Use AI to draft follow-ups and tidy CRM notes from facts you have logged, never to quote a rate, payment or approval. Store consent, licence number and do-not-contact status in CRM fields, give the assistant a fixed list of banned claims, and put every outgoing message through a short compliance check before it sends.
A rate in a message is more than a number. Advertising rules attach mandatory wording to a stated rate and treat a stated payment, down payment or loan term as a trigger for further disclosures, a campaign or follow-up that promotes a loan can count as advertising, and an assistant has no pricing engine, lock date or credit profile to quote from. Rules also differ by jurisdiction and by whether you are an independent broker or a lender's employee, so your compliance contact outranks this tutorial. The method below locks the facts in the CRM, constrains the draft, checks it and logs the send.
What the assistant may touch, and the hard stops
Draw the line before you open the CRM. This is the division I would use in a small brokerage; adjust the right-hand column with your compliance contact.
| Task | The assistant's part | Hard stop |
|---|---|---|
| Follow-up drafts | Writes from CRM facts and a template | No rate, payment, fee or APR; no "approved" or "qualify" |
| CRM notes from calls | Summarises what the borrower said | Records statements as stated, never as verified; no eligibility conclusions |
| Document chasers | Turns the lender's condition list into plain language | Never invents deadlines or consequences |
| Compliance pre-check | Reads a draft for banned claims and missing lines | Flags only; never clears or approves |
| Lead sorting | Orders contacts by activity such as replied or uploaded | Never uses or infers protected characteristics or their proxies |
| Rate or payment quotes | None | Come from the pricing engine and are sent by a person |
Everything else stays with you: pricing, product fit, disclosures and any view on credit or underwriting. A useful test for any new task is whether a wrong answer could change what a borrower believes about their loan. If it could, a person writes it or checks it.
Rule families to verify for your licence, and where each lives
Which rules apply depends on where you are licensed, whether you broker or lend, and which lenders take your files. I have grouped them by family rather than by name, because names change with the jurisdiction and the habit does not. For each family, put the middle-column question to your compliance contact or to the regulator that licenses you, then build the answer into the place in the right-hand column.
| Rule family | What to verify | Where it lives in the CRM |
|---|---|---|
| Advertising content | Which stated terms (rate, payment, down payment, loan term) bring mandatory wording or extra disclosures, and which claims count as misleading | Banned-claims list in the prompt and the pre-check; a "contains stated terms" tick |
| Eligibility claims | When wording such as pre-approved, qualified or guaranteed is allowed at all | The same banned list; the stage field limits which template is available |
| Fair lending | What counts as discouraging or steering a prospect on a protected basis | Written lead-sorting rules; a periodic review of who receives which message |
| Licence display | Where your licence number must appear: ads, websites, first contact, email signatures, cards | A footer field per officer, pasted by the template, never written by the assistant |
| Calls and texts | Consent needed for each channel, do-not-contact lists, automated or AI-voiced contact | Consent fields with source, date and wording version; one do-not-contact flag |
| Referral and co-marketing | What may be given or accepted for referrals or shared marketing | A source field; an approval step for any partner campaign |
| Records | How long communications, scripts and training material must be kept | Archive of each materially different message and prompt version, with a keep-until date |
| Lender terms | Whether a lender's marketing guide, logo, rate sheet or conditions may be shown to borrowers or fed into a tool | A folder of approved lender wording; nothing else gets pasted |
Two details are worth checking against the rule text itself. Fair-lending rules reach statements made to prospects, not only credit decisions: one widely used rule covers any spoken or written statement, in advertising or otherwise, that would lead a reasonable person to believe they would be refused, or offered worse terms, because of a protected characteristic. That rule was reworded in 2026, so ask your compliance contact for the current text rather than relying on older training material. And one widely used set of mortgage advertising rules requires keeping every materially different communication, plus sales scripts and marketing and training material, for 24 months after last use. That is a minimum to ask about, not your answer. If you work for a lender rather than as an independent broker, most of this arrives as a marketing guide from its compliance team; where the guide is stricter than the rule, the guide wins.

The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowSet up the CRM before writing a single prompt
The assistant is only as safe as the facts it starts from, so put the compliance facts in fields it can read, not in your memory. For a two-officer brokerage I would add these to every contact:
- Consent by channel: email, text and call each get a yes or no, a date, the source (web form, event, referral) and the version of the consent wording the person saw.
- Do not contact: one flag that stops every sequence and every manual send.
- Licence line: the footer text for the sending officer, including the licence number, pasted in by the template and never written by the assistant.
- Rate or payment discussed: yes or no, with a date, set by the officer so the notes show when and by whom a figure was given.
- Stage: one of the stages in the table below, which decides which templates the assistant may use.
- Source: where the lead came from, so referral rules can be checked whenever a partner is involved.
