Three AI Agents, One Mortgage Brokerage, 63% More Applications Processed

A four-person mortgage brokerage in Hertfordshire was processing 35 applications a month. The ceiling wasn't the market — enquiries were coming in faster than they could be turned around. The ceiling was 29 hours of weekly admin: document chasing, lender research, and application preparation that no one had time to automate. Three AI agents removed that ceiling. Ninety days later, they were processing 57 applications a month, the team had recovered 23 hours a week, and the whole system was running at £80 a month. Mortgage broking is a relationship business. The value a broker delivers is advice, judgement, and the ability to navigate a market where product criteria change weekly and lender appetite shifts without notice. The problem is that most of a broker's day isn't actually spent advising. It's spent chasing documents, cross-referencing income against lender affordability calculators, reading policy guides, and drafting submission notes that say largely the same thing in slightly different words for different lenders. That's the gap AI fills — and it's a larger gap than most brokers realise until they time it. <h2>The Problem Nobody Talks About in Mortgage Broking</h2> <figure> <img src="https://images.unsplash.com/photo-1454165804606-c3d57bc86b40?w=1200&q=80" alt="Broker working through stacks of mortgage documents — representing the manual admin burden that caps throughput in UK mortgage brokerages" width="1200" height="800" loading="lazy" /> </figure> Before we built anything, we mapped how the team at this Hertfordshire brokerage actually spent their time. Four brokers, all experienced, all generating good client outcomes. The mapping exercise produced a number that surprised even them. Of the 40 working hours per week each broker logged, roughly seven were spent in genuine advisory conversations — fact-finds, recommendations, follow-ups where their expertise mattered. The other 33 were split across four categories of repeatable admin: <ul> <li>Document collection and chasing. Clients submit incomplete document packs. Brokers chase by email and WhatsApp. Documents arrive in the wrong format. Someone has to rename, organise, and file everything before it can be reviewed. Average: 9 hours per broker per week.</li> <li>Income verification and affordability checking. Reading payslips, bank statements, P60s, and self-assessment returns to extract the figures each lender's affordability model needs. Average: 7 hours per broker per week.</li> <li>Lender research and product matching. Cross-referencing client profiles against lender criteria — LTV limits, income multiples, adverse credit policies, property type restrictions — to find the right products. Average: 8 hours per broker per week.</li> <li>Application preparation. Drafting submission notes, completing decision-in-principle forms, and preparing supporting case notes for complex applications. Average: 9 hours per broker per week.</li> </ul> That's 33 hours of repeatable, structured work per broker per week. Work that follows rules. Work that an AI agent can do. <blockquote>The brokers weren't slow. They were doing the right things. The problem was they were doing them by hand when software could do it faster, more consistently, and without fatigue.</blockquote> This pattern is consistent with what the wider mortgage sector is now reporting. Research published in early 2026 found that 55% of UK mortgage brokers now use AI tools in some form — but most are using point solutions for individual tasks rather than an integrated system that handles the full admin workflow. The gap between using AI and having AI run your operations is where most practices are stuck. <h2>What the Three AI Agents Do</h2> <figure> <img src="https://images.unsplash.com/photo-1518770660439-4636190af475?w=1200&q=80" alt="Three interconnected AI agent nodes forming a workflow — illustrating the document intelligence, lender match, and application preparation agents built for this UK mortgage brokerage" width="1200" height="800" loading="lazy" /> </figure> We built three agents that together cover the bulk of that admin load. Each one does a specific job. Together, they form an AI Operating System that runs alongside the brokers and hands them work that's ready to review — not work that still needs to be done. <h3>Agent 1: The Document Intelligence Agent</h3> This agent monitors the client documents inbox. The moment a document arrives — payslip, bank statement, P60, SA302, proof of address, passport scan — it opens it, reads it, and extracts the relevant data into a structured record. For payslips: employer name, employment type, gross pay, net pay, bonus amounts, deductions. For bank statements: income credits, consistent expenditure, and any anomalies lenders typically query — gambling transactions, undisclosed credit repayments, irregular large outgoings. For SA302s: self-employed income figures by year, the trend, and any points the lender will want explained in the case notes. The agent also tracks what's missing. It compares what's been submitted against the required document checklist for the client's application type and sends a specific, plain-English message to the client listing exactly what's still needed — not a generic "please send your documents" email, but "we still need your March 2026 payslip and six months of bank statements for your Halifax account." Document chasing time: from 9 hours per broker per week down to under 90 minutes. <h3>Agent 2: The Lender Match Agent</h3> Once the document intelligence agent has extracted and structured the client's financial profile, the lender match agent goes to work. It takes the client's income, deposit, property type, credit history, employment status, and purchase or remortgage objective, and cross-references them against the current criteria for 40+ lenders on the brokerage's panel. It returns a shortlist of the most suitable products — typically three to five — ranked by overall suitability, with a brief rationale for each one. The broker reviews the shortlist, makes the final recommendation with their judgement and knowledge of the client, and proceeds. The 7-8 hours of manual criteria research collapses to 20 minutes of review. This agent is updated weekly as lender criteria change. Criteria changes are the single biggest source of wasted effort in mortgage broking — a broker spends an hour researching a lender, only to discover on submission that a policy changed two weeks ago. The agent catches those changes before the broker starts the work. <h3>Agent 3: The Application Preparation Agent</h3> The third agent handles the part of the process that requires the most skilled human time but the least human judgement: drafting the submission pack. It pulls the structured data from the document intelligence agent, the selected lender from the broker's recommendation, and the client's fact-find data, then produces a first draft of the full submission: completed application form data, case notes explaining the client's circumstances, a supporting income narrative where required, and a cover note