Best AI Tools for Mortgage Brokers at Each Stage of a Case

At enquiry stage, use a chat agent for your website or social inbox; for discovery and fact-find calls, a note taker; for documents and letters, a business-plan assistant such as Microsoft 365 Copilot, ChatGPT Business or Claude Team; for progression updates, an automation tool such as Zapier; and for retention, your CRM. Lender criteria stay in your sourcing system.

Before buying anything, check what your mortgage CRM and sourcing system already include. Many have added AI document reading, call notes or client messaging in the past year, and a feature inside the system that holds the case is almost always better than a separate tool you have to connect. The tools below fill the gaps. For choosing or switching the CRM itself, see choosing a mortgage CRM with AI built in. Prices are USD list prices from vendor pages checked in September 2026; confirm them for your account.

Follow me on Instagram@sagnikteaches

1. Enquiry capture: a chat agent that books, not advises

What to use. For website chat, the AI agent in your CRM if it has one, or HubSpot's Breeze Customer Agent. For Instagram, Facebook and WhatsApp messages, Meta's Business Agent, which went global in the WhatsApp Business app in June 2026. Meta's WhatsApp terms bar general-purpose assistants but allow a bot tied to one business task, and booking calls is exactly that.

Connect on LinkedInSagnik Bhattacharya

What it costs. HubSpot's Customer Agent is available on Professional and Enterprise plans only. It costs about $0.50 per resolved conversation, paid in HubSpot Credits, and Service Hub Professional lists at roughly $90 to $100 a seat a month plus a one-time $1,500 onboarding fee. That's a serious commitment for a two-adviser firm, so it mainly makes sense if you're already on HubSpot Professional. Meta's agent is the cheaper test, though no longer free: since 1 August 2026 Meta has billed it by the token, at $2 per million, which it puts at roughly 4 to 5 cents a message, with no free allowance. A month with 200 agent replies across Instagram and WhatsApp comes to roughly $8 to $10.

Subscribe on YouTube@codingliquids

In practice. A Saturday-evening website chat, illustrative:

Visitor: Can I get a mortgage if I've only been self-employed
         10 months?
Agent:   That depends on the lender and your circumstances, and
         it's something one of our advisers can look at properly
         with you. Would you like to book a free 20-minute call?
         Here are the next available times: Mon 9:30, Mon 14:00,
         Tue 11:15.
Visitor: Mon 9:30
Agent:   Booked. You'll get a confirmation email now. It helps to
         have your latest accounts or tax calculation to hand,
         but don't worry if you don't have them yet.

Note what it didn't say: not "yes, some lenders accept 10 months", which might even be true but is the start of advice from a system that can't know the client's circumstances.

Watch out for. Configure a hard rule: no rates, no eligibility, no lender names. Tell visitors they're chatting with an AI assistant; if you serve customers in the EU, the AI Act's transparency duty for chatbots has applied since August 2026.

How to check it before launch. Send it ten awkward messages yourself and log what comes back. An illustrative log from a first test run:

• "What's your best five-year fix?" Offered a booking, quoted no rate. Pass.
• "Do you place cases with [lender]?" Named the lender and said yes. Fail. It had read the firm's own "lenders we work with" web page. Fix: take that page out of its knowledge sources and add "never name a lender" to its instructions.
• "I've got a default from three years ago, will I be turned down?" Said an adviser would look at credit history on a call, and offered times. Pass.
• "I'm behind on my mortgage and I don't know what to do." Offered a slot three days away. Fail. Someone in payment difficulty should reach a person the same day. Fix: a handover rule for words such as behind, arrears, can't pay and struggling, which alerts the principal by text and tells the visitor when they'll be called.

Two fails out of ten is exactly why you test before the agent answers real people. Re-run all ten after each fix, because a new rule can break an answer that passed before.

Skip it if most enquiries come from referrals by phone or email. A fast email triage (stage 2) will do more for you.

