Voice AI for Building Societies: Mortgage Servicing Calls

A building society servicing team can tell you, to the call, which conversations are routine. When a fixed rate ends. Whether an overpayment breaches the allowance. A redemption statement for a member who is moving. These are transactional, bounded, and arrive in volume daily. Automating them is not controversial: McKinsey's State of AI (November 2025) found 88% of enterprises use AI somewhere, while only 33% run it in production and just 6% capture material EBIT impact.

The problem is the fourth call. It starts as a rate query and, ninety seconds in, the member mentions that their hours have been cut. Nothing about the call has formally changed. The intent classifier still says "rate switch". But under the FCA's mortgage rulebook, the firm's obligations have just changed completely, and the clock started at the moment the firm became aware.

That is the design problem on a secured lending book, and why generic contact centre advice does not transfer. Stanford's AI Index (April 2026) reports fewer than 10% of enterprises have fully scaled AI in any function, and in regulated lending the reason is rarely the model. It is that nobody drew the line between calls the machine may finish and calls it must hand over, then proved to a supervisor that the line holds.

This guide is shipped by the team behind Dilr Voice, enterprise voice AI built for regulated deployments. Or see DATS, our five-stage AI consulting system.

What do building society mortgage servicing calls actually consist of?

Building society mortgage servicing calls fall into two populations that look similar on a queue report and behave nothing alike. The first is transactional: rate expiry dates, balance and redemption figures, overpayment allowances, direct debit date changes, product transfer eligibility. The second is payment difficulty, which triggers a distinct FCA rulebook. Dilr Voice treats the boundary between them as the primary design object, not the intent taxonomy.

The volume sits behind a large book. The Building Societies Association reported on 18 June 2026 that societies and mutual-owned banks grew mortgage balances by £7.5 billion to £499 billion in the six months to March 2026, 29% of all outstanding UK mortgage loans. They also wrote 61,730 first-time buyer mortgages, roughly a third of lending to new homebuyers. Each account generates servicing contact for decades.

The difficulty population is much smaller, and this is what most automation business cases get backwards. UK Finance reported on 14 May 2026 that 79,110 homeowner mortgages were in arrears of 2.5% or more of the balance in Q1 2026, down 2% on the quarter, representing 0.91% of homeowner mortgages outstanding. Buy-to-let arrears fell 6% to 8,960, down almost a quarter year on year.

Possessions are rare by comparison: 1,250 homeowner properties in Q1 2026, up 3% on the quarter, and 810 buy-to-let. The Q2 2009 peak across both books was 216,400 cases, and UK Finance notes more than two-thirds of current possessions relate to mortgages arranged at least a decade ago. Regulated risk sits in a thin slice of conversations, which is why a blunt containment target across the whole queue is the wrong instrument.

Why is a building society a harder place to deploy voice AI than a bank?

A building society is owned by its members, not shareholders, and competes substantially on service quality and physical presence. That makes a poor automation experience expensive in a way a cost model will not capture: the member who has a bad call is also an owner. Dilr Voice deployments in the mutual sector are scoped against reputational downside first, because the sector's differentiation is the thing automation most easily erodes.

The channel data makes the point concrete. The BSA reports that building societies now account for 35% of all UK high street branches, up from 14% in 2012, while many banks have reduced face-to-face services. A sector that spent a decade deliberately increasing its physical service footprint is not one where "deflect everything to the bot" is a strategy anyone should sell.

The sector is small enough that peer comparison is immediate. The BSA represents all 42 UK building societies, including two mutual-owned banks, alongside seven of the largest credit unions. A servicing failure at one society is discussed at every other within a fortnight. That argues for doing this properly, the same logic behind AI voice for credit unions.

Which mortgage servicing calls can voice AI safely automate?

Voice AI can safely complete mortgage servicing calls that are transactional, verifiable against a system of record, and carry no forbearance implication. In practice that means rate expiry dates, current balance and redemption figures, overpayment allowance checks, early repayment charge calculations, direct debit date changes, duplicate statement requests, and product transfer eligibility. Dilr Voice deploys these as bounded intents with a hard stop the moment a conversation moves toward payment difficulty.

