Estate Agent AML Checks: Organise Due Diligence Files With AI

An estate agency can use AI to rename, sort and summarise due diligence documents, draft source-of-funds notes from evidence on file, and list what is missing, while the nominated officer reads the evidence and makes every accept, escalate or report decision. AI organises the file. It never decides the risk rating or the outcome.

That split is the whole design. Anti-money-laundering duties sit with the business and with named people in it, and a model cannot hold them. What a model can do is turn a pile of PDFs, photos and email attachments into a file a human can review in ten minutes instead of forty.

Follow me on Instagram@sagnikteaches

What the rules ask for, in plain terms

This is a working summary, not legal advice, and your supervisor's current guidance is the authority. Estate agency businesses are among the sectors that must be supervised for anti-money-laundering purposes, so the agency registers with its supervisor. The core duties the guidance describes are these:

Connect on LinkedInSagnik Bhattacharya

• Customer due diligence. Identify the customer and verify who they are. The supervisor's sector guidance says this applies on both sides of a sale, buyer and seller, and that for a buyer it must be completed before a binding contract exists, or if none arises, before the sale becomes irrevocable. The same guidance asks agencies to begin buyer checks as soon as possible after first contact, ideally before an offer is accepted, so the checks never hold up the sale.
• Source of funds. Where the guidance calls for it, you obtain information on the source and origin of the money the customer will use, and enhanced checks go further into where funds came from.
• Risk assessment. A business-wide assessment, and a risk view of each customer that is updated when circumstances change.
• Nominated officer. A named person to whom staff report suspicion, who decides whether to report onward.
• Records. Kept for five years from the end of the relationship or completion of the transaction, after which the sector guidance says the data is deleted.
• If checks cannot be completed. The supervisor's guidance says the business must not go ahead with the customer, must end any existing relationship, and must consider whether a suspicious activity report is required.

None of this changes because AI is involved. What changes is how fast the paperwork is assembled, and what new mistakes appear.

Subscribe on YouTube@sagnikteaches

Who does what: the split that keeps decisions human

The last row is deliberately empty. Do not ask a model whether a customer looks suspicious. It will answer fluently, and neither a yes nor a no is evidence.

Set up the file structure before you touch any AI

AI sorting only works if the destination is fixed. Agree one folder layout, one naming convention and one checklist, and use them for every file, human or machine.

/AML/2026/
  0412_Property-ref_Sale/
    SELLER_Surname-Initial/
      01_ID-photo
      02_ID-address
      03_Ownership-evidence
      04_Checklist-and-notes
    BUYER_Surname-Initial/
      01_ID-photo
      02_ID-address
      03_Source-of-funds
      04_Checklist-and-notes
    00_Officer-decision-log

Names above are placeholders. The numbered prefixes mean any file can be audited at a glance: if 03_Source-of-funds is empty on a buyer, that is visible without opening anything.

A worked example: one buyer file, start to decision

Say a small agency has an offer accepted on a flat at $385,000. The buyer sends a mortgage agreement in principle, two months of bank statements, a letter from a relative about a gift of $40,000, and photos of two documents, all by email. Illustrative numbers throughout.

Step 1: sort. A staff member, not the model, first confirms that every attachment belongs to this buyer. Then the model is asked to propose a filename and a document type for each attachment, from text the staff member has pasted or from a tool approved for the content. Output:

Attachment 1 -> BUYER_Surname-A / 03_Source-of-funds / mortgage-AIP_2026-10-30
Attachment 2 -> BUYER_Surname-A / 03_Source-of-funds / bank-statement_2026-09
Attachment 3 -> BUYER_Surname-A / 03_Source-of-funds / bank-statement_2026-10
Attachment 4 -> BUYER_Surname-A / 03_Source-of-funds / gift-letter_relative
Attachment 5 -> BUYER_Surname-A / 01_ID-photo / UNCLEAR: image cropped, name not legible
Attachment 6 -> BUYER_Surname-A / 02_ID-address / utility-bill, date missing

Step 2: gap list. The model compares the folder against your checklist:

MISSING: legible ID photo (attachment 5), dated proof of address (attachment 6),
evidence of the relative's funds, evidence of how the deposit will be paid (account in buyer's own name?)

