AI for mortgage brokers works when it handles the conversation before advice and the chasing of documents, and stops dead at the point where anything resembles advice. On the levelpartner.ai platform that means two agents: a Conversation AI bot on chat, SMS and messaging channels, and a Voice AI agent that answers the phone after hours. Both are trained only on content your compliance lead has approved, and both hand over to a person the moment a client asks which product, lender or rate suits them. This article sets that boundary, then shows how to build each agent to respect it.
Where AI fits in a brokerage, and where it does not
A brokerage runs on three kinds of work. There is advice, which is regulated and belongs to a qualified adviser. There is administration around the case: collecting documents, booking appointments, chasing lenders, updating clients. And there is the relationship between cases: keeping in touch with a client whose fixed rate ends in three years, or who might buy an investment property next spring.
AI is good at the second and third kinds. It answers at 9pm on a Sunday, asks the same fact-find questions every time, and never forgets to send a reminder. It is poor at the first kind, and in a regulated market it has no business there. An AI agent that tells a first-time buyer which product suits them has given advice, and your firm owns the consequences.
So the rule is simple. AI gathers, schedules, reminds and routes. Advisers recommend.
Where most brokers draw the line in the wrong place
Most firms write the boundary as a list of banned topics. The agent is told not to recommend products, and then a client asks "would I be better on a two-year or five-year fix?" and the agent, trying to help, explains the difference in a way that reads like a recommendation. A banned-topics list fails because clients rarely use the words on it.
Draw the line by what the agent is for instead. Its job is to collect facts, explain your process and book time with an adviser. Anything that asks the agent to compare, choose or predict goes to a person, however it is phrased. That is easier to instruct, easier to test and easier for a compliance lead to sign off.
Step 1: build a knowledge base of approved content only
An agent answers from what it has been trained on. If you point the crawler at your whole website, it learns your blog's opinions on rate movements and your old product pages. So build a separate knowledge base containing only process content your compliance lead has read: how an appointment works, what documents you need, your fee structure as published, your opening hours, how complaints are handled.
- Go to AI Agents, then Knowledge Base.
- Create a Knowledge Base or open an existing one.
- Click Add Source.
- Select the source type: File Upload, Tables, Rich Text or Google Drive.
- Add the approved content and wait for processing to finish.
Add the questions clients ask most as FAQ entries, each with a short answer in your own words. When you do use the web crawler, choose Exact URL for each approved page rather than crawling the whole domain, and skip anything with market commentary. Before connecting the knowledge base, open the Knowledge Base Retrieval Tester and ask the awkward questions: "which lender is cheapest?" should retrieve nothing useful, which is the point.
Step 2: set up the Conversation AI bot for pre-advice enquiries
The chat and SMS bot handles enquiries out of hours, asks the opening fact-find questions (buying, remortgaging or moving, rough timescale, best time to talk) and offers an appointment with an adviser. It also chases documents: a client who replies "what counts as proof of address?" gets an answer from your approved list.
- Go to AI Agents, then Conversation AI, and click Create Bot.
- Select Prompt Based Bot.
- Choose Appointment Booking.
- In the bot editor, configure Bot Settings and Bot Training, selecting your approved Knowledge Base.
- Configure Bot Goals using prompts, model settings and actions, including Human handover.
In the prompt, state the role in one line: "You collect enquiry details and book appointments for a mortgage brokerage. You never compare, recommend or predict products, lenders or rates." Then set the human handover to fire whenever a client asks a question of that kind, and say what the bot tells the client: that their adviser will cover it in the appointment, and here are three times.
New bots start in Off. Move to Suggestive first, so a case handler sees and approves each reply for a week, and only then to Auto-Pilot.
Step 3: add a Voice AI agent for after-hours calls
Calls that ring out after six are often the most motivated: a buyer with an offer accepted, a client whose lender has written to them. A Voice AI agent answers those, takes the reason for the call, and books a callback.
- Go to AI Agents, then Voice AI, and open the Agent List tab.
- Click Create Agent, select Create Custom Agent and click Continue.
- Enter an Agent Name, then select the Voice, Model and Language.
- Write the prompt: the agent's role, the information to collect, allowed topics, topics to avoid, when to transfer and how the call ends.
- Set the Welcome Message for inbound calls.
- Click Save.
- Open Knowledge Base, select your approved knowledge base and describe when the agent should search it.
- Open Actions, click New Action and add a Call Transfer with its trigger condition and destination number.
Then, on the agent's Deploy tab, select the phone number, choose Use as Backup under Call Routing, and set Working Hours to a Custom Schedule covering evenings and weekends. Test with a Phone Call rather than a Web Call, because transfers only run on a real call, and ring the number outside office hours to confirm the agent picks up.
The welcome message should say plainly that the caller is speaking to an automated assistant and that an adviser will call back. Instruct the agent not to guess when information is missing. A voice agent improvising an answer about stamp duty is the failure you are building to prevent.
