From Enquiry to Fact Find: An AI Workflow for Mortgage Brokers

Mortgage brokers can use AI at five hand-offs between enquiry and fact find: triaging the enquiry, replying and booking the first call, sending a document list matched to the case, extracting figures from payslips and statements into a draft fact find, and writing the adviser's call brief. The adviser still asks the questions, checks every figure and gives the advice.

This stretch is the safest place to start because nothing in it is advice yet. It's also where cases leak: an enquiry answered the next afternoon instead of within the hour, a client who sends four of the nine documents and goes quiet, an adviser re-keying payslip figures at 9pm. The line to hold is that AI organises information and never tells the client what they can borrow or which lender suits them; what mortgage brokers can automate and what stays advice sets out that boundary in detail.

Follow me on Instagram@sagnikteaches

The five hand-offs at a glance

Hand-off 1: triaging the enquiry in the first minute

Enquiries arrive messy. The triage step turns them into a consistent record your CRM can use and flags what's urgent. Run it through an automation (Zapier or Make watching a shared inbox or form) calling a business-plan AI model, or through your CRM's own AI features if it has them. The prompt:

Connect on LinkedInSagnik Bhattacharya

Classify this mortgage enquiry. Use only what the enquiry says.
Return:
- case_type: purchase / home mover / remortgage / further
  borrowing / investment property / unclear
- applicants: number, if stated
- employment: per applicant, as stated (do not infer)
- deadline: any date or timescale mentioned
- urgency: high (deadline under 8 weeks or offer accepted),
  normal, or unknown
- missing: key facts not given (deposit, income, property
  value, current lender, etc.)
- flags: anything an adviser should see first
Do not estimate borrowing, rates or eligibility.

Enquiry:
[paste]

Here's a real-looking enquiry and what comes back (illustrative):

Subscribe on YouTube@codingliquids

"Hi, me and my partner are looking to move, we've had an offer accepted on a place for 385,000. Our current mortgage has about 2 years left on the fixed rate I think? I'm self employed since last year. Can you help? Need to move quick as the sellers want to complete within six weeks."

case_type: home mover
applicants: 2
employment: applicant 1 self-employed, about 1 year;
            applicant 2 not stated
deadline: completion within about 6 weeks
urgency: high
missing: deposit/equity, incomes, current property value and
         sale status, current lender and balance
flags: possible early repayment charge on current fixed rate;
       short self-employed history; tight timescale

Useful, with one correction. "Self-employed since last year" could mean eight months or twenty; the output's "about 1 year" turns vagueness into a figure, and trading history changes which conversations the adviser needs to have. Change the prompt rule to "quote employment wording exactly" and the problem goes away.

A second enquiry shows why the flags field matters more than the case type. This one arrives through the website form:

"Fixed rate ends in March. We'd like to remortgage and borrow an extra 30k or so to pay off a couple of credit cards and the car loan, as the payments are getting a bit much. Both employed."

case_type: further borrowing (with remortgage)
applicants: 2
employment: "Both employed"
deadline: fixed rate ends in March
urgency: normal
missing: current balance and lender, property value, incomes,
         amounts owed on each debt
flags: debt consolidation requested; "payments are getting a
       bit much" may indicate financial strain - adviser to
       review before any reply is sent

The case type alone would have sent this down the plain-remortgage route, where replies go out automatically. The flag stops that. Consolidating unsecured debt into a mortgage needs a careful conversation about the costs of spreading it over a longer term, and a client describing payments as "getting a bit much" deserves a person's reply, not a template. Make "debt consolidation" and signs of strain two of the flags that always hold the automatic reply.

The triage record then drives routing, which is where most of the time saving in this first step actually comes from. Three rules cover most small brokerages: urgent cases (a deadline under eight weeks or an offer accepted) go to whichever adviser has the earliest free slot and trigger a phone notification; complex cases (self-employed under two years, adverse credit mentioned, investment property portfolios) go to the adviser who handles them; everything else joins the normal queue. Write the triage fields into your CRM as a new lead so nobody retypes them, and keep the original enquiry attached underneath, because the client's own words often matter later.

Search the CRM before creating anything. In an illustrative first month, a past client whose five-year fix was ending emailed from a new personal address. Triage created a fresh lead, a different adviser picked it up, and the client got a "thanks for your enquiry, first step is a 20-minute call" reply from a firm that had arranged her last two mortgages. Add a matching step before the lead is created: look up the name, phone number and any property address in the CRM, and if there's a likely match, attach the enquiry to the existing client and send it to their adviser. Returning clients are the enquiries you least want to treat as strangers.

Check the labels for the first month. If one in ten is wrong, look at which kind: a model that keeps calling further-borrowing requests "remortgage" needs an example of each in the prompt, not a different tool. Enquiries that arrive as Instagram or Facebook messages can feed the same triage if your inbox tool forwards them, but keep automated replies on those channels just as short.