Step 3: draft the source-of-funds note. Here the model writes only what the documents show, in neutral words.

Funding as evidenced on file (illustrative):
- Mortgage: agreement in principle, stated amount $308,000 (80% of $385,000), dated 30 October.
- Gift: letter stating $40,000 from a relative; no supporting statement for the donor on file.
- Savings: balance on September and October statements consistent with a further $37,000; origin of balances not explained.
Total evidenced or stated: $385,000. Items without supporting evidence: gift, savings origin.

A check on the sum: $308,000 plus $40,000 plus $37,000 is $385,000, so the note adds up, but two of the three components rest on statements rather than evidence. That is a fact for the officer, not a verdict from the model.

Step 4: the officer reads and decides. They open the documents, not the summary. They ask the buyer for the donor's evidence and an explanation of the savings, record the request in the decision log with a date, and only then clear or escalate. The AI's note sits in the file as a working paper, clearly labelled as such.

Prompts you can adapt

Sorting and naming

You help an estate agency organise anti-money-laundering files. For each document text below, return: customer (buyer or seller, as labelled), document type from this list [photo ID, proof of address, ownership evidence, bank statement, mortgage evidence, gift letter, other], date on the document, and a filename in the format TYPE_YYYY-MM.
Rules: do not assess the customer, do not judge whether a document is acceptable, and write UNCLEAR if any field cannot be read. Never guess a name or date.

Gap list

Here is our checklist for a buyer file and the list of documents currently on file. Return only the checklist items with no matching document, and any document whose date, name or type does not match the checklist requirement. Do not comment on risk.

Source-of-funds note

Using ONLY the documents pasted below, write a neutral note headed "Funding as evidenced on file". List each source, its amount as stated, the document that supports it, and what is not evidenced. Do not say whether the explanation is plausible, and do not use words such as suspicious, clean, low risk or satisfactory.

That last instruction matters. Banning the evaluative words stops the model from slipping a rating into the note where a hurried reader would take it as fact.

Seller files are different, and often forgotten

Agencies focus on the buyer, and the sector guidance is clear that checks cover the seller too. The seller file has its own questions: is the person instructing us the owner, and does the ownership evidence match? A model can compare the name on the instruction form, the ownership document and the ID and list every difference, including middle names, initials and spelling. Say the instruction form gives only an initial and a surname, while the ownership document gives a full first name, a middle name and the same surname: the model lists the difference, and the officer decides it is the same person after a short check, not before.

Where the seller is a company or a trust, the file grows: who owns and controls it, and how to verify them. Treat those files as officer-led. Use AI for the sorting and gap list only.

Record keeping: what to retain, and how to retire it

Guidance for supervised businesses keeps records for five years counted from the end of the relationship or completion of the transaction. Build that into the folder routine rather than hoping someone remembers:

1. Put the completion date on the file index when the sale completes.
2. Add a review-for-deletion date five years on, in a diary your officer controls.
3. Keep the officer's decision log with the file for the same period.
4. Decide whether AI-produced working papers (sorted names, gap lists, draft notes) are kept as part of the record or discarded once the officer's note exists, and write the rule down. The safe pattern is that the officer's dated note is the record and the AI draft is not relied on.
5. Check where any AI tool stores what you give it, and switch off retention for client files where the tool allows.

A second file: a buyer purchasing through a company

Take an illustrative case a small agency might meet. An offer of $620,000 is accepted from a limited company that says it is buying a flat as an investment. The director sends a certificate of incorporation, a shareholder list showing two people at 50 percent each, ID for the director only, and a bank statement for the company showing a balance of $240,000. The rest is to come from a loan.

The model's gap list is the useful part here. Asked to compare the folder against a company-buyer checklist, it can report: no ID or verification for the second shareholder who holds half the company; no evidence of the loan; no explanation of how the company's balance was built up; and no confirmation that the person instructing is authorised to act for the company. It cannot say whether any of this is worrying, and it should not try. It has turned a muddled email thread into four questions the officer can put to the director in one message.

The funding arithmetic helps too. $620,000 less the $240,000 balance leaves $380,000 to come from a loan, so the note should state that the loan is unevidenced and that the origin of the company balance is unexplained. Whether that points to a higher-risk rating is the officer's call, using your own rating sheet.

