What Mortgage Brokers Can Automate With AI, and What Stays Advice

Mortgage brokers can safely automate the administrative and informational side of a case: enquiry handling, document collection and sorting, data extraction into the fact-find, case status updates and deal-expiry reminders. What stays advice is anything that recommends: which lender or product, remortgage versus product switch, term, repayment method, debt consolidation and protection needs.

The test is whether a client could reasonably read a message as a recommendation about their own circumstances. "Your fixed rate ends on 1 March, so let's review your options" is information. "Rates have fallen, you should switch to a five-year fix" is advice, whether a broker, an assistant or an automation sent it. Rules differ between markets, and your compliance adviser or network sets the exact boundary, but that reading test catches most of the problems before they reach a client.

Follow me on Instagram@sagnikteaches

The advice line, drawn task by task across a case

Laying a typical residential case out stage by stage shows how much of it sits safely on the automation side:

Connect on LinkedInSagnik Bhattacharya

Read across the rows and a pattern appears. The "Automate" column is almost entirely moving information to or from the client. The "Assists" column is preparing the broker to advise. Nothing in the "Stays advice" column is a document task; it's all judgement about a person. For the front end in detail, an AI workflow from enquiry to fact-find walks through the first two rows.

Subscribe on YouTube@codingliquids

Three automations that quietly cross the line

Most brokers don't set out to automate advice. It creeps in through automations that start out as information and grow a sentence too far.

"How much can I borrow?" on the website

A chatbot or calculator that answers this is useful and popular. The problem is specificity. An illustrative before and after of a chatbot's reply:

Crosses the line: "Based on your joint income of $95,000, you can borrow up to $427,500 with Lender A at 4.5 times income. That would suit the flat you mentioned."

Stays information: "Lenders often lend somewhere between four and five times income, but it depends on your outgoings, credit history, deposit and the lender's own rules, so a figure from a calculator is only a rough guide. A broker can check what's realistic for you. Would you like to book a call?"

The first names a lender, gives a precise figure and comments on suitability. The second explains how lending generally works and hands over to a person.

Deal-expiry reminders

Reminders are one of the most valuable automations a broker can run, and spotting remortgage opportunities in your client bank covers how to find the right clients. The risk is in the message:

Crosses the line: "Good news, rates have dropped! Switching now could save you $140 a month. Reply YES and we'll move you to a new five-year fix."

Stays information: "Your current rate with your lender ends on 1 March 2027. Most people review their options three to six months before a deal ends. Would you like to book a review with [broker]? It takes about 20 minutes."

The first makes a saving claim nobody has checked, recommends a product type and invites the client to agree by text. The second states a fact about their mortgage and offers a conversation.

The chatbot that "just explains"

Explaining the difference between fixed and tracker rates is fine. The trouble starts when the client adds their own details. "I'm self-employed and might move in two years, should I fix?" is a request for advice. A well-built bot recognises the shift and replies along the lines of: "That depends on your plans and circumstances, which is exactly what a broker will go through with you. In general, fixed rates give certainty of payments; trackers move with the base rate. Shall I book you in?" Test for this specifically by asking your bot personal questions dressed as general ones.

A short test sheet makes that repeatable. Here's an illustrative one, filled in after a firm's first run:

The failure on the second row came from the bot doing arithmetic on figures the client supplied. The fix was one instruction: "never calculate an income multiple or borrowing figure from anything the user tells you". Re-run the whole sheet after every change to the bot's instructions, because fixing one answer can shift another.

Wording that stays on the right side

A quick reference for anyone writing templates, chatbot instructions or automated messages:

Put the left-hand column into your chatbot and automation instructions as banned phrasing, and test for it before launch.

Documents and the fact-find: where AI earns its keep

Document handling is the biggest safe time saving in a mortgage case. AI can read payslips, bank statements, tax documents and credit reports, extract the figures into your fact-find and flag things the broker should ask about. A prompt for bank statements:

From the three months of bank statements attached, list:
1. Regular income: payer, amount, frequency, dates
2. Regular commitments: loans, credit cards, finance
   agreements, child maintenance, with amounts
3. Items a broker should ask the client about, with date and
   amount: returned payments, overdraft use, payday or
   short-term lenders, gambling transactions, large or
   unexplained transfers
Quote the statement description for each item. Do not
interpret or judge; just list.