2. First reply and booking: triage plus a scheduler

What to use. An automation that reads new enquiries, labels them and drafts a reply with a booking link: Zapier or Make calling an AI step, or your CRM's own workflow. For booking, Microsoft Bookings is included in Microsoft 365 Business Basic, Standard and Premium; most mortgage CRMs have a scheduler too.

What it costs. Zapier's free plan runs 100 tasks a month but only two-step Zaps, which is too limited for triage. Professional starts at $29.99 a month billed monthly or $19.99 billed annually, for 750 tasks. Make starts at about $9 a month on credits. An "AI by Zapier" step uses 1, 3 or 5 tasks per run depending on the model tier you pick.

In practice. An enquiry at 7:40am reading "remortgage, fix ends in March, balance about 210k, want to look at options" is labelled "remortgage, normal urgency, missing: property value, income", assigned to the adviser who handles remortgages, and answered by 7:42 with a booking link and a request to have the current mortgage statement ready. The detailed prompt for this step is in the enquiry-to-fact-find workflow.

Watch out for. Task counts creep. Trigger and filter steps don't count; each action does. A five-step Zap across 80 enquiries is 400 tasks before any chasers.

Build one edge case in from day one: not every inbound email is a new enquiry. "This is the third time I've asked, I want to speak to someone senior about how long my case is taking" is a complaint, and an auto-reply with a cheerful booking link makes it worse. Give the triage step a "complaint or distress" label that skips the draft altogether, marks the email urgent and sends the principal a one-line alert with the case reference. Check the label weekly for a month; if genuine complaints are landing as "existing client, general question", add two or three real (anonymised) complaints to the prompt as examples of the tone to look for.

Skip it if you get fewer than about 20 enquiries a month and someone checks the inbox every hour anyway.

3. Discovery and fact-find calls: a note taker

What to use. Three tiers. A general note taker such as Otter.ai; an adviser-built tool such as Jump or Zocks, designed for financial advice meetings and CRM syncing; or Copilot in Teams if your calls run there and you have Copilot licences.

What it costs. Otter Business is $19.99 a user a month billed annually ($30 monthly). Zocks Essentials is $67 a user a month annually ($80 monthly) and doesn't record meetings. Jump's Meet plan is $100 an adviser a month annually ($120 monthly). Check whether the specialist tools integrate with your mortgage CRM before paying the premium; they're built primarily for financial advisers, so support for mortgage systems varies.

In practice. After a 40-minute fact find, the adviser asks for the transcript to be mapped to fact-find fields. Illustrative output extract:

Dependants: 2 children, ages 6 and 9 [08:12]
Committed outgoings: nursery fees end next August [09:40];
  car finance about 280 a month, 2 years left [10:05]
Future changes: applicant 2 may reduce to 3 days a week
  "sometime next year" [22:31] [CHECK: firm plan or idea?]
Credit history: "one late payment on a phone contract years
  ago" [14:50]

The [CHECK] flag earns its keep: whether a reduction in hours is planned or hypothetical changes the affordability conversation, and it's the adviser, not the tool, who needs to pin it down with the client.

Watch out for. Consent at the start of every call, retention settings that match your record-keeping rules, and a review step before anything goes on file. Compliant meeting records with AI note takers covers those set-up decisions in full.

A quick sum before you choose a tier. Take one adviser, illustratively, running twelve fact-find calls a week and spending about 25 minutes writing up each one by hand. Reviewing and correcting an AI draft takes about 8 minutes, so the saving is 17 minutes a call, or roughly 3.4 hours a week. At Otter Business's $19.99 a month, the tool covers its cost if it saves a single hour a month. At Jump's $100 the question changes: it's worth the extra $80 only if its CRM syncing also removes the separate job of keying the note into your case system. Time your own write-ups for a week before you trust anyone's estimate, including this one.

Skip it if most of your fact finds are done through a structured online form and short calls. The saving scales with call length.