Hand-off 2: a reply within the hour and a booked call

Most enquirers contact more than one broker, so speed matters more than polish. Have AI draft the reply from the triage output and a fixed template, with a booking link. Microsoft Bookings is included in Microsoft 365 Business Basic, Standard and Premium, and most mortgage CRMs have their own scheduler.

Before, a typical reply sent the next morning read: "Thanks for your enquiry, one of our advisers will be in touch." After, sent eleven minutes after the enquiry:

"Thanks for getting in touch. With completion in about six weeks we'll want to move quickly, so the first step is a 20-minute call to understand your plans, your current mortgage and your income. You can book a time that suits you here: [link]. If you can, have a rough idea of your deposit and your current mortgage balance to hand. Nothing you tell us on that call commits you to anything."

Notice what it leaves out: no rates, no "you should be fine", no lender names. Decide which enquiry types can send automatically (a plain remortgage enquiry, say) and which wait for a person to approve (anything flagged urgent or complex). If the client books nothing within a day, a single follow-up is enough; a second automated nudge the same week starts to feel like a call centre.

Hand-off 3: a document list that matches the case

A generic "please send the following" list is why clients send the wrong things. After the discovery call, generate the list from the case facts. Filled in for the illustrative couple above:

Documents for your home move: [applicant 1, self-employed]
[ ] Photo ID (passport or driving licence), in date
[ ] Proof of address dated in the last 3 months
[ ] Accounts or tax calculations for each year you've traded,
    plus your accountant's contact details
[ ] Last 3 months' business bank statements
[ ] Last 3 months' personal bank statements
[ ] Latest statement for your current mortgage
[ ] Proof of deposit: savings statements, or details of the
    sale of your current home
[applicant 2: we'll confirm after you tell us about your work]
Upload here: [secure portal link]. Please don't email these.

Schedule chasers on day 2, day 5 and day 9, each listing only what's still outstanding. A day-5 chaser for the same couple, written from the portal's live status:

"Thanks for uploading your ID, proof of address and personal bank statements. To keep your six-week timescale on track, we still need: your accounts or tax calculations, the last three months' business bank statements, and your latest mortgage statement. If the accounts aren't finished yet, just reply and tell us; that's useful to know now."

The last sentence does more work than it looks. It gives a client who is stuck a way to say so without feeling they've failed.

After day 9, stop automating and have a person call: silence at that point usually means a problem (the accounts aren't done, a partner is unsure) that no reminder will fix.

Hand-off 4: extracting figures into a draft fact find

This is where the real time goes, and where AI mistakes are most expensive. Use a tool approved for client data: your CRM's document features, or Microsoft 365 Copilot or another business-plan assistant inside your own tenant. Ask for figures with their source:

From the attached documents for applicant 1, extract into the
table below. For every value give the document name, page and
the exact line it came from. If a value isn't present, write
"not found". Do not calculate totals or averages.

Fields: employer or business name; pay frequency; gross pay per
period; net pay per period; overtime/bonus/commission per
period (separately); pay date; regular outgoings on bank
statements (payee, amount, frequency); ID document type and
expiry date.

An illustrative slice of the result for an employed applicant, and the mistake it contained:

The year-to-date mix-up is the classic payslip error, because many payslips print "Gross Pay" twice, once for the period and once cumulative. The source column is what made it a ten-second catch rather than a mis-keyed application. Never let extracted figures flow straight into anything a lender sees; they pre-fill a draft fact find that the adviser confirms with the client.

Company directors produce a different error. For an illustrative applicant who runs a small limited company, the extraction from two years' accounts and tax calculations returned "income, year 2: 48,000". The source lines showed that figure was salary of 12,000 plus dividends of 36,000, added together by the model despite the instruction not to total anything. Meanwhile the company's net profit for the year, 61,000, sat on the accounts' profit and loss page and wasn't extracted at all, because "net profit" wasn't in the field list. Lenders look at these figures in different combinations, so the fact find needs each one on its own line: salary, dividends and net profit, per year, each with its source. Add all three to the field list for any director, and treat a single "income" figure for a self-employed applicant as a sign the extraction has merged something.

Bank statements hold the other thing worth extracting carefully: commitments the client didn't mention. Ask for a list of regular payments with payee, amount and frequency, and a separate list headed "payments the adviser may want to ask about". In one illustrative run, that second list held a new monthly payment of 145.00 to a car finance company that started in July, and three payments to a buy-now-pay-later provider. Neither appeared in the enquiry or on the discovery-call notes. The AI lists them without comment; whether they matter, and how to raise them, is for the adviser at the fact find.

AI won't judge whether a document is genuine, whether income is sustainable or whether a lender will accept it. Those stay with your existing checks and the adviser's judgement.

