To spot remortgage opportunities, start with a clean list of every client's deal end date, early repayment charge end date, lender, balance and estimated property value. Spreadsheet formulas flag anyone inside your contact window; AI earns its place by pulling missing dates out of old offer documents and drafting personalised review invitations. Whether and how a client remortgages stays your advice.
Most brokers who lose clients at the end of a deal don't lack a tool; they lack the data. The dates exist, but they sit in PDFs of mortgage offers, in case notes typed three years ago, or in a CRM field half the team never filled in. Check first whether your mortgage CRM already has product-expiry alerts, because many do. If it does, AI's job is filling the fields those alerts depend on. If it doesn't, a spreadsheet plus AI does the job well enough for a bank of a few thousand clients.
Six triggers worth watching in a client bank
| Trigger | Data needed | When to act | What you'd offer |
|---|---|---|---|
| Deal ends | Product end date, lender | Inside the lender's rate-lock window, usually some months ahead | A review before they drift onto the lender's standard rate |
| Early repayment charge ends | ERC end date, if different from product end | A few weeks before | A review now that switching carries no penalty |
| Loan-to-value band improves | Balance, estimated value, lender's bands | When the estimate crosses a band threshold | A check on whether a lower band improves the options |
| Circumstances change | Notes from annual reviews or client emails | When noted | A review: new baby, income change, plans to move |
| Term or interest-only period ends | Repayment type, term end | Well ahead, often a year or more | A repayment strategy conversation |
| Rates move relative to theirs | Client's current rate | When the market moves meaningfully | Only where the ERC and costs make a switch worth checking |
The first two produce most of the opportunities. The third is the one competitors miss: a client who borrowed at 88% loan-to-value four years ago may be under 75% now through repayments and price growth, without having noticed.
The fourth trigger lives in free text, which is exactly what AI reads well and formulas can't. Paste the last annual review note (by client reference, no name) with a narrow instruction:
From the review note below, list only changes in circumstances the
client actually stated that could affect their mortgage. For each,
quote the words used. Do not predict income, affordability or
intentions that were not stated.
NOTE: [client ref] [paste note]An illustrative result for a note that read "Second child due in April. Partner going back part-time after leave. Mentioned possibly doing a loft conversion next year if finances allow":
1. Family change - "Second child due in April"
2. Working pattern change - "Partner going back part-time"
3. Possible home improvement - "possibly doing a loft conversion
next year if finances allow"
4. Household income likely to fall - inferred from item 2Items 1 to 3 are useful: the loft conversion could mean further borrowing, and the working pattern matters for any new application. Item 4 is the model doing the adviser's job without the facts. Part-time after leave may be more hours than during it. Strike inferences like that and keep the three quoted items as prompts for the review conversation.
Audit what your records actually hold
Export the client bank from your CRM and count what's missing before planning anything. An illustrative audit for a brokerage with 900 completed cases still on its books:
| Field | Filled | Missing | Where the missing data probably is |
|---|---|---|---|
| Lender | 897 | 3 | Case notes |
| Product end date | 641 | 259 | Mortgage offer PDFs |
| ERC end date | 312 | 588 | Offer PDFs (often the same as product end) |
| Loan amount at completion | 880 | 20 | Offer PDFs |
| Property value at completion | 702 | 198 | Valuation reports, fact finds |
| Marketing consent recorded | 655 | 245 | Client agreements, email history |
A quarter of the bank has no product end date: that's roughly 250 clients the business can't currently spot in time. That gap, not the lack of a clever tool, is what the rest of this tutorial fixes. If the CRM itself is a mess of duplicates and old contacts, sort that first; cleaning up a messy CRM with AI covers it.
Recover missing dates from old case files
This is where AI saves the most time. A person reading 250 offer documents to find three dates in each is a week's work. An AI assistant on a business plan can extract them in an afternoon, with a person checking a sample. The prompt, run one document at a time or in small batches:
From the mortgage offer below, extract:
- client reference (from the file name)
- lender
- loan amount
- initial rate and rate type (fixed, tracker, discount)
- the date the initial rate period ENDS
- the date early repayment charges END
- property value used by the lender
- whether the loan has more than one part
Quote the sentence each date came from. If a value is not stated,
write NOT STATED. Do not calculate dates.
DOCUMENT: [client ref]
[paste or attach offer]For one case file, the result looks like this (illustrative):
Client ref: C-0417
Lender: [lender name]
Loan: $312,000
Rate: 4.19% fixed
Initial period ends: 30/11/2026 - "The fixed rate applies until
30/11/2026"
ERC ends: NOT STATED
Property value: $395,000
Parts: 2 - "Part A $262,000... Part B $50,000 (additional
borrowing)"
Offer valid until: 14/03/2022Two things to check. The two-part loan matters: Part B may have its own rate and end date, and the extraction only gave one. Go back to the document for the second part's terms. And the "offer valid until" date is a classic trap: in other outputs from the same batch, the model sometimes put that date in the "initial period ends" field when the fixed-rate wording was unusual. Quoting the source sentence is what lets a checker spot it in seconds. Check a random 1 in 10 extractions against the documents, and every one where the quoted sentence doesn't obviously match the field.