Say the firm misses around 60 calls a month after six. If the Voice AI agent answers as backup on evenings and weekends and 45 of those callers leave their details and accept a callback slot, advisers would start each morning with a list of named callbacks and a summary of each call instead of a voicemail inbox. These figures are illustrative.
Where the agents earn their keep
Out-of-hours enquiries and missed calls are the obvious jobs. Document chasing is the quieter one. Cases stall when a payslip is missing and nobody has asked twice. A workflow sends the list of outstanding items, and the bot answers the replies that follow from your approved list. Mortgage document collection sets out the checklist, upload form and reminder sequence behind that. For how enquiries move from first contact to an adviser's diary, see running mortgage lead management in one system. If you want a lighter first step before a voice agent, missed call text back covers the text-only version.
The bigger return usually sits in the existing book: clients whose fixed rates end in the coming year. A workflow can start a rate-review conversation months ahead and the bot can book the review, provided it stays on the booking and leaves the review itself to the adviser. The timing and ownership behind that are covered in a mortgage retention strategy for past clients.
Suppose 600 clients have a product ending in the next 12 months and the team reaches 200 of them before the lender does. If a bot-led booking sequence raises that to 450, that is 250 extra conversations without one new enquiry. If one in three of those became a case at an illustrative £1,600 in total fees, that would be roughly £133,000 of extra fee income. These numbers are hypothetical.
What you give up by working this way
There is a trade-off, and you should go in with your eyes open.
Keeping AI out of advice means your agents are deliberately narrow. They will decline questions a client wants answered: "Which lender is cheapest?" gets "Your adviser will go through that with you", not an answer. Some clients find that frustrating, and a competitor running a looser chatbot may seem more helpful in the moment.
You also give up some speed of setup. A narrow agent needs a clear prompt, defined handovers and tested responses for the questions it must refuse, plus a week in Suggestive mode. That takes longer than switching on a general chatbot. The extra week is worth it, because the alternative is an agent your compliance lead will eventually switch off.
Letting the AI agent quote rates or name lenders "just as a guide". Rates change daily and a named product is advice. Give the agent a firm instruction to hand any product question to an adviser, and test it with ten awkward phrasings before it goes live.
Compliance and record-keeping
In the UK, the US and most other markets, mortgage advice is regulated, and so is how you contact people. Before any AI agent sends a message, make sure your consent records are in the CRM, your opt-out handling works on every channel, and your adviser hand-off is logged.
The advantage of running AI inside your CRM is that every AI conversation sits in the client record, with a timestamp, next to the adviser's notes. When you need to show what a client was told and when, the full thread is there. A standalone chatbot rarely gives you that.
Keep a written policy for what your AI agents may and may not say, and review it whenever you change a script. Your compliance lead should sign off each agent before it goes live, the same way they would sign off a new marketing email.
How long it takes to run each week
Once set up, AI agents do not run themselves. Plan for a short weekly review: read a sample of conversations, check that hand-offs reached the right adviser, and adjust any reply that confused a client. For a brokerage with a handful of agents, that review usually fits inside an hour a week for one person who knows the business.
The setup itself is the larger investment. Writing scripts, mapping hand-offs and loading product end dates into the CRM is a few days of focused work. Getting those dates into the system cleanly is the part most firms underestimate, because the data often sits in old case files and spreadsheets.
When not to use AI
AI is the wrong tool in a few situations:
- When your CRM data is unreliable. An agent that messages a client about a product they redeemed last year does more harm than silence.
- When a client is vulnerable or distressed. Your workflow should route anything that mentions bereavement, separation, illness or arrears straight to a person.
- When the case is complex. Self-employed income, adverse credit or unusual property types need an adviser from the first message.
For everything else, the test is whether a well-trained junior administrator could do the task from a script. If yes, AI can probably do it. If the task needs judgement, keep it with a person. Our guide to qualifying leads before they reach an adviser covers what the opening questions should establish.
Frequently asked questions
Can an AI agent give mortgage advice to clients?
No, and you should configure it so it cannot. Its job is to collect information, answer process questions and book time with an adviser. Any question about products, lenders or suitability should trigger a hand-off to a qualified person.
How do we stop the agent drifting into advice?
Train it only on an approved knowledge base, state in the prompt that it never compares or recommends, and set a human handover for any question of that kind. Then test ten awkward phrasings before it goes live, and repeat the test after every change.
Will clients know they are talking to AI?
They should. Tell them plainly at the start of the conversation and give them a way to reach a person at any point. Clients are generally comfortable with an AI assistant handling booking and reminders when it is honest about what it is.
Do we need a separate AI tool on top of our CRM?
Not on the levelpartner.ai platform. Conversation AI, Voice AI, workflows and the CRM sit in one system under one licence, so every conversation is written back to the client record. That shared record sits at the centre of a sound CRM strategy for a growing firm.
Where should a brokerage start with AI?
Start with the Conversation AI bot in Suggestive mode on out-of-hours enquiries. It is contained, easy to measure and every reply is checked. Add the after-hours Voice AI agent once the bot has run cleanly for a month.