- AI-assisted and last review: a tick when AI touched a message, and the date a person last checked the file.
Where the assistant sits matters as much as the fields. Pasting facts into a separate assistant keeps the CRM clean but means you carry data across by hand, so paste a first name and the fields the prompt needs and nothing more. A CRM's built-in drafting feature keeps the data inside a system you have already secured, though you should check what it can read and whether it can send by itself. An automation that drafts into the CRM from a trigger is the fastest and the riskiest, because it removes the pause in which you would have caught a problem; keep it to drafts that wait for approval. If you have not chosen a CRM yet, choosing a mortgage CRM with AI built in covers what matters, and the comparison of AI CRMs for loan officers puts three side by side.
With the fields in place, the stage decides what the assistant may draft:
| Stage | AI may draft | Never from AI |
|---|---|---|
| Enquiry | Acknowledgement and call booking | Any rate, payment or "you qualify" |
| Application started | Document checklist in plain language | Comments that the file "looks fine" or "will be approved" |
| Documents outstanding | Reminders based on the lender's condition list | Invented deadlines or consequences |
| Submitted | Neutral status updates such as "submitted on [date]" | Predictions of the decision or timing the lender has not given |
| Conditions | Plain-language list copied from the lender's wording | Changing what a condition means |
| Clear to close | Scheduling and what to bring | Fee or cash-to-close figures |
| Closed | Thank-you note | Referral or review incentives |
| Past client | Annual check-in from one approved template | Refinance savings claims or rate comparisons |
A follow-up prompt with the banned claims built in
This prompt writes one follow-up from CRM fields. The banned list is the part to tune with your compliance contact; start strict and loosen it only with their agreement. The tutorial on emails and case summaries for mortgage brokers has more prompt patterns once this one is working.
You write a short follow-up message for a mortgage broker. Use ONLY the CRM facts below. If a fact is missing, write [MISSING: name it].
CRM FACTS
- First name: [first name]
- Stage: [stage]
- Last contact: [date and channel]
- What the borrower asked or said: [...]
- Open items from our side: [...]
- Open items from the borrower: [...]
- Next step and date agreed: [...]
END CRM FACTS
RULES
1. Under 110 words. Plain language. One request.
2. No interest rates, APRs, payment amounts, fees, down payment figures or loan terms of any kind.
3. Never say or imply approved, pre-approved, qualified, eligible or guaranteed, or that the file will be accepted.
4. Do not use: lowest, best, save, no cost, rates are rising, lock now, limited time.
5. Do not reassure the borrower about any outcome.
6. Do not mention any personal characteristic of the borrower beyond what they raised themselves.
7. Do not predict what a lender or underwriter will decide, or when.
8. If the borrower asked about rates or payments, write the HOLDING REPLY below instead of answering.
9. End with the LICENCE LINE and OPT-OUT LINE below, unchanged.
HOLDING REPLY: Rates and payments depend on your credit profile, the property, the loan amount and the lender's pricing on the day, so I do not give a figure in writing until I have reviewed your details. [one proposed call time and a list of what I need]
LICENCE LINE: [officer footer]
OPT-OUT LINE: [approved wording]
Rule 5 exists because of drafts like the next one. Here is a first-run draft for a borrower at the documents stage who asked whether being self-employed would be a problem (illustrative):
Hi [first name], thanks for sending the first three documents. Don't worry about being self-employed, most lenders are comfortable with it and you should qualify without trouble. To keep things moving I still need your last two years' tax returns, with every page, and your most recent business bank statement. Could you upload them by Friday? [licence line] [opt-out line]
What I would change, in order:
- Delete the reassurance sentence. It predicts eligibility, nobody has assessed the file yet, and the assistant volunteered it. Rule 3 bans the word "qualified", but a phrase such as "should qualify without trouble" can slip past it, which is why rule 5 bans reassurance outright.
- Replace "Friday". The assistant invented a deadline. Use the date the lender actually gave, or ask when the borrower can upload.
- Check each document name against the lender's condition list. "Tax returns" and "business bank statement" were on it here; if one had not been, it would not go in the message.
The borrower's real question still needs an answer, so it goes onto the officer's call list rather than into the assistant's mouth.
Answering a rate question without quoting one
The most tempting message for an assistant is the reply to "what is your rate today?" It will want to be helpful, and that is the problem. Left alone it produces something like this (illustrative, with a placeholder where it would invent a number):
Great question! Today's rates start around [x]% on a fixed loan, so it's a good time to lock in.
Almost every phrase is a problem. The figure is invented or stale, "start around" is a headline-rate claim, "a good time to lock in" is urgency, and a stated rate can turn an ordinary reply into a regulated advertisement. The holding reply in the prompt does better:
Thanks for asking, [first name]. Rates and payments depend on your credit profile, the property, the loan amount and the lender's pricing on the day, so I do not give a figure in writing until I have reviewed your details. Could we speak on [day] at [time], or would you rather send me [list] first? Once I have them I can send figures for your scenario.
