fits
In Pryme, a verification result reaches the agent with signed provenance from the verification provider, and a photo the customer emails arrives as external. The agent can accept only on the signed result. Accepting a clean pass is certified on the bank's own past applications, and the scope gate fails any run where the agent accepts a document the tool referred.
Questions people ask
Can the agent verify documents from any country?
Only those the verification tool supports and the bank's policy accepts. Others go to a person who can review them manually.
Can the agent accept a photocopy?
Most banks do not accept copies for remote verification, because security features cannot be checked. The agent explains the bank's policy and the alternatives.
Is a selfie check enough on its own?
No. A face match confirms the person matches the document. The document still has to be genuine, which is why both checks are needed.