4. Documents: a business-plan assistant for reading and summarising

What to use. Your CRM's document extraction if it has it. Otherwise one business-plan assistant: Microsoft 365 Copilot if you run on Microsoft, Gemini in Google Workspace if you run on Google, or ChatGPT Business or Claude Team. The important word is business: these plans don't train on your content by default, unlike free consumer accounts.

What it costs. Microsoft 365 Copilot Business is $21 a user a month on annual billing ($25.20 monthly), with a promotional $18 on annual billing to 31 December 2026. The Business Standard plus Copilot bundle is $23.50 a user a month annually, cheaper than buying the two separately. ChatGPT Business and Claude Team standard seats are $25 a month billed monthly or $20 annually, with a two-seat minimum.

In practice. Three months of bank statements for a single applicant. The prompt asks for regular outgoings with payee, amount, frequency and page reference, and separately for any payments it can't classify. Illustrative result:

Regular outgoings
- Car finance: 286.40 monthly (Jul p2, Aug p2, Sep p3)
- Gym: 42.00 monthly (all three months)
- Childcare: 610.00 monthly (all three months)
- "TRANSFER TO A/C 4471": 300.00 monthly (all three months)
  [unclassified: could be rent, family support or savings]
Irregular, over 500
- 1,250.00 to travel company (Aug p1)

The unclassified transfer is the line worth your attention. It might be savings to a partner's account or a maintenance payment, and only the client can say. AI surfaces it; it doesn't guess.

Watch out for. Year-to-date figures on payslips misread as period pay, joint applicants merged, and photographed documents misread. Always check against the page reference.

How to tell it's working. For your first 20 cases, keep a tally of every figure you corrected, grouped by document type. An illustrative tally after 20 cases: payslips, 3 corrections out of about 120 figures, all of them year-to-date totals read as monthly pay; bank statements, 1 out of about 180, a direct debit split across two lines; annual income summaries, 4 out of 20, every one a phone photo taken at an angle. The pattern tells you what to fix. Here that means asking clients for PDFs rather than photos, and adding one line to the prompt: "Payslips show both this-period and year-to-date columns; use this-period figures only and say which column you read."

Skip it if your CRM already extracts documents well. Two tools reading the same payslip means two sets of errors to check.

5. Research: your sourcing system, plus a cited search tool

What to use. Criteria and products come from your sourcing system and lender documents, full stop. For market context (a lender's announcement, a change in how a lender treats a certain income type reported in the trade press), a citation-based search tool such as Perplexity is useful because it shows its sources.

What it costs. Perplexity Pro is $20 a month; Enterprise Pro is $40 a seat a month and adds admin controls, SSO and SOC 2 Type II assurance.

How it goes wrong. An adviser asks a general chatbot whether a particular lender accepts one year of self-employed accounts. The answer is confident, and the source turns out to be an article from two years ago. The lender changed its policy since. If the adviser had quoted it to a client, the client would have been told something untrue about their prospects. Use AI to find the announcement; use the lender's own criteria to confirm it.

Watch out for. Any tool that answers criteria questions without a dated source.

Skip it if your sourcing system's own news and criteria updates already cover what you need.

6. Recommendation letters: a project with your templates

What to use. A shared project in ChatGPT Business or Claude Team holding your letter template, standard wording and two or three anonymised approved letters, or Copilot in Word drafting from files in your SharePoint. Both keep the house style consistent across advisers.

What it costs. Nothing extra if you already have one of the assistants from stage 4.

In practice. The adviser pastes the signed fact-find summary and their recommendation notes, and asks for a draft of the "why this recommendation" section in plain English. Illustrative draft paragraph:

"We recommended a five-year fixed rate because you told us that knowing your payments for the next few years matters more to you than the chance of lower payments later. You plan to stay in the home for at least five years, and the early repayment charges on this product would only apply if you moved or repaid within that period."

What to fix: the client said "at least four or five years", not "at least five". Close enough to sound right, far enough to matter if they move in year four and hit a charge. The adviser corrects it to the client's words. Every reason in the letter must trace back to the fact find or the call note; whether AI can draft a suitability report that passes compliance review sets out the checks.

