A KYC file has two lives. The first is the day the customer opens an account, when your Customer Identification Program has to collect and verify who they are. The second is everything after: the reviews, refreshes and event-driven updates that keep that file true for as long as the account is open. Most of the manual hours sit in both, and in the hand-offs between the tools that serve them.
This guide is for a BSA officer, head of onboarding or compliance operations lead at a US bank or fintech. It covers how automation works in those two stages, which steps software can run, and which decisions stay with your analysts. It stays on CIP and refresh on purpose. Beneficial ownership under the CDD rule is covered in CDD and beneficial ownership: automating FinCEN CDD rule workflows, and business customers in KYB automation for banks and fintechs. For transaction monitoring and the wider program, see the AML automation hub.
A note on names: this is Zamp (zamp.ai), which builds AI employees for back-office work. It is not Zamp HR (payroll and PEO) or the zamp.com sales-tax platform.
How does KYC automation work?
What does the CIP rule actually require?
For banks, the rule is 31 CFR 1020.220 (eCFR). Fintechs that offer accounts through a sponsor bank usually work inside the bank's CIP, so the same requirements reach them by contract. The rule asks for these things:
- Collect four data points before the account opens. Name, date of birth for an individual, address, and an identification number. For a US person that is a taxpayer identification number. For a non-US person it can be a passport number and country of issuance, an alien identification card number, or another government-issued document number.
- Verify identity within a reasonable time after the account is opened. The program must use risk-based procedures and may rely on documents, non-documentary methods or both. Examples in the rule include an unexpired government-issued photo ID, and checks against a consumer reporting agency or public database.
- Handle failure. The program must say when the bank will not open an account, when it may be used while verification continues, when to close it, and when to consider a Suspicious Activity Report.
- Compare against government lists of known or suspected terrorists, where Treasury has designated such a list, and give customers notice that you are asking for this information.
- Keep records. The identifying information is kept for five years after the account is closed. Descriptions of documents relied on, verification methods and results, and how discrepancies were resolved are kept for five years after the record is made.
The rule does not say how you collect or check any of this. That is why automation fits: the requirements are specific, the evidence is structured, and every step has to be recorded.
What steps of CIP can software run?
Step | What the software does | What a person still decides |
|---|---|---|
Collect | Pre-fills the four required data points from the application, flags missing fields, requests documents | Policy on what you accept when a customer has no residential address |
Extract | Reads data from ID images and compares it with what the customer typed | Cases where extraction confidence is low |
Verify | Runs document and non-documentary checks through your identity provider and records the result | Mismatches, thin-file customers, anything the policy sends to manual review |
Screen | Runs the name through your sanctions and PEP screening tools and the government-list comparison | Whether a partial match is a true match |
Score | Applies your customer risk policy to produce a tier | Overrides, and any customer your policy treats as high risk |
Record | Writes inputs, results and reasoning to the case file | Nothing; this should never depend on someone remembering |
Identity verification vendors cover the verify step well. The slow part is usually the work between steps: moving data from the application to the verification tool to the screening tool to the case system, chasing a missing document, and assembling the file before an analyst can look at it. An AI employee works across those systems the way an analyst would, following your written procedure and logging each action. It is not a replacement for the verification vendor or the screening tool. It runs the process that connects them.
Where does periodic refresh fit?
CIP is a point-in-time requirement. What keeps a file current is your customer due diligence program, which the AML program rule at 31 CFR 1020.210 describes as risk-based procedures for ongoing customer due diligence. That includes understanding the nature and purpose of the relationship to build a customer risk profile, and ongoing monitoring to report suspicious transactions and, on a risk basis, to maintain and update customer information (eCFR).
Two points follow from that wording.
First, no regulation sets a refresh interval. Your risk tiers do. Many institutions review high-risk customers more often than low-risk ones, and the schedule belongs in your written policy.
Second, a calendar review is not the only trigger, and for beneficial ownership it is not a trigger by itself. FinCEN's CDD FAQs (FIN-2018-G001, Question 14) say institutions need not solicit or update beneficial ownership information as a matter of course during periodic reviews, absent specific risk-based concerns. The obligation arises when normal monitoring turns up information relevant to the customer's risk profile. On February 13, 2026, FinCEN issued exceptive relief order FIN-2026-R001, which lets covered institutions limit beneficial ownership identification and verification to when a legal entity customer first opens an account, when facts call earlier information into question, and as needed under risk-based CDD procedures, instead of at each new account; our CDD post covers that and the ownership rules in depth.