Illustrative output (extract):
1. Income: ACME LOGISTICS SALARY, 2,850.00 monthly (28th).
   ACME LOGISTICS O/T, 410.00 (Jun), 380.00 (Jul), 620.00 (Aug)
2. Commitments: car finance, 289.00 monthly; credit card
   minimum payments, approx. 60.00 monthly
3. To ask about: overdraft used 11-19 Jul (lowest -312.40);
   transfer out 2,000.00 on 3 Aug, "SAVINGS JH"

Useful, and one line needs a broker's correction. The AI listed overtime under "regular income". Whether and how much overtime counts depends on the lender's rules, how long it's been received and whether it's guaranteed, which is exactly the kind of judgement that stays with the broker. The prompt gets a fix ("list overtime, bonuses and commission separately from basic pay"), and the broker decides how to treat it. The $2,000 transfer to "SAVINGS JH" is likely the client's own savings, but it's the broker who asks. Choosing a tool for this step is covered in the best AI tools for each stage of a case.

Helping the broker advise, without advising

AI can make the advice itself faster to prepare while leaving the decision alone. Three safe patterns:

• Criteria summaries from lender documents. Paste in a lender's published criteria for self-employed income and ask for a summary of the rules that apply to a client with two years' accounts. The source is the lender's document, not the AI's memory. Never ask a general model what a lender's criteria are without giving it the document; it will answer confidently and may be out of date.
• Comparing sourcing results. Export your sourcing system's shortlist and ask for a plain comparison of fees, rates, early repayment charges and incentives over the client's chosen period. The broker checks the arithmetic and makes the choice.
• Drafting the recommendation letter. The broker writes the reasons in a few bullet points; AI turns them into a clear letter for the client. The reasons are the broker's, the words are shared.

The sourcing comparison needs the broker's arithmetic check more than anything else on this list. Take an illustrative remortgage of $250,000 over 25 years, comparing a two-year fix at 4.1% with a $1,495 fee against one at 4.3% with no fee. Asked for "a comparison", the AI's summary led with "Product A has the lowest rate and the lowest monthly payment", which is true: about $1,333 a month against $1,361. Over the 24 months, though, A's $28-a-month saving adds up to about $670, less than half its fee. Payments plus fees come to roughly $825 more on A (about $500 more once you allow for the slightly lower balance A leaves at the end). The fix is in the prompt: ask for total cost over the fixed period, fees included, and for the balance remaining at the end, and never let the summary open with a ranking.

The recommendation letter has a subtler failure. A broker's illustrative bullet points for a self-employed client read:

- Lender B: accepts one year's accounts; client only has 18 months trading
- 5-year fix: client wants fixed payments while the business grows
- Overpayments up to 10% a year allowed; client expects lumpy income
- Not the cheapest rate on the sourcing results; see criteria point above

The AI's draft (illustrative) turned those into a clear letter but added a sentence of its own: "This product also offers excellent value compared with the wider market." Nobody said that, and the broker's last bullet says the opposite. Check every reason in the draft against the bullets, line by line, and delete anything that isn't in them. A reason the broker didn't give is the most dangerous sentence a recommendation letter can contain, because the client may rely on it later.

How the line moves with the type of case

The admin-versus-advice split holds across mortgage work, but the proportions and the risky moments change:

• First-time buyers. Lots of explaining, lots of documents, lots of anxious "what happens next?" messages. Automation pays off in education and updates: a plain-English explainer sequence on how the process works, sent after the first meeting, answers questions before they're asked. The advice risk is in affordability: nervous buyers ask chatbots "can I afford this?", and the answer must always route to the broker.
• Product transfers and remortgages. The highest-volume automation opportunity, because the trigger is a known date. The danger is the one described above: reminder messages that drift into recommending a switch. Keep reminders factual and put the "stay, switch or remortgage" conversation with a broker, even when the answer looks obvious.
• Landlords and buy-to-let. Document-heavy: rental statements, portfolio schedules, company accounts for limited-company landlords. AI extraction saves real time building a portfolio schedule from a dozen tenancy agreements and mortgage statements. The advice sits in structure and lender choice, which vary a lot with the landlord's circumstances.
• Later-life lending. Clients may be more vulnerable, conversations often involve family members, and the products carry long-term consequences. Keep automation to scheduling and document handling, and be slower to introduce chatbots or AI-drafted client messages here than anywhere else in the business. Family involvement creates its own edge case: a client's son emails asking "where are we with Mum's application?", and an automated status update goes back to him with the case details. Unless the client has given written authority for him to receive information, that's a disclosure you shouldn't have made. Record third-party authority on the case, and have the automation send updates only to addresses on that list.