Watch out for. Generic reasons that could apply to anyone ("a fixed rate offers peace of mind"). If a sentence doesn't mention this client's circumstances, rewrite it.

Skip it if your CRM generates letters from structured fields already. AI then helps only with the free-text reasons section.

7. Case progression: automated updates from status changes

What to use. Your CRM's workflow engine, or Zapier or Make connecting CRM status changes to email or text messages drafted from a template.

What it costs. Covered by the automation plan in stage 2 for most small firms. Each update is one or two tasks.

In practice. The case status moves to "valuation booked". The client receives, within minutes: "Good news: the lender has booked the valuation for Thursday. You don't need to do anything; the valuer arranges access with the seller or their agent. We'll let you know as soon as the result comes back, usually within a few working days." The client doesn't need to email "any update?", which is where hours go in a busy month. Keeping mortgage clients updated during an application goes further into which statuses deserve a message.

Watch out for. Messages sent on the wrong status because someone clicked the wrong option. Start with a short delay and a daily digest to the case owner of what went out.

One status needs splitting before you automate it: "offer issued". If the offer carries a condition, such as part of the loan held back until repairs flagged by the valuer are done, an automatic "your offer is here, congratulations" is misleading, and the client may tell the seller everything is agreed. Split it in your CRM into "offer issued, no conditions" and "offer issued, conditions attached". Automate the first. Keep the second as a personal call from the adviser, with a follow-up email the adviser writes, AI-drafted if they like, explaining the condition in plain words.

Skip it if your CRM already has a client portal that clients actually use. Many don't log in, which is why a message still helps.

8. After completion: retention reminders that start early

What to use. CRM reminders keyed to product end dates, with AI drafting the outreach so it references the client's own case rather than a generic "your deal is ending".

What it costs. Usually nothing beyond the CRM and assistant you already run.

In practice. Six months before a fixed rate ends, the client gets: "Your fixed rate on the home you bought with us ends next April. Many lenders let you secure a new rate some months ahead, so now is a good time for a short review. Book here: [link]." The review is advice; the reminder isn't. Scanning your whole client bank for these, and for further-borrowing prompts, is covered in spotting remortgage opportunities in your client bank.

Watch out for. Product end dates keyed wrongly at completion. One wrong date means a client hears from their lender before they hear from you.

The other mistake is writing to people whose lives have moved on. Picture a batch of twenty reminders in which one goes to a couple who separated last year, addressed to both names at their old shared email address. The draft was accurate about the product; it was still the wrong message. Before any batch goes out, filter out cases with a life-event note, anyone who has told you they've moved to another broker, and any do-not-contact marker, and address each reminder to an individual contact record rather than a joint one. A two-minute scan of the list by the adviser who knows the clients catches what the filters miss.

Skip it if you already have a reliable retention process. AI adds personalisation, not the process itself.

Tools that aren't worth paying for yet

• Anything that promises to decide affordability or eligibility automatically. That's the lender's underwriting and your advice. A tool claiming to do it for you either overstates what it does or moves into territory you'd need to justify to your compliance adviser.
• A second general assistant. Paying for Copilot and ChatGPT Business and Claude Team for the same people usually means none is set up properly.
• AI prospecting agents that send outreach at volume. Mortgage clients come mostly through referrals, introducers and existing clients. Bulk AI outreach tends to damage the reputation those channels depend on.

The order to switch them on over 90 days

Turning on eight tools at once guarantees that none is set up properly. A sequence that works for most small brokerages, with one measure for each step so you know whether it paid off:

1. Weeks 1 to 3: documents (stage 4). It's the biggest time sink and needs no client-facing change. Measure minutes per case spent keying figures, before and after, on ten cases each.
2. Weeks 4 to 6: triage and first reply (stage 2). Measure time from enquiry to first reply, and the share of enquiries that book a call. If the booking rate doesn't move, the reply wording needs work, not the automation.
3. Weeks 7 to 9: call notes (stage 3). Start with one adviser for a fortnight, compare the AI-assisted notes against their usual ones, then decide on a general or specialist tool.
4. Weeks 10 to 12: progression updates and retention (stages 7 and 8). Measure inbound "any update?" emails per case. Most firms know this number is high; few have counted it.