How do you automate periodic refresh?
A refresh program has two engines: a calendar and a set of events. Software can run both.
- Schedule by tier. Each customer carries a next-review date driven by risk tier. When it comes due, the system opens a case, pulls the current file, and compares it with the latest data from your core system.
- Watch for events. Triggers can include a changed address or name, a new product, a change in expected activity, a new screening hit on the customer, or an adverse media alert. Each opens a case in the same queue. Teams sometimes call this perpetual KYC.
- Re-screen. The customer is run against current sanctions and PEP data. Screening the customer base when lists change is a separate job from screening at onboarding, and the sanctions screening glossary entry explains the difference.
- Request what is missing. If a document has expired or an attestation is due, the system sends the request, tracks the response and follows up.
- Prepare the file. The AI employee summarizes what changed, notes inconsistencies between systems, and lists open questions, so the analyst starts with a reviewable packet.
- Decide. An analyst confirms the refresh, escalates the customer, or recommends exit. If the facts raise suspicion, the file goes to your investigations team and the SAR decision follows that process.
A customer whose profile has not changed and whose screening is clean can be refreshed with no analyst touch if your policy allows it. That policy choice is yours to document.
What should stay with people?
Automation is a good place to be strict about this, because examiners ask who made each decision.
- High-risk and PEP customers. Software prepares the file. A named person approves the relationship.
- Ambiguous screening matches. An analyst decides, and the reason is recorded.
- Failed verification. The decision to decline, restrict or close an account, and any SAR consideration, sits with people.
- Overrides of the system. Any disagreement is logged with a reason.
- Policy. Risk tiers, refresh intervals and exception thresholds are set by your compliance function, not by the vendor or the model.
How do model-risk expectations apply?
Identity verification scores, risk-scoring models and name-matching engines can be models in the supervisory sense. On April 17, 2026, the Federal Reserve and the OCC issued SR 26-2 and OCC Bulletin 2026-13, revised model risk guidance that replaced SR 11-7 and SR 21-8 (the 2021 interagency statement on models supporting BSA/AML compliance). The OCC bulletin also rescinds OCC Bulletin 2021-19. The new guidance takes a risk-based approach, covers vendor models, and says generative and agentic AI models are outside its scope. It also says it does not set enforceable standards.
In practice, treat your conventional scoring and matching models under that guidance, and treat any generative or agentic component with your own governance: documented scope, approval gates, testing and logging. Ask your vendors which parts of their product are conventional models and which are generative, because the expectations differ.
What are common failure modes?
- Automating around bad data. If customer records disagree across systems before you start, automation moves the disagreement faster.
- One queue for every exception. If screening hits, missing documents and discrepancies land together without priority or context, analysts spend their time sorting.
- Weak case logs. Recording the outcome without the steps behind it fails an exam even when the decision was right.
- Refresh with no event triggers. A calendar alone misses the customer whose risk changed in month three of a 24-month cycle.
- Treating high false-positive rates as a vendor problem. Tuning thresholds is a program decision that someone has to own.
How does this look with an AI employee?
Zamp's AI employees follow your written procedure, work across your onboarding, document, screening and case systems, and escalate when confidence is low or a threshold is crossed. The same pattern appears in our published case studies of regulated back-office work, for example multi-entity billing operations and freight order management, where each action is logged and low-confidence cases go to a person. For the compliance analyst role in a bank, see AI compliance analyst and the banking and financial services compliance guide.
Sources
- eCFR: 31 CFR 1020.220 (CIP), 31 CFR 1020.210 (AML program), 31 CFR 1010.230 (beneficial ownership).
- FinCEN, Exceptive Relief from Requirement to Identify and Verify Beneficial Owners at Each Account Opening, FIN-2026-R001 (February 13, 2026).
- FinCEN, Frequently Asked Questions Regarding Customer Due Diligence Requirements, FIN-2018-G001 (April 3, 2018).
- Federal Reserve, SR 26-2, and OCC, Bulletin 2026-13 (April 17, 2026).