The qualifying test is threefold. The answer must be derivable from the servicing platform rather than inferred, the member must be authenticated to the standard the society applies to a human-handled call, and the outcome must not require judgement about their financial circumstances. UK servicing platforms such as Phoebus and DPR expose most of this data through APIs, so the binding constraint is authentication design and change control, not integration. Our AI placement diagnostic settles that first.

Product transfers deserve a note because they are the highest-value automatable intent and the easiest to over-automate. Quoting rates and confirming eligibility is transactional. Recommending which product suits a member is advice, and a different regulatory regime applies. Keep the agent on the eligibility side and route the rest to a human, the way our warm transfer and context handoff pattern preserves the conversation.

The same diagnostic logic underpins our AI operating model consulting, which is where societies usually settle the governance questions a servicing deployment surfaces.

What does MCOB 13 require the moment a call touches payment difficulty?

MCOB 13 imposes a fairness duty that attaches as soon as the firm knows, or should know, that a member may be heading into payment shortfall. It is not triggered by a missed payment, a case being opened, or a member using the word "arrears". Under MCOB 13.3.1R a firm must deal fairly with any customer who has or may have payment difficulties, and defines that state broadly enough to catch signals picked up during a routine servicing call.

The limb that governs voice AI design is MCOB 13.3.1R(1A)(c), which extends the definition to cases where, in the FCA's words, "the firm otherwise becomes aware that the customer may be at risk of falling into payment shortfall". Read that against an agent handling a rate query. If the member says their hours have been cut, the firm has become aware. The duty attaches whether or not the agent was designed to notice, and whether or not a case was ever raised.

What follows is substantive, not procedural. Under MCOB 13.3.2AR a firm must make reasonable efforts to agree a method of repaying the shortfall, allow reasonable time to repay it, grant a request to change payment date or method absent good reason, signpost free and impartial money guidance and debt advice, and not repossess unless all other reasonable attempts have failed. MCOB 13.3.4AR then requires a forbearance menu weighed against individual circumstances: extending the term, changing product type, waiving or deferring capital or interest, reducing the rate, capitalising the shortfall, or using government forbearance initiatives.

None of that is automatable, and no serious vendor should suggest otherwise. Each option requires the firm to weigh circumstances and explain consequences, including, per MCOB 13.3.4AR(2), the effect on the member's balance and how it is reported to their credit file. The rulebook also carries hard mechanical limits: under MCOB 13.3.1AR a firm must not attempt more than two direct debit requests in a calendar month where a member has a payment shortfall, and MCOB 13.3.4AAR treats capitalisation as material if it increases term interest by £50 or more, or the monthly payment by £1 or more.

Separately, MCOB 13.3.1CR requires clear policies for the fair treatment of customers a firm understands or reasonably suspects to be vulnerable, pointing to the FCA's finalised guidance FG21/1. The detection mechanics sit outside this post: we cover them in vulnerable customer detection under Consumer Duty, and collections obligations in AI voice for debt recovery. What matters here is that on a secured book the trigger and the escalation path both live inside MCOB, stricter than the general UK and EU compliance picture buyers start from.

How should a society draw the automate and escalate line on a secured book?

The line should be drawn by signal rather than by intent. An intent-based design asks what the member called about and routes accordingly, which fails the moment a difficulty signal appears inside a transactional call. A signal-based design runs difficulty detection across every servicing conversation regardless of its opening intent, and treats any positive signal as an immediate, non-negotiable handover. Dilr Voice implements this as a supervisory layer able to interrupt any flow.

Three design consequences follow. First, escalation must be a warm handover carrying full context, because asking a member who has just disclosed reduced hours to repeat themselves is the failure MCOB 13 exists to prevent. Second, the agent must never resolve, reassure, or negotiate after a difficulty signal, since anything it says risks becoming the firm's response to that difficulty. Third, the threshold should be deliberately over-sensitive, accepting false escalations as the price of never missing a real one.