The figures themselves come from your pricing engine, are sent by a person, and carry whatever disclosures your rules attach. Tick "rate or payment discussed" in the CRM when you send them. One related habit for brokers: check each lender's terms before a rate sheet goes anywhere near a consumer message or a third-party tool, because wholesale pricing may come with conditions you have agreed to. Do not assume; read the terms.
Document chasers: before and after
Chasers are the safest high-volume job for an assistant, provided it works from the lender's own list. Compare these two versions of the same nudge (illustrative):
Before: Hi, we're still waiting on your paperwork. The lender will decline your file if we don't have it by Friday.
After: Hi [first name], to keep your file moving I still need your last two bank statements, all pages, and the signed disclosure form. The lender's condition list asks for both before they can review the file. Could you upload them by [date the lender gave], or tell me if something is holding you up?
The first version invents a consequence and a deadline, and neither came from the lender. The second repeats only what the condition list says and leaves room for the borrower to explain a delay. Put the condition list in the CRM as a checklist and let the assistant draft only from unticked items. The same rule applies when it simplifies a condition into plain language: the meaning belongs to the lender and the assistant only changes the wording. If you cannot tell whether the meaning shifted, send the lender's wording instead.
CRM notes from a call: a template that keeps the record clean
Notes are what you will be shown in a complaint, an audit or a dispute over who said what. Give the assistant a structure that separates what the borrower said from what you have seen:
- Stated by the borrower: income, debts, property, timeline, in their words and marked unverified.
- Documented so far: only what has actually arrived.
- Open items: ours and theirs, each with a date.
- Rate or payment discussed: filled in by the officer only, with what was said and when.
- Consent and channel: recording consent, preferred channel, any do-not-contact request.
- Next step and date.
Here is a realistic slip. On a call, a borrower says: "I make about 6,200 a month before tax, I've got 1,150 a month in debt payments, and it's the first place I've bought on my own." The assistant's note reads: "First-time buyer, qualifies for a standard loan, DTI 38%." (DTI is the debt-to-income ratio.) Three things are wrong. "First-time buyer" is a programme label the borrower never claimed; they said they had not bought alone before, and they may have owned a home jointly. "Qualifies" is a conclusion nobody has reached. And the 38% cannot be derived from what was said: 1,150 divided by 6,200 is 18.5% before any housing payment, and the housing payment is unknown. The corrected note: "Stated, unverified: gross income about 6,200 a month; debt payments about 1,150 a month; first purchase made alone, previous joint ownership to be asked. No eligibility view given."
If a tool records or summarises the call itself, settle consent first. Is it safe to let AI listen to mortgage advice calls? goes through that decision, and the short version is to tell the borrower at the start, get a clear yes, record it in the consent field and keep the transcript only as long as your retention rules say.
The compliance pre-check: a second reader, never a signatory
A second prompt reads a draft against your banned list. It will not catch everything and it must never approve anything, but it catches the slip you stopped seeing after the fortieth message that week.
You are a first-pass compliance reader for a mortgage broker's outgoing messages. You do not approve anything. Read the message and list only problems, each with the exact words and the rule family it may touch.
Check for:
1. Any rate, APR, payment amount, fee amount, down payment figure or loan term.
2. Any statement that the person is approved, pre-approved, qualified, eligible or guaranteed anything.
3. Superlatives or comparisons: lowest, best, better than, save, no cost.
4. Urgency: rates rising, lock now, limited time, today only.
5. Any reference to race, colour, religion, national origin, sex, marital status, age, receipt of public assistance, the neighbourhood a person lives in, or any other characteristic our fair-lending rules protect.
6. Missing lines: licence number, opt-out wording, company name.
7. Any claim not supported by the facts supplied.
Output a numbered list of flags, or the single word CLEAR. Never rewrite the message.
Run on a campaign line (invented): "You're pre-approved for up to $400,000 and rates are at a three-month low. Lock in today!" It should return four flags: the pre-approval claim with a stated amount, the unsupported rate comparison, the urgency, and the missing licence and opt-out lines. CLEAR means only that the reader found nothing. Approval stays with a named person, and for campaigns that may mean your compliance contact or the lender's marketing team.
Past-client and partner campaigns need an approval step
One-to-one follow-up and campaigns are different animals. A message to 220 past clients about refinancing is advertising on any sensible reading, so it gets more scrutiny than a document chaser. The flow I would use:
- Write the template once, with the assistant if you like, from approved company and lender wording.
- Run the pre-check and fix every flag.
- Send it to your compliance contact, or to the lender's marketing review where that is required.
