Chasing paperwork is the part of KYC an agent can take almost entirely. It finds what is missing or expired on each file, sends the request and reminders on an approved schedule, checks that what comes back is legible, current and the right type, and passes it to an analyst. Verifying the document and restricting an account stay with people.
What is the process, and who owns it?
Customer files go incomplete for ordinary reasons: a passport expires, an onboarding exception was granted on a promise to supply proof of address later, a periodic or trigger-event KYC review asks for something new. Under a risk-based approach, the firm must keep its customer due diligence up to date, and an incomplete file is a finding waiting to happen.
The KYC operations team usually owns the chase, with a team lead accountable for overdue files and the MLRO setting the policy on what happens when customers do not respond. In most firms the work is repetitive and badly tracked: emails sent from personal inboxes, reminders forgotten, uploads sitting unread.
What does the agent prepare, and what does a person approve?
- Agent, depth 1: finds the gaps. Reads each customer file against the document requirements for that customer's type and risk rating, and lists what is missing, expired or due to expire.
- Person, KYC lead: approves the request templates and the chase schedule. Once, not per customer. The schedule sets how many reminders, at what interval, through which channels.
- Agent, depth 3, within bounds: sends the requests and reminders. Using only approved templates, only through the customer's verified contact details, and only on the approved schedule. Anything outside that, such as a customer who replies with a question, goes to a person.
- Agent, depth 1: checks what comes back. Right document type, legible, in date, name and address match the file, all pages present. Each upload is marked as declared, since the customer supplied it and nobody has verified it yet.
- Person, KYC analyst: verifies. Judges authenticity, accepts or rejects, and records the decision.
- Agent, depth 2: drafts the escalation for customers who have not responded by the end of the schedule, with the history of every contact.
- Person, team lead or MLRO delegate: decides the consequence. A further chase, a restriction, or a move to exit.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Gap finding | Lists missing and expiring items | 1 | Setting the requirements |
| Templates | Nothing | None | Approving wording and schedule |
| Chasing | Sends approved requests and reminders | 3 | Replies that need judgement |
| Upload check | Tests type, legibility, dates, match | 1 | None |
| Verification | Nothing | None | Accepting or rejecting |
| Non-response | Drafts escalation with history | 2 | Restricting or exiting |
Which controls apply, and what evidence does each leave?
Templates and schedule approved once, enforced every time. The agent's sending capability is bounded by the approved templates and the schedule. A request in different wording, or a fourth reminder when three are allowed, is refused. Evidence: the approved templates with their version, and a contact log showing every message matched one.
Verified channels only. Requests go to contact details already verified on the file. A customer who replies from a new email address is not treated as verified. Evidence: each sent message shows the channel and when that channel was verified.
Provenance on uploads. A customer-supplied document is declared until an analyst verifies it; a document pulled from a registry or a trusted issuer may carry a stronger provenance. Evidence: each document's provenance and the analyst who changed it.
No redaction by guesswork. Customers often send more than asked, such as a full bank statement when only the address page was needed. Where your policy requires minimisation, the redaction approach in whether an agent can redact personal data from documents applies, with a person checking.
Consequences decided by people. Restricting an account affects the customer's access to their money. The agent drafts; a named person decides. Evidence: the escalation draft and the decision, filed together.
What should you measure?
- Files complete within the schedule, by document type.
- Uploads rejected at the agent's check versus at verification. Many analyst rejections means the upload check is too loose.
- Customer replies the agent could not handle, and why. These show where templates are unclear.
- Overdue files at month end, the number your MLRO actually watches.
- Complaints about chasing, including tone and frequency. Gathering this is covered in collecting feedback on an AI agent's work.
Where does it usually go wrong?
Chasing for things the file does not need. If the requirements table is wrong, the agent chases efficiently for the wrong documents. Test the gap list on a sample of files before sending a single request.
Messages that look like phishing. A request for identity documents with a link is exactly what fraudsters send. Use channels customers can verify, avoid links where possible, and tell customers how to check the request is genuine.
Vulnerable customers treated as non-responders. A customer in hospital or in difficulty is not ignoring you. The agent should flag known vulnerability markers before escalating, and a person should decide.
Uploads accepted because they are tidy. A clean, legible document is not a genuine one. The agent's check is about completeness; authenticity is the analyst's call.
Chase history scattered. If some reminders go from the agent and some from a person's inbox, the history is incomplete when you need it. Route all contact through one record.
Where this sits within compliance work overall is covered in AI compliance agents.
Questions people ask
Can the agent accept a document without an analyst?
Not for verification. It can confirm that an upload is complete and in date, which removes most of the back-and-forth, but accepting a document as evidence of identity or address is an analyst's decision.
Should the agent restrict an account when the deadline passes?
No. A restriction is a consequential action for the customer and must be decided by a named person, who can weigh vulnerability, history and risk. The agent prepares the case for that decision.
How many reminders should the agent send?
That is a policy choice for the KYC lead, set once and enforced. A short, clear schedule with a final notice that explains the consequence usually works better than many gentle reminders.