A before and after for the notes themselves

The wording of a note decides whether it helps a reviewer. Compare two versions of the same fact, written about the same illustrative gift.

BEFORE (AI default): "The buyer's funds appear legitimate and the gift from a family member is a low-risk source."

AFTER (with the neutral-note prompt): "Gift of $40,000 stated in a letter from a relative dated 2 November. No statement from the donor on file. Relationship stated, not evidenced. Not yet verified."

The first version reads like a conclusion and quietly invites the reader to stop. The second states what exists and what does not, and sets up the next request. Make the second form the house style, and treat any note containing words like "legitimate", "low-risk" or "satisfactory" as a draft that has not yet been through the prompt.

Factors to list against your own rating sheet

Your risk assessment and your supervisor's guidance decide what raises or lowers risk. The model's job is only to lay the facts on the file beside the factors your sheet already names. The supervisor's sector guidance itself singles out cash purchases, gifted deposits and a last-minute change of mortgage lender as points for extra care, so rows like those belong on most sheets. A sheet might look like this, with the right-hand column filled in by the model from the documents and checked by the officer.

The last two rows show where the model stops. "In line with stated means" is a judgement, so the column says "officer to assess". If your sheet has judgement rows, make the model leave them blank on purpose.

A short briefing for the team

Staff need three sentences they can remember. Put them on a card by the desk: the AI tidies and lists, it does not decide; anything that worries you goes to the nominated officer the same day, whatever the AI said; and nobody uploads identity images to a tool that is not on the approved list. Then run a ten-minute session using the example folder above, with names changed, and ask each person to find the gaps before the model does. Then compare each list with the model's: a gap that one side missed teaches more about why a human check stays than any policy paragraph.

It is also worth agreeing what staff do when the AI and their own eyes disagree. The rule is simple: the person trusts their eyes and tells the officer. A model contradicting a human on identity or funds is a reason to look harder, never a reason to relax.

A four-week pilot plan

1. Week 1. Agree the folder layout, naming and checklist. Run them manually on live files. No AI yet.
2. Week 2. Use the sorting and gap prompts on five closed files, using redacted text where you can. Compare with the file as staff built it and record every difference.
3. Week 3. Use the prompts on live files for one negotiator, with the officer reading the underlying documents as usual. Time each file and count the corrections.
4. Week 4. The officer runs the half-hour review above, decides whether to widen, and updates the prompts for every mistake seen.

Stop or slow down if corrections are frequent, if anyone starts relying on the AI note instead of the documents, or if identity material has gone anywhere it should not.

When a customer cannot or will not provide evidence

Sooner or later a buyer sends half the documents and then goes quiet, or objects to being asked. The model can draft a polite chaser listing exactly what is outstanding, using the gap list. Keep the wording factual and the same for every customer, so nobody can say they were treated differently:

Hi [first name], to move your purchase forward we still need: a clear photo of your passport photo page, and a utility bill dated within the last three months. Please send both by [date]. We are required to complete these checks before the sale can proceed. Thanks, [your name]

If the evidence does not arrive, the model has no further role. The officer decides, in line with your supervisor's guidance, whether the transaction can go on, which as described earlier is generally not at all without completed checks. Record the chasers sent and the dates in the decision log so the file shows the effort made.

Notice also what the chaser leaves out. It does not hint at suspicion, and it does not say why a particular item matters beyond the requirement. Anything that touches concerns about a customer is spoken or written by the nominated officer, not drafted by a model and sent by a junior colleague.

Where AI gets AML files wrong

Names that nearly match

Models treat "Jon" and "John", or a maiden and married name, as the same person without saying so. Tell the prompt to list every near-match as a difference, not to reconcile it.

Expiry and date maths

A model will sometimes decide a proof of address is "recent" by guessing today's date. Give it the date and your own rule, for example "utility bill dated within the last three months", or do the date check by hand.

Polished summaries that hide thin evidence

A fluent source-of-funds note can make weak evidence read as strong. The "what is not evidenced" line is there to prevent that, and the officer should read it first.

Mixed-up files

If two customers share a surname or address, one customer's statement can land in the other's folder. A staff member confirms ownership of every attachment before anything is sorted, and the officer spot-checks a sample each month.

