AI in KYC for a Small Wealth or Advice Firm: Checks That Stay Human

A small wealth or advice firm can automate the capture of ID documents, the first-pass checks on them, form pre-filling and reminder chasing, but must keep human decisions on source-of-wealth credibility, risk rating, screening hits, escalation and the final approval to onboard. AI collects and organises evidence; a named person decides.

The reason is practical as much as regulatory. Know-your-customer checks exist so that someone at the firm can say, with evidence, who the client is, where the money comes from and why the firm accepted them. A model can read a passport image or tidy a questionnaire. It cannot be accountable for a judgement. In a review or an investigation, "the software said it was fine" is not an answer; "I reviewed this, here is my reasoning and the documents" is.

Follow me on Instagram@sagnikteaches

Where a human has to stay: the short list

Start with the decisions, because a firm that is clear about these can automate everything else with less worry.

Connect on LinkedInSagnik Bhattacharya

1. Accepting or declining a client.
2. Assigning the client's risk rating and deciding what extra checks that triggers.
3. Judging source of wealth and source of funds: is the explanation credible and consistent with the documents and the person's profile?
4. Clearing or escalating screening hits for sanctions, politically exposed persons and adverse media.
5. Handling anything unusual: complex ownership structures, vague explanations, pressure to hurry, evidence that does not fit.
6. Deciding whether to report a suspicion through the firm's internal reporting route.
7. Periodic reviews and trigger events, where a change in circumstances needs judgement.

Everything else in onboarding sits around those decisions. Which exact checks, thresholds and records your firm must follow depends on its regulator and its own policy, so treat what follows as a design pattern to discuss with your compliance lead, not as legal advice.

Subscribe on YouTube@sagnikteaches

Step by step: automate, assist or keep human

Illustrative setting: a seven-person advice firm onboards about 12 new clients a month. Before changes, each file takes an adviser and the compliance officer roughly 3 hours of combined effort. The aim of this design is to reduce the clerical share, perhaps by a third, without touching the judgement share.

Document capture and ID checks

This is the safest place to automate because the task is narrow: is this document genuine, current and consistent with what the client told us? Dedicated identity-verification providers do this with document scanning, data extraction and sometimes a selfie match with liveness detection. These are regulated-sector products with their own pricing, usually per verification, so ask for quotes and check what the provider does with the images. A general chatbot is not a substitute: it is not built to detect forged documents, and its privacy terms are not written for identity data. Use the dedicated tool for authenticity and use general AI, if at all, for admin around it.

What the human does here: look at every result that is not a clean pass. A "pass with warnings" on a passport might be a worn chip or a doctored page. Rules for the reviewer, written down:

• Any failed authenticity check goes to the compliance officer, not back to the client for a retake without review.
• If a name differs between documents, ask for the explanation (marriage, spelling, transliteration) and record it.
• If a client cannot or will not provide something the firm needs, record the refusal and escalate rather than working around it.

Form pre-filling and chasing

Here is where a general AI assistant earns its keep, using the minimum data. Using a business plan the firm has approved, you can ask a model to turn a client's free-text answers into the structured fields of your fact-find, or to draft a polite message listing the three documents still missing. Client data should be reduced to what the task needs. For a chaser you only need first name and the list of items, so there is no reason to paste an account number. Illustrative chaser prompt:

Write a short, friendly email asking [first name] for the documents below to complete our onboarding checks. Say we cannot proceed until we have them, offer a secure upload link placeholder [link], and give a deadline of [date]. Do not mention any reason for the request beyond "regulatory onboarding checks". Under 100 words.
Missing: proof of address dated within the last three months; evidence for the source of the $60,000 deposit.

What you fix: check the tone does not hint at suspicion, that the deadline is realistic and that the request matches what your policy actually requires. The adviser sends it. For further onboarding automation ideas in a similar regulated practice, see automating client onboarding in a small accounting practice.

Screening hits: speed up the reading, never the decision

Screening tools match names against lists and news. They return many near-misses: a common name, a similar date of birth, a relative. The temptation is to let AI clear those. Resist it. Use AI, if at all, to put the information in front of the reviewer more clearly: a summary of why each hit matched, what differs (date of birth, country, occupation), and what extra information would settle it. The reviewer then looks at the underlying record and decides.

A worked example (illustrative): a new client named in a screening hit as a possible match for a politically exposed person. The tool shows a match on name and a birth year within two years. The reviewer pulls the screening record, which shows a different middle name, a different occupation and a country the client has no connection with. The reviewer clears the hit and records: "Different middle name and occupation; client has no link to the country listed; cleared by [reviewer], [date]; evidence saved as file 14." A second case: a hit on adverse media about a company in which the client is a director. That does not clear; it escalates to the compliance officer, who requests more information from the client and decides whether enhanced checks apply. The point of recording both is that you can show your reasoning later.

Source of wealth: where AI helps and where it must not

Source of wealth is the hardest part to automate because it is a story that has to fit evidence. A client says the $400,000 they want to invest came from selling a business three years ago. You need the sale agreement or completion statement, evidence the proceeds reached an account, some sense that the business was real, and consistency with their age and occupation.