Hand-off 5: the adviser's call brief

The day before the fact find, generate a one-page brief from the triage record, emails, call notes and the draft fact find: what's confirmed, what's missing, what conflicts, and questions to ask. An illustrative extract:

Confirmed: 2 applicants; purchase price 385,000; completion
target about 6 weeks; applicant 1 self-employed.
Missing: applicant 2's income; deposit amount; sale status.
Conflicts: enquiry said "2 years left on the fix"; mortgage
statement shows fixed period ends in 14 months.
Ask: when did applicant 1 start trading, and are accounts filed?
Is the current home under offer? Any change expected to
applicant 2's work (the email of 12 Sep mentions "after the
baby")?

That last line is the kind of thing the brief is for. A planned family change affects income and possibly circumstances an adviser should explore carefully. The brief lists it as a question; it doesn't draw a conclusion. The fact-find meeting itself can then be captured with a note taker, which brings its own questions covered in whether it's safe to let AI listen to mortgage advice calls.

An illustrative two-adviser brokerage, before and after

Take a brokerage with two advisers and one administrator handling about 60 enquiries a month, of which about 35 reach a fact find. Before the workflow, admin between enquiry and fact find averages around two and a half hours per case: replying, booking, requesting and chasing documents, keying figures, preparing. That's roughly 87 hours a month.

With the five hand-offs running, a realistic figure is 70 to 80 minutes per case, most of it the extraction check and the brief read-through. Call it 44 hours a month: about 43 hours returned, a little over a week of one person's time. Costs on list prices: Microsoft 365 Copilot Business for three people at $21 a user a month on annual billing is $63, and Zapier Professional at $19.99 a month on annual billing, or $29.99 monthly, covers modest volumes. Five action steps per case across 60 enquiries is about 300 tasks a month, well inside the 750 included, provided each AI step counts as one task; that holds on the cheapest AI tier or with your own API key, while higher tiers use 3 or 5 tasks a run. If your CRM does triage and extraction itself, the Zapier line may disappear. For choosing the rest of the stack, see the best AI tools for mortgage brokers at each stage, and if your CRM is up for renewal, choosing a mortgage CRM with AI built in.

Where the pre-advice workflow breaks

• Joint applicants merged. Two payslips from different employers become one applicant's income. Extract per applicant, in separate runs, and name the applicant in the prompt.
• Scanned and photographed documents. A payslip photographed at an angle on a kitchen table produces more misreads. Ask clients to upload PDFs where possible, and expect to check photos more carefully.
• Documents arriving by the wrong channel. Clients will send bank statements by message or personal email regardless of what you ask. Have a rule for moving them into the portal and deleting the originals.
• Chasers that don't know what arrived. A reminder asking for ID the client uploaded yesterday destroys trust quickly. Chasers must read the live document status, not a fixed schedule.
• Auto-replies that drift. Someone edits the template to add "rates from X%". Lock the template and review it quarterly.
• Self-employed income. AI can read accounts and tax calculations, but which figures a lender uses is a judgement. Extraction gives the adviser the numbers; it doesn't choose between them.

Once the workflow runs, the same pattern extends forward into the application stage, where AI can keep clients updated while the case progresses.

Questions brokers ask about automating the pre-advice stage

Can the AI reply to enquiries with indicative rates?

Don't let it. Rates, eligibility and lender names in an automated reply can read as advice or a promise, and they go out of date within days. Keep automated replies to acknowledgement, next steps and a booking link. Anything about what the client might borrow or which lender might suit belongs in a conversation with a qualified adviser.

Is it safe to put payslips and bank statements into an AI tool?

Only into a tool your firm has approved for client data, on a business plan that doesn't train on your content, ideally inside your CRM or your Microsoft or Google tenant. Never into a free consumer chatbot. Check where documents are stored, how long they're kept, and whether you can delete them when the case closes.

Will AI spot a fake payslip?

Don't rely on it. A language model reading a document extracts what's printed; it isn't designed to judge authenticity, and a well-made forgery reads the same as a real one. Keep your existing fraud checks and lender verification. AI may flag inconsistencies, such as totals that don't add up, and those flags are worth a look.

Further reads

• Using AI to Spot Remortgage Opportunities in Your Client Bank — The same automation mindset applied to existing clients.
• How to Reduce No-Shows With AI Reminders and Automatic Rebooking — Reminder patterns that stop booked fact-find calls going missing.
• Questions to Ask Before Buying AI That Touches Client Data — Checks before any tool reads client payslips and statements.
• Why Is AI Bad at Maths? What to Check in Quotes and Invoices — Why extracted figures still need a human check.
• How to Manage Your Inbox With AI: Triage, Drafts, and Follow-Ups — Triage the rest of the inbox the same way.
• AI Note Takers for Financial Advisers: Compliant Meeting Records — Compliant notes for the fact-find meeting itself.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Learn, Microsoft Bookings availability by plan; Microsoft 365 Copilot Business and Zapier pricing from vendor pages, checked September 2026.