to the lender for complex cases. The broker reviews the draft. In straightforward cases — roughly 70% of the workload — they approve it with minimal changes. In complex cases, they edit the sections that require their specific knowledge and judgment. Application preparation time drops from 9 hours per broker per week to under 2 hours. <h2>How We Built It</h2> <figure> <img src="https://images.unsplash.com/photo-1563986768609-322da13575f2?w=1200&q=80" alt="Technical workflow build process showing the six-week implementation of an AI Operating System for a UK mortgage brokerage" width="1200" height="800" loading="lazy" /> </figure> The build ran over six weeks. The technology stack is straightforward: a workflow engine (n8n), a vision-capable AI model for document reading, a custom lender criteria database updated weekly, and integration with the brokerage's existing CRM and document storage. Weeks one and two: we built and tested the document intelligence agent on a sample of 200 historical client documents — payslips, bank statements, and SA302s from real anonymised cases. Extraction accuracy came in at 97.3% on clean PDFs and 94.1% on phone-camera scans. We added an OCR pre-processing step for scanned documents, which brought accuracy up to 96.8%. Week three: the lender criteria database. We structured the criteria for 40 lenders — LTV bands, income multiples by employment type, adverse credit tolerances, property type restrictions — into a queryable format. This is the highest-maintenance part of the system, so we built a weekly update process that reviews published criteria changes and updates the database. Week four: the lender match logic. This isn't pure AI — it's a combination of rules-based filtering and an AI ranking layer that assesses suitability based on the full client picture, not just headline numbers. Rules eliminate ineligible lenders. AI ranks the eligible ones. Weeks five and six: the application preparation agent, integration testing, and broker training. The brokers needed to trust the outputs before they'd act on them — which is the right instinct. We ran two weeks of parallel operation, where the agents did their work and the brokers checked it against their manual process. By the end of week six, the team was confident enough to go live. For the technical architecture behind this kind of multi-agent system, our <a href="/blog/multi-agent-orchestration-patterns">guide to multi-agent orchestration patterns</a> covers the four approaches we use and when each one applies. <h2>The Numbers After 90 Days</h2> <figure> <img src="https://images.unsplash.com/photo-1611974789855-9c2a0a7236a3?w=1200&q=80" alt="Results dashboard showing 63% more applications, 23 hours recovered, and 91% first-submission success rate after 90 days of running AI agents in a UK mortgage brokerage" width="1200" height="800" loading="lazy" /> </figure> We tracked four metrics across the 90-day live period. Applications processed. The brokerage went from 35 mortgage applications a month to 57. That's a 63% increase with the same four-person team. The bottleneck wasn't capacity — it was admin throughput. Remove the admin, and the throughput follows. Hours recovered. Across the four brokers, total weekly admin time dropped from 29 hours to 6 hours. That's 23 hours a week returned to the team — the equivalent of adding a full-time broker's worth of productive hours without the salary. First-submission success rate. Applications accepted by lenders without an information request improved from 74% to 91%. The document intelligence agent catches missing information before submission. The application prep agent produces more consistent case notes than a tired human at 4pm on a Friday. Running cost. The full system — AI API calls, n8n cloud hosting, weekly criteria database updates — runs at £80 a month. Against the revenue generated by processing 22 additional applications a month at an average broker fee of £500–700 per completed case, the ROI is not subtle. <blockquote>The number that landed hardest with the principal broker wasn't the applications processed. It was the first-submission success rate. That number represents real money — lender arrangement fees paid faster, client satisfaction higher, broker reputation with lenders better.</blockquote> If you want context on how running costs at this level are achieved, our piece on <a href="/blog/ai-agent-cost-optimisation-uk">AI agent cost optimisation</a> explains the three-pillar approach we use to keep LLM costs low even as volume scales. <h2>What to Do If You Run a Mortgage Brokerage</h2> <figure> <img src="https://images.unsplash.com/photo-1560520653-9e0e4c89eb11?w=1200&q=80" alt="Confident mortgage broker at a clean desk with a laptop — representing the freed-up capacity that AI agents deliver to UK brokerages" width="1200" height="800" loading="lazy" /> </figure> The three agents we built here are directly applicable to any FCA-authorised mortgage broker business, regardless of size. The document intelligence agent works whether you're processing 10 applications a month or 150. The lender match agent makes more difference as your lender panel grows. The application prep agent pays back fastest where your application volume is highest. The regulatory question — and it's the right one to ask — is about human oversight. FCA rules require a regulated person to own the advice process. These agents don't give advice. They prepare information for a regulated person to review, recommend, and sign off. The broker's judgement stays at the centre of every client outcome. The AI handles the retrieval, extraction, and drafting. That's the right division of labour, and it's FCA-compliant. The most common objection we hear from brokers considering this is that their cases are "too complex" for AI to help with. The irony is that the agents are most valuable precisely in complex cases — where document reading is more detailed, the lender search is harder, and the case notes require more careful drafting. Straightforward first-time buyer cases don't need an AI agent to process them faster. Complex remortgages, self-employed income cases, and adverse credit applications are where the time saving is largest. For context on how AI Operating Systems scale across different service businesses, our <a href="/blog/ai-agents-vs-automation">guide to AI agents versus traditional automation</a> explains the core distinction — and why it matters when you're building something that has to be FCA-compliant and audit-ready. <h2>Ready to Remove the Admin Ceiling?</h2> If you run a mortgage brokerage and want a clear picture of what an AI Operating System would look like in your practice — which agents, what they'd handle, what the build would cost, and what you'd reasonably expect after 90 days — that's exactly the conversation we have with every firm before we start a build. <a href="/contact">Book a free 30-minute call</a> and we'll map your current workflow, identify the admin load that AI can take off your plate, and give you an honest estimate of what to expect. No pitch, no pressure — just a clear picture of what's possible for your practice.