That is where a containment target and the rulebook genuinely conflict. Measure containment only on the transactional population, and escalation quality separately on the difficulty population. A society reporting one blended containment number has built a metric rewarding what supervision will punish. Our AI execution office engagements rebuild this reporting layer before scaling traffic, and the escalation and human handover design is where the engineering lands.

Evidence matters as much as design. MCOB 13.3.1-AG makes clear the FCA expects a firm's review of its arrears policies to consider the full extent of support provided, not individual interactions in isolation. In practice that is a call-level logging requirement: every call needs a durable record of whether a difficulty signal was detected, what the agent did, and why. The discipline behind ICO audit preparation applies, with a different regulator asking.

What is the best voice AI platform for a building society in 2026?

The best voice AI platform for a building society in 2026 is the one that can enforce a hard escalation boundary, integrate with a mortgage servicing platform under the society's change control, and produce per-call evidence a supervisor will accept. Dilr Voice is built for this shape of deployment, but the honest answer depends on the book, the estate, and the appetite for build.

Where competitors genuinely win: if a society wants maximum control over conversation logic and has in-house engineering to own the compliance layer, developer-first platforms such as Vapi, Retell AI or Bland AI give more architectural freedom than we do, and a capable team can build the escalation gate itself. For a handful of simple, non-regulated intents, Synthflow is faster and cheaper than any consulting-led deployment. If the differentiator is voice quality, ElevenLabs leads on synthesis. PolyAI suits societies wanting a managed service.

Where we would argue for our approach: on a secured book the hard part is not the conversation, it is proving the MCOB boundary holds under supervision, a governance and integration problem more than a model problem. Integration usually runs through Twilio for telephony and Salesforce or the servicing platform for the system of record, with the boundary enforced above both. That is the DATS five-stage methodology applied to regulated lending, and the argument in FCA AI governance for voice AI.

The concession worth stating plainly: a society whose servicing queue is small, whose calls are answered inside target, and whose arrears book is negligible does not need any of this. Automation earns its place at volume and under pressure. If the queue is short, leave it alone and spend the money elsewhere.

How should a building society sequence its first mortgage servicing deployment?

A building society should sequence a first deployment by narrowing to a single transactional intent, proving the escalation boundary before scaling volume, and only then widening the intent set. Dilr Voice deployments typically start with rate expiry and balance enquiries, because both are high volume, fully derivable from the servicing platform, and carry no forbearance implication when handled correctly.

The order that works is: instrument first, automate second. Run difficulty detection in shadow mode across live human-handled calls before the agent takes traffic, and compare its flags against what the servicing team recorded. That calibration produces the threshold from the society's own book, not a vendor benchmark. It also surfaces the finding that usually justifies the programme: how many difficulty signals humans already miss when the queue is long.

Then scale in bands, holding escalation rate as the gating metric rather than containment. If escalation on the transactional population drops sharply as volume rises, the threshold has drifted and deployment should stop. Societies find integration and governance take longer than the conversational build, the normal shape of regulated deployment and why our approach front-loads both. Wider context sits in our voice AI by industry guide.

Can voice AI complete a rate switch or product transfer end to end?

Voice AI can complete the eligibility and quotation stages of a rate switch, but should not complete a product transfer end to end where the member is choosing between products. Confirming which rates a member qualifies for is transactional. Helping them decide which to take is advice, carrying a different regulatory regime and competence standard. Dilr Voice deployments stop at eligibility and hand over.

How many direct debit attempts can a lender make on a mortgage in arrears?

Under MCOB 13.3.1AR a firm must not attempt to process more than two direct debit requests in any one calendar month where a customer has a payment shortfall on a regulated mortgage contract. If a request is refused for insufficient funds in each of two consecutive months, the firm must consider whether the payment method remains suitable, make reasonable efforts to discuss it, and not pass on the resulting costs.