Template drift is the product-transfer trap, and it creeps in through small edits. A firm sets up an automation that emails clients three months before their deal ends. Someone edits the template to be "more engaging" and adds: "With rates lower than when you fixed, now is a great time to lock in a new deal." For most clients that's harmless. For one, whose circumstances had changed and who needed to move lender rather than take the product transfer on offer, it reads as the firm telling them to stay put and fix. They reply "yes please", and the case is nearly processed as a straight transfer before a broker reads the file. The fix: templates are locked, edits need broker approval, and any "yes" reply to an automated message opens a review appointment, never a product switch.

Controls that keep automation on the admin side

1. A broker approves every template before it's used in any automation, and re-approves it when it changes.
2. No free-text AI replies to clients without review, unless the channel is limited to scheduling and factual status. Anything conversational goes through an approval step.
3. Banned phrasing in every set of instructions, using the table above, and tested with personal questions.
4. A log of automated messages kept with the case, so you can show what the client was sent and when.
5. Disclaimers aren't a fix. "This is not advice" at the bottom of a message that says "you should switch" doesn't make it information. Fix the message.

One regulatory point for brokers with customers in the EU: the EU AI Act lists AI systems used to evaluate the creditworthiness of individuals as high-risk (Annex III), with those obligations deferred to 2 December 2027 for stand-alone systems. That's mainly aimed at lenders' decision systems, but if you use a tool that scores or ranks applicants' credit profiles, ask the vendor how they're preparing for it.

A two-broker firm's automation map

Picture a firm of two brokers and one administrator, completing around 25 cases a month. Before automation, each case takes, say, 10 hours of total staff time, of which roughly four are advice (fact-find discussion, research, recommendation, client calls on decisions) and six are admin (documents, chasing, updates, data entry, reminders).

A sensible automation plan targets the six admin hours: document requests and sorting, extraction into the fact-find, status updates and deal-expiry reminders. If those drop from six hours to about three and a half, the firm gets back around 60 hours a month across 25 cases, much of it administrator time. The four advice hours barely move, although AI-drafted recommendation letters might trim half an hour. Keeping the updates themselves accurate is its own topic, covered in keeping mortgage clients updated during an application.

The numbers are illustrative. What carries over is the shape: most of a mortgage case's hours are admin, most of the automation value is there, and the advice hours are both the smallest block and the one you shouldn't try to shrink with AI. If you're choosing systems to support this, choosing a mortgage CRM with AI built in covers what to look for, and if call recording is part of the plan, letting AI listen to advice calls safely deals with the data risks.

Further reads

• Can Financial Advisers Use ChatGPT? What Compliance Allows — Compliance ground rules that apply to broker firms too.
• How to Add Human Approval Steps to AI Automations — Put a broker's approval in front of any client-facing message.
• How to Anonymise Client Data Before You Paste It Into AI — Test automations on real cases without exposing client data.
• AI Document Extraction vs Traditional OCR: Which Should You Use? — Which document-reading approach suits payslips and statements.
• Should You Tell Customers You Use AI? A Disclosure Guide — How open to be with applicants about automation.
• A Simple AI Risk Register for Small Businesses (With Template) — Log the advice-line risks before anything goes live.
• How to Handle Sensitive Customer Conversations Without AI — Which customer conversations to keep away from AI, how to route them to a named person quickly, and what that person says when they pick up.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: EU AI Act high-risk categories (Annex III, creditworthiness) and deferred dates; vendor pricing as listed September 2026. Practical guidance only; mortgage advice rules differ by market, so confirm boundaries with your compliance adviser or network.