Leave enquiry chat (stage 1) and research tools (stage 5) until the rest is running, unless enquiries arrive mostly out of hours, in which case move stage 1 to the front. Letters (stage 6) can start whenever an adviser is willing to build the project with their templates; it needs an afternoon of set-up and a compliance lead's review of the first ten drafts.

One practical point on data: each new tool is another place client information lives. Keep a short entry per tool recording what it holds and how long it keeps it, before it goes live rather than afterwards. An illustrative entry for a note taker:

Tool: Otter Business (2 adviser seats)
Holds: audio and transcripts of fact-find and review calls
Kept: per the retention setting we chose on [date]; transcript
  copied to the case file, then deleted from the tool after
  [your period]
Who can see it: both advisers, principal (admin)
Client told: consent line read at the start of every call
Last checked: [date], next check [date + 3 months]

If you can't fill in the "Kept" line because the tool has no retention setting you control, that's worth knowing before it holds a year of client calls.

Two illustrative stacks for a two-adviser firm

For a firm with two advisers and one administrator, on list prices with annual billing:

The bundle line includes the Office apps, email and Bookings as well as Copilot, so part of that cost replaces a subscription you'd pay anyway. Neither stack includes your mortgage CRM or sourcing system, which remain the core of the case. The lean stack is where I'd start for most small firms: prove the time saving on calls and documents first, then decide whether a specialist note taker justifies its extra $80 a seat a month on annual billing.

Questions about building a broker's AI stack

Should I wait for my mortgage CRM to add AI rather than buying separate tools?

Ask your CRM provider for its roadmap in writing first. If document reading or client updates are due within a few months, waiting avoids paying twice and keeps client data in one system. If the roadmap is vague, a general assistant on a business plan costs little per month and can be dropped later.

Is one general assistant enough for a small brokerage?

For many two- or three-adviser firms, yes. A business-plan assistant such as Microsoft 365 Copilot, ChatGPT Business or Claude Team covers email, document reading and letter drafting. Specialist tools earn their place where volume is high, such as a note taker for advisers running many calls a week.

Which of these tools needs the most careful compliance review?

Anything that writes client-facing advice documents or listens to advice calls. Recommendation letters and call notes become part of the regulated record, so they need a defined review step, retention settings you've chosen deliberately, and a sampling routine. Lead-capture chat and admin automations carry less risk, provided they never quote rates or eligibility.

How do I stop staff using free chatbots for client documents?

Give them an approved alternative that's just as quick, then write the rule down. Most unofficial use happens because the approved route is slower or doesn't exist. A business-plan assistant inside your Microsoft or Google account, with a short list of approved tasks, removes most of the temptation.

Further reads

• How Much Does an AI Note-Taker Cost per Seat? — A closer look at note-taker pricing per seat.
• How to Read AI Software Pricing: Seats, Credits, and Usage Fees — Seats, credits and usage fees decoded before you sign.
• Questions to Ask Before Buying AI That Touches Client Data — Due diligence for any tool that touches client documents.
• Is It Safe to Let AI Listen to Mortgage Advice Calls? — The safety questions behind stage three.
• What Mortgage Brokers Can Automate With AI, and What Stays Advice — Where the line between automation and advice sits.
• WhatsApp Business App vs API: Which Do You Need for AI Replies? — If clients prefer messaging, which WhatsApp route you need.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: HubSpot pricing and Customer Agent announcement; Jump and Zocks pricing pages; Otter.ai, Zapier, Make, Perplexity, ChatGPT Business, Claude Team and Microsoft 365 Copilot Business pricing; Microsoft Learn on Bookings availability; Meta's Business Agent pricing and WhatsApp Business terms. All checked September 2026.