AI can: draft the question list for the client; extract amounts and dates from the documents so the adviser can compare them; produce a timeline from the adviser's notes; list gaps ("no evidence yet of how proceeds reached the account"). It must not: say whether the story is plausible; estimate what a business might have sold for; fill gaps with assumptions; or produce a conclusion the reviewer signs without reading the documents.

Below are my notes on a client's stated source of wealth and the list of documents received. Do not assess plausibility. Produce:
1) A dated timeline built only from the notes.
2) A table of each claim in the notes and which document, if any, supports it.
3) A list of claims with no supporting document.
4) Questions I could ask the client to close each gap.

This prompt keeps the model at clerical work. The adviser then reads the documents, decides whether the picture holds, and writes the conclusion in their own words, with the reasoning. If you would like a deeper look at the governance around this kind of use, the governance checklist for small advice firms is the companion piece.

Records: what to keep so you can explain every check

For each client, keep a file that lets a stranger follow what happened. Typical contents (follow your own regulator's and firm's retention rules):

• The documents received, with dates and who checked them.
• The result of the document verification, including failures and retakes.
• Screening results with the reviewer's name, the decision and the reasoning for each hit.
• The source-of-wealth and source-of-funds evidence and the adviser's conclusion.
• The risk rating, the factors behind it and any override with reasons.
• The approval, with name and date.
• Where AI helped: the tool, its role (for example, "drafted chaser", "tidied fact-find answers") and who checked the output.

The last line is worth including. If a regulator or auditor asks how AI is used in onboarding, a file that says exactly where and who checked is a strong answer. Do not rely on a chat history as the record: histories sit in personal accounts and can be deleted. Save what matters to the client file at the time.

Choosing tools without taking on new risk

Questions to put to any provider that touches identity data are covered in questions to ask before buying AI that touches client data. For KYC specifically, ask: where are images stored and for how long; can we delete on request; who are the sub-processors; does the provider use our data to train its models; what are the false-reject and false-accept patterns you can show us on a sample of our own clients; and can the human reviewer see why a check failed? Business plans of general assistants such as ChatGPT Business and Claude Team say they do not train on your content by default, but that does not make them suitable for identity documents. Use them for text drafting under your policy, and use the verification provider for identity. If your CRM already has onboarding checks built in, compare it first; what is built in versus bolted on in adviser CRMs explains the options.

A worked example: one new client from first call to approval

Say a client in her fifties wants to move $250,000 into an advised portfolio after the sale of a small property (illustrative). Here is how the checks run.

Day 1. The adviser sends the secure form link. The client completes it and uploads a passport image and a recent utility bill. The verification provider returns a pass on the passport but a mismatch on the bill, which shows her former surname. The system flags it; no AI judgement is involved.

Day 2. The adviser reads the flag, sees the former surname, and asks for the marriage certificate or a second address proof. A chaser, drafted with AI and read by the adviser, goes out. The client replies with a bank statement in the current name.

Day 3. Screening returns one possible match on adverse media. The compliance officer reads the article, sees that it concerns a different person of the same name in another sector, clears it and records the reasoning with the article saved.

Days 4 to 6. The source-of-funds question is the sale. The client provides the completion statement. The adviser uses the timeline prompt to list claims and supporting documents. The statement supports the amount and date; the proceeds reached an account in her name, shown by a bank statement. One gap: the purchase price of the property when she bought it years ago, which is not needed here. The adviser notes that and moves on, without letting the model "infer" a history.

Day 7. The risk rating applies the firm's scoring rules. The adviser reviews the factors, agrees with the standard rating and writes two sentences of reasoning. The compliance officer approves.

The clerical work took about 90 minutes of adviser time against perhaps two hours before, and the human work took the same time as before, which is the intended outcome. The records show the surname mismatch, the cleared hit and the evidence trail.

A file note a stranger could follow

Ask the model to draft the file note from your own bullet points, then edit it for truth. A good note (illustrative) reads:

Client: [code]. Onboarding approved [date] by [approver].
Identity: passport verified by [provider], pass. Address: bank statement dated [date], matches current name. Utility bill showed former surname; explained by marriage; evidenced by bank statement.
Screening: one adverse-media hit, reviewed by [reviewer] on [date]; different individual (different sector and location); cleared; article saved to file.
Source of funds: sale of property, $250,000; completion statement and bank statement on file; consistent with client's stated circumstances.
Risk rating: standard. Factors: [list]. No override.
AI use: chaser email drafted with [tool], reviewed and sent by [adviser]. Timeline of source-of-funds claims drafted with [tool] from adviser notes and checked against documents by [adviser].

The model improves the prose; it cannot supply the content. Every sentence has a document or a named person behind it, and the AI-use line states what it did.

Ongoing monitoring and triggers

KYC is not a one-off. Clients change address, receive an inheritance, become public officials or alter how they use the account. AI can help the firm notice and prepare, for example by drafting the annual refresh questionnaire, listing clients whose documents are about to expire, and summarising meeting notes for anything that sounds like a change of circumstances (a large unexpected receipt, a new country of residence). It must not decide that nothing has changed. A person reads the flag, asks the question and decides whether to refresh the checks. Record the review date and the reviewer, and schedule reviews by risk rating: higher-rated clients more often. The ongoing client conversations are the best source of signals, and the meeting-records tutorial for advisers, AI note takers for financial advisers, covers recording them properly.

