KYC queues are long because of evidence, not judgement. Most of an analyst's time goes on chasing documents, reading them, extracting details, matching them against records and screening names. The risk judgement at the end takes a fraction of that time. Agents can carry the gathering and checking. The judgement about the customer's risk stays with a person.
Where does the time in KYC actually go?
Know your customer is the work of establishing who a customer is, understanding the nature of their business or activity, and assessing the risk they present. Under the FATF Recommendations, the international AML standard, this is applied on a risk-based approach through customer due diligence.
On paper, the decision is the centre of the process. In practice, the decision comes at the end of a long trail of collection and checking. For a business customer, that trail can include certificates of incorporation, registers of directors and shareholders, ownership charts through several layers of companies, identity documents for each beneficial owner, proof of address, source of funds explanations, and screening of every name against sanctions, politically exposed person and adverse media lists.
Each item has to be requested, received, read, checked and filed. Many arrive incomplete or out of date and have to be chased. That is the bottleneck.
What can an agent carry?
The evidence work is repetitive, rule-bound and document-heavy. It is the part of KYC where agents help most.
| KYC task | What the agent does | Depth |
|---|---|---|
| Requesting documents | Works out what is needed for this customer type and risk, and requests it | 3 |
| Chasing missing items | Sends reminders and tracks outstanding requests | 3 |
| Reading and extracting | Pulls names, dates, numbers and addresses from documents into structured fields | 2 |
| Checking against strong sources | Compares extracted details with sealed records such as official registers | 3 |
| Mapping ownership | Builds the ownership chart and identifies who meets the beneficial ownership threshold | 2 |
| Screening | Runs names through screening and gathers context on potential matches | 2 |
| Summarising | Drafts a case summary with every fact, its source and any discrepancy | 2 |
The question of whether an AI agent can carry out KYC checks is answered by this table: it can carry out the checks, and it presents the results.
Why does the judgement stay with a person?
Three parts of KYC involve judgement that an agent should not make.
Risk rating. Deciding that a customer is low, medium or high risk shapes every later control on the relationship: monitoring intensity, review frequency, product access. It weighs factors the policy describes but cannot fully specify.
Resolving potential matches. A screening hit may be the customer or someone with a similar name. Deciding which, especially for politically exposed persons and sanctions, is a judgement with serious consequences either way.
Accepting or declining the customer. Onboarding a high-risk customer, or refusing a customer, is a decision the bank must own. It also affects the person's access to banking.
These decisions sit with trained analysts and, for higher risk, with senior approvers. The agent's job is to put a complete, checked file in front of them.
What does a strong evidence file look like?
The value of the agent's work depends on what reaches the analyst. A good file does four things.
- Shows every fact with its source and class. A director's name from an official register is sealed. A source of funds explanation from the customer is declared. The analyst should see the difference at a glance.
- Lists every discrepancy. Different dates of birth across documents, an address that does not match the register, a shareholder missing from the chart.
- States what is missing. Anything requested and not received, and anything the policy requires that could not be checked.
- Separates fact from inference. "The register lists three directors" is a fact. "The ownership structure appears designed to obscure control" is an inference, and should be marked as one.
The KYC refresh checklist for agent-assisted reviews applies the same structure to periodic reviews.
How does this change the analyst's day?
The analyst stops being a document chaser and becomes a reviewer. Their time goes on discrepancies, potential matches and risk judgement, which is where their expertise is needed.
Throughput rises because the evidence arrives complete, not because the analyst works faster. Quality can rise too, because a person who is not exhausted by collection has more attention for the parts that need it. The risk to manage is the familiar one: an analyst presented with a polished summary may skim it. Override and challenge rates are worth watching.
Periodic refresh is where the gains often show first, because the volume is large and most reviews find little change; see KYC refresh with an agent: what stays with a person.
Where Pryme fits
In Pryme, KYC evidence work runs as separate capabilities, each certified on your own cases: document requests and chasing can earn depth 3, extraction and summary sit at depth 2. Every fact carries its provenance class into the analyst's view. Risk rating and onboarding decisions are held for a named analyst or approver. See AI agents in banking: what they may and may not do.
Questions people ask
Can an agent onboard low-risk customers automatically?
A bank may choose to let simplified checks for clearly low-risk products run end to end under a published rule. The rule is the bank's decision; the agent applies it. Anything outside the rule goes to a person.
Does the agent decide what documents to ask for?
It applies the bank's policy to the customer type and risk indicators. The policy itself is set and owned by the bank's financial crime function.
Can an agent dismiss false screening matches?
It can gather the evidence that a match is false and recommend dismissal. The decision to dismiss, especially for sanctions, should be made and recorded by a person.