Offer documents hold personal and financial data. Use only a business plan that doesn't train on your content by default, identify files by client reference rather than name, and delete the working chats once the data is back in your CRM. For larger volumes, AI document processing for PDFs and scans compares dedicated extraction tools.
Build the watchlist
With the dates recovered, a spreadsheet does the flagging. Columns: client reference, adviser, lender, product end date (D), loan balance estimate (E), estimated current value (F), loan-to-value at completion band (I), ERC end date (J), marketing consent, last contact date. Then formulas:
Days until deal ends:
=D2-TODAY()
Window flag (adjust 180 and 270 to your lenders' windows):
=IF(D2-TODAY()<0,"EXPIRED",IF(D2-TODAY()<=180,"IN WINDOW",
IF(D2-TODAY()<=270,"PREPARE","")))
Estimated loan-to-value now:
=E2/F2
Current band:
=IF(G2<=0.6,"60",IF(G2<=0.75,"75",IF(G2<=0.85,"85","90+")))
Band moved since completion:
=IF(H2<>I2,"BAND MOVED","")
ERC ends within 90 days:
=IF(AND(J2>=TODAY(),J2-TODAY()<=90),"ERC ENDS SOON","")The balance and value estimates are estimates. Balance can come from the original loan, rate and term (your CRM may calculate it, or an AI assistant can write the formula, which you then spot-check against three client statements). Value can come from a house price index or an automated valuation. Label both columns "estimate", because they're for spotting who to call, never for quoting to a client.
A typical balance formula an assistant will suggest, with the original loan in A2, the annual rate in B2, months since completion in C2 and the term in years in D2:
Estimated balance now (repayment loans only):
=-FV(B2/12,C2,PMT(B2/12,D2*12,A2),A2)It's a fair estimate for a repayment mortgage that has stayed on one rate, and it ignores overpayments and rate changes. The mistake that shows up in practice is running it on every row. Interest-only loans don't reduce, but the formula shrinks them anyway, so an interest-only client appears to have dropped two loan-to-value bands when the balance hasn't moved. Add a repayment-type column and set the estimate to the original loan wherever it says interest-only. On part-and-part loans, estimate only the repayment part.
Once the estimates are in, check one flagged row by hand. C-0662 in the table below borrowed $243,000 against a lender valuation of $270,000, which is 90%. The formula puts the balance at about $226,000 now, and the index-based value estimate is $305,000: 226,000 divided by 305,000 is 0.741, or 74%. That's two bands down, and a sum the adviser can sanity-check in ten seconds before deciding it's worth a conversation. If you want help writing formulas like these, using ChatGPT with Excel covers the prompting.
Filtered in September, the watchlist might show (illustrative):
| Client | Deal ends | Flag | LTV then / now | Consent | Action |
|---|---|---|---|---|---|
| C-0417 | 30 Nov 2026 | IN WINDOW | 79% / 71% | Yes | Review invitation this week; check Part B terms |
| C-0288 | 31 Jan 2027 | IN WINDOW | 85% / 83% | Yes | Review invitation |
| C-0662 | 31 May 2027 | PREPARE | 90% / 74% | No | Service message only; adviser to decide |
| C-0105 | 15 Oct 2026 | IN WINDOW | 60% / 52% | Opted out | No contact |
C-0662 is the kind of row that justifies the exercise: two band moves since completion, and the deal doesn't end for eight months, so there's time to prepare properly.
Contacting clients: consent, timing and wording
Three rules before AI drafts a single message. First, respect consent: a client who opted out of marketing doesn't get a "have you thought about remortgaging?" email, however valuable the opportunity. Whether a message about the deal you arranged counts as a service message or as marketing depends on your terms, what you told the client and the electronic marketing rules you're under, so agree the line with your compliance adviser once and apply it every time. Second, time it to the lender's window, not your sales calendar. Third, invite a review; don't recommend a product in an automated message.
Here's the AI's first attempt and the version that went out:
AI's first draft: "Great news! Rates have dropped and your fixed rate ends soon. You could save hundreds a month by switching now. Book your remortgage today!"
After editing: "Your fixed rate with [lender] ends on 30 November. Many lenders let you secure a new rate a few months before a deal ends, so now is a good time to review your options, whether that's a new deal with your current lender or elsewhere. If you'd like me to do that, reply to this email or book a 20-minute call here [link]. If your plans have changed, for example if you're thinking of moving, let me know and we'll factor that in."