- Segment by consent: only contacts with consent for that channel and no do-not-contact flag.
- Send in batches and check replies and opt-outs daily for the first week.
- Archive the final version, the segment rules and the send dates.
Anything involving a referral partner, shared marketing or a payment between businesses has rules of its own. The checklist for referral and marketing rules and referral partner updates drafted with AI go through them, and the safe boundary for the assistant is simple: it may draft a status update about a shared client, and it never drafts anything about what a partner receives.
What to log for every AI-assisted message
A message you cannot reconstruct is a message you cannot defend. The log can live as a note on the contact record, as long as every entry has the same shape:
| Field | Example entry (illustrative) |
|---|---|
| Date and time | 12 Nov, 10:42 |
| Contact | Contact ID 20418, stage: documents outstanding |
| Template and prompt | Document chaser v3, prompt v5 |
| Facts supplied | CRM fields only; no pasted documents |
| Pre-check result | Two flags, both fixed before sending |
| Approved by | Officer initials |
| Consent | Email, web form, 3 Oct, wording v2 |
| Final text | Attached to the contact record |
Keep each entry for the period your rules and lenders require. One widely used set of advertising rules sets 24 months after last use as its minimum for communications, scripts and training material, and some licensing and investor requirements run longer. Keep prompt versions as well: if a prompt shaped what a borrower read, the version that did is part of the record.
A 30-day trial on 35 files, with the sums
Take a two-officer brokerage with 35 active files and 220 past clients. Each active file gets about three messages a week, so 105 a week. These timings are assumptions; swap in your own.
- By hand at 5 minutes each: 105 x 5 = 525 minutes, about 8.75 hours a week.
- With a draft from CRM fields (1 minute) plus the banned-list check and pre-check (2 minutes): 105 x 3 = 315 minutes, about 5.25 hours a week.
- Difference: 3.5 hours a week, or 14 hours over four weeks.
- Less about 3 hours of set-up (fields, templates, banned list) and 2 hours of audit: roughly 9 hours net over the month.
The 220 past clients cost one reviewed template and one approval a quarter, not 220 drafts. In week four, read 20 sent messages against the banned list and count three things, each with a target of zero: banned claims found after approval, messages missing the licence line, and contacts messaged without consent. Also note how many drafts you rewrote heavily, because that is where the prompt needs work.
If any of the three zeros fails, pause the sequences, fix the field or the prompt, and rerun the week before going on. If you would rather map this with someone, that is the kind of job my AI implementation consultation covers.
Loan officer questions on AI follow-up
Can AI reply to borrower texts automatically?
Only with fixed acknowledgements that were approved in advance, such as confirming a message was received. A free-running auto-reply will sooner or later state a rate, an eligibility view or a deadline, and rules on automated texts and calls are their own subject. Anything that answers a question about the loan should be drafted for a person to approve.
Can AI decide which leads I call first?
It can order contacts by behaviour you can explain, such as replied, uploaded documents or opened a message. Keep protected characteristics, and proxies for them, out of the rules, write the ordering rules down, and review now and then who is getting which message. Fair-lending rules cover how you treat prospects, not only how you decide applications.
Do I need to keep my prompts?
Treat a prompt like a script: if it shaped what a borrower read, keep the version that did. One widely used set of mortgage advertising rules makes you keep materially different communications, sales scripts and training material for 24 months after last use. Your licence, lenders or investors may ask for longer, so confirm the period with compliance.
What borrower data is safe to paste into an assistant?
Use a first name and the CRM fields the prompt needs. Keep national ID numbers, account numbers, full credit reports and complete bank statements out of any tool your compliance contact has not approved. Check the plan's terms on training, retention and who can read inputs, and check whether your lenders allow their documents in third-party tools.
Further reads
- AI Loan Processing: Catch Missing Documents Before Underwriting — Catch missing documents before underwriting, once the file leaves your follow-up.
- Can Loan Officers Use AI Videos for Market Updates? Rules to Follow — The rules for AI-made market update videos before you post one.
- What Mortgage Brokers Can Automate With AI, and What Stays Advice — Where automation ends and advice begins in a broker's day.
- HubSpot for Mortgage Brokers: Leads, Referrers and Case Stages — Case stages and referrer tracking if your CRM is HubSpot.
- Mortgage Broker Social Media With AI: A Month of Compliant Posts — A month of social posts that pass the same banned-claims check.
- AI Tools for Mortgage Lenders: Processing, Disclosures, Audit Trails — The lender side: processing, disclosures and audit trails.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: published mortgage advertising and fair-lending rule text (triggering terms, misrepresentation categories, discouragement, 24-month records rule), public licensing guidance on displaying licence identifiers. Rules differ by jurisdiction and lender, so confirm each point with your compliance contact.