Protecting the data you are handling

Identity documents, addresses and bank statements are exactly the material you least want in the wrong place. Pick one of three arrangements and write it down: an approved tool with data terms you have read and retention controls; your existing checking or CRM system's own AI features if it has them; or text-only prompts where staff type or paste the minimum fields and never upload images. For the general approach to rules of this kind, see the AI compliance checklist for small businesses.

A monthly review the officer can run in half an hour

• Pull five files at random and compare the AI gap list against what a person finds.
• Check that every file with an AI-drafted note also has a dated officer decision.
• Count files where the checks were not finished before contracts were exchanged, and ask why.
• Look for any file where a document was sorted into the wrong customer folder.
• Confirm that deletion dates exist on completed files.

Keep the findings short and dated. If the sample shows errors, tighten the prompt or the checklist before widening use.

Rolling it out in a small agency

Start with the folder layout and checklist, which need no AI and improve things on their own. Run the sorting prompt on closed files for a fortnight and compare with what staff did. Only then use it on live files, one branch or one negotiator first. If you are working out wider staff rules for AI, the governance checklist for advice firms shows how short a useful policy can be, and sorting incoming documents with AI covers the general sorting method in more detail.

The officer's decision log is the piece auditors and colleagues will care about most. The sector guidance frames the test plainly: if your supervisor asked what you did to establish source of funds, could you explain and evidence the decision, and why you made it? A good entry has four parts: the date, what the officer reviewed (documents, not summaries), the decision, and the reason in a sentence. Keep it consistent so a stranger could follow it.

If you would like help mapping where AI fits in your own file process, my AI implementation consultation begins with the way your files move today.

Questions agency owners ask about AI and AML files

Can AI approve or reject a customer?

It should not. The decision to accept a customer, escalate a concern or report suspicion belongs to a named person, usually the nominated officer. AI can sort documents and list gaps, but its output is an input to that person, never the outcome.

Is it safe to upload passports to a chatbot?

Not to a general consumer tool. Identity documents are sensitive. Use only a tool approved for that data under terms you have read, or keep identity images in your checking system and give the AI text fields only.

How long must the file be kept?

Government guidance for supervised businesses says records are kept for five years, counted from the end of the business relationship or the completion of the transaction. Check your supervisor's current guidance, and build deletion into the routine.

What if the AI summary misses a red flag?

That is why the officer reads the underlying documents for any file rated above routine risk, and why the AI summary is never the record. The record is the documents and the officer's dated decision.

Further reads

• How Conveyancers Use AI to Cut Admin on Each Transaction — Another property-side workflow with human checkpoints.
• How Estate Agents Use AI to Prepare Valuation Appointments — Where the instruction process begins.
• How to Automate Buyer and Vendor Follow-Up With AI — Chasing missing documents politely and on time.
• AI Listing Mistakes That Can Mislead Buyers, and How to Check — Another place where unchecked AI output causes harm.
• AI Tools for Estate Agents: A Stack by Job, Valuation to Completion — A stage-by-stage AI stack for a small estate agency: valuation, listing, viewings, offers and progression, with costs and the checks each needs.
• Due Diligence Checklist for Selling a Business: Build One With AI — A seller-side due diligence checklist by category, plus how a broker uses AI for the data room index, the gap tracker and the buyer question log.
• Landlord Compliance Checklist: Track Gas Safety and EICRs With AI — Read certificates with AI, log expiry dates in one tracker, set reminders and keep a short exception list, with a human check on every certificate.
• Letting Agent Software With AI: A Checklist Before You Switch — A vendor question list for letting agencies: data export, client money, what the AI really does, per-property pricing and a demo script to score.
• Can AI Help With Right to Rent Checks? Where a Human Must Sign Off — Where AI can help with right to rent checks (first-pass checklists, record tidying, reminders) and the sign-off points a human must keep.
• Which AML Client Checks Can an Accountant Automate With AI? — A step-by-step map of client due diligence in an accounting practice: what to automate, what AI may only draft, and what must stay with a person.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: government guidance on money laundering regulations for supervised businesses (record keeping, nominated officer, source of funds) and the supervisor's estate agency sector guidance (timing of customer due diligence), read in November 2026. Check current guidance before relying on this.