Four realistic mistakes

Auto-clearing low-score hits. A firm configures the tool to clear any hit scoring below a threshold. Months later a true match turns out to have scored low because of a spelling variant. The fix: a person samples cleared hits, and every hit above a minimum is reviewed by name.

Pasting a whole client file into a chatbot. An adviser asks a free assistant to "summarise this client's source of wealth" with the passport number attached. The fix: approved tools only, minimised data, and a rule that identity documents never go to a general chatbot.

The summary that hides a gap. A model's neat paragraph says "funds derive from the sale of a property" when the file only holds the client's own statement. The fix: every claim in a file note links to a document or is labelled "client statement, unevidenced".

Treating speed as the goal. A firm promises same-day onboarding and shortens review. The fix: measure time to a correct decision, not time to a decision, and let escalations take as long as they take.

A first-month plan for a small firm

1. Week 1: map your current onboarding steps from first call to approval, and mark each with the table above. Agree with the compliance lead which steps may be automated.
2. Week 2: automate the clerical layer: the online form, upload link and the chaser reminders. Draft the chaser prompts and store them with the compliance officer's sign-off.
3. Week 3: trial an identity verification provider on a handful of consenting new clients, alongside the existing manual check. Compare outcomes and time.
4. Week 4: write the review rules for failures and screening hits, the file checklist and the AI-use line for the record. Review one completed file as a team to test whether a stranger could follow it.

Measure sensible things: time from first document to approval, number of files returned for gaps, number of hits reviewed and escalated, and whether every file passes the stranger test. If you would like to talk through your own onboarding with someone, that is what my AI implementation consultation covers. A related piece on the document that follows onboarding is drafting proposals with a compliance review.

Questions small firms ask about AI and KYC

Can AI approve a new client?

No. AI can capture and pre-check evidence, but approval, risk rating and the decision to onboard, decline or escalate should rest with a named person. Regulators expect firms to own and explain these decisions, so a tool's output cannot be the decision itself.

Is a document check by software enough to verify identity?

It is a strong first layer for authenticity and consistency, but your firm's rules decide what counts as verified. A person still reviews failures and borderline results, and should be able to explain what was checked and why it was accepted.

Who reviews sanctions or adverse-media hits?

A trained person. Screening tools produce many false positives, especially on common names, and a missed true match is serious. Record who reviewed each hit, the evidence used to clear or escalate it, and the date.

Can a general chatbot summarise a source-of-wealth file?

It can tidy your own notes and draft questions, but it must not judge whether the explanation is credible or fill gaps. Use an approved business tool, minimise personal data and check the output against the documents. Credibility is a human decision.

Further reads

• Can Financial Advisers Use ChatGPT? What Compliance Allows — What compliance typically allows before any tool is switched on.
• How Advisers Use AI to Prepare for Annual Client Reviews — Ongoing review work after onboarding.
• Can AI Draft a Suitability Report That Passes Compliance Review? — The next compliance-heavy document after KYC.
• AI Financial Planning Software for Advisers: What to Test First — Vendor questions to test before software touches client data.
• Does a Five-Person Business Really Need an AI Policy? — A short policy covering approved tools and banned uses.
• AI Agents for Financial Advisers: What They Can Do Without Advising — A task-by-task map of what AI agents can do in an advice practice without advising, checked against vendor policies and backed by a supervision log.
• AI Automation for Financial Advisers: Onboarding Without Re-Keying — An entry-point map for adviser onboarding: form to CRM to planning software, consent boxes, an AI file note for review and ten error checks.
• AI Software for Wealth Management: A Small Firm's Evaluation Checklist — A three-pass checklist for small wealth firms: knockout questions, tests on your own meetings and reports, and the full cost including archiving.
• AI Tools for Investment Managers: Client Reports and Fact Sheets — How a small investment manager can draft quarterly reports and fact sheets with AI while the numbers stay locked and compliance gets a reviewable packet.
• AI Tools for UHNW Wealth Managers: Seven Jobs With Privacy Rules — A job-by-job guide to AI in a small wealth team, including access limits, sample outputs and review gates.
• Companies House ID Verification: A Client Chaser Workflow for Accountants — A tracking sheet, a five-status chase sequence and email templates for getting every director and PSC verified, with AI kept away from ID documents.
• AI in Wealth Management Examples: Five Admin Jobs Worth Trying — Meeting notes, follow-ups, review packs, onboarding checklists and newsletters: five admin jobs a small wealth firm can give AI, plus the records caveat.
• Can AI Write an Investment Proposal? Draft Narrative, Check Numbers — A practical adviser review of statement extraction, portfolio arithmetic, proposal wording and sign-off.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business and Claude pricing pages (training defaults), checked 4 November 2026; general customer due diligence practice. Your regulator's rules and your firm's policy decide the specific requirements.