The first draft makes an unsupported savings claim, implies advice before any assessment, and would struggle with your regulator's rules on financial promotions. The second states facts from your records, offers a review and leaves the advice where it belongs. Have your compliance lead approve each template once; then the AI only fills in the date, lender and link. What mortgage brokers can automate and what stays advice draws the wider line. If you want different wording for first-time buyers and long-standing clients, write a separate approved template for each rather than letting the AI vary the wording freely.
Two situations need their own approved wording, and a rule that the AI never drafts for them unprompted:
- Deals that have already ended. The first extraction run always turns up some EXPIRED rows: clients who may be on the lender's standard rate now. A deadline-style message is wrong for them. An illustrative approved version: "Our records show your fixed rate with [lender] ended in [month]. If you haven't arranged a new deal since, you may now be on the lender's standard rate. I'm happy to review where you are, with no obligation. Reply here or book a time [link]." Send these first; they're the clients paying most right now.
- Joint borrowers whose notes suggest a separation. An automated email addressed to both is the wrong move and can cause real harm. Flag any case note that mentions separation, divorce or one borrower moving out, take it out of the automated run, and let the adviser decide how to make contact.
A two-adviser brokerage's first quarter
Go back to the illustrative brokerage from the audit: two advisers, an administrator and 900 cases. In the first month, the administrator runs the extraction on the 259 cases without an end date, spending about 12 hours including checking. 231 end dates are recovered; 28 files are incomplete and go to the advisers to check with clients at their next contact.
The watchlist then shows 146 clients whose deals end in the next nine months, 38 of whom the brokerage would previously have missed entirely, plus 22 clients whose estimated loan-to-value had crossed a band. After consent filtering, 131 are contacted over the quarter using the approved template, at the point each enters its window.
The outcomes to track are review meetings booked, cases submitted, and clients who went elsewhere. Don't pre-judge the numbers: the value shows up over a full year of deals ending, not a single quarter. What the brokerage knows for certain after three months is that no client in the bank will reach the end of a deal without being asked whether they'd like a review. The same watchlist feeds your annual reviews, which preparing for annual client reviews with AI covers.
Mistakes that make a watchlist unreliable
- The wrong date in the right column. Offer expiry, completion date and product end date all appear in the same documents. Quoted source sentences and sampling catch most of these.
- Multi-part loans treated as one. Additional borrowing often has its own rate and end date. Give each part its own row.
- Clients who already moved. Some will have remortgaged through another broker or directly with the lender. Mark them when you find out rather than deleting them; they may come back next time.
- Duplicate clients. Joint borrowers entered twice get two emails. Deduplicate on the case, not the person.
- Stale value estimates. Refresh them quarterly at most, and never quote them.
- Nobody owns the list. A watchlist that isn't reviewed every month becomes another spreadsheet nobody trusts.
A monthly hour that keeps the watchlist trustworthy
Make the data complete at the source: when a case completes, the administrator records product end, ERC end and each loan part before closing the file. Once a month, one person refreshes the estimates, re-sorts the watchlist and assigns the new "IN WINDOW" clients to advisers. That's about an hour a month for a bank this size, and it's the difference between a one-off clean-up and a retention routine.
Use part of that hour to test the watchlist against what actually happened. Take every client whose deal ended last month and ask one question of each: were they contacted before their window opened? An illustrative March check: 14 deals ended; 12 clients had been invited four or more months ahead; one had opted out, correctly skipped; one was missed. The missed one had "NOT STATED" in the ERC column and a product end date typed as 2062 instead of 2026, so the formula thought it was decades away. One finding like that is worth a new rule: a column that flags any end date more than ten years out, because no fixed deal in the bank runs that long.
If your CRM is due for replacement, choosing a mortgage CRM with AI built in lists the retention features worth paying for.
Further reads
- From Enquiry to Fact Find: An AI Workflow for Mortgage Brokers — Capture complete data at the start so the watchlist stays full.
- Best AI Tools for Mortgage Brokers at Each Stage of a Case — Tools for each stage of a case, including retention.
- Can AI Keep Mortgage Clients Updated During an Application? — Keep clients informed once a remortgage case is under way.
- AI Governance Checklist for Small Financial Advice Firms — The governance checks an advice firm needs around AI.
- How to Win Back Lapsed Customers With AI-Personalised Emails — Re-engage clients who went elsewhere last time.
- Is It Safe to Let AI Listen to Mortgage Advice Calls? — Whether AI should be near the advice conversation itself.
- Do You Need a Lot of Data to Use AI in Your Business? — How much data each common AI task needs, from three example emails to two years of sales history, and why scattered data is the bigger problem.
- 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 Team plan pages (business-data defaults), checked September 2026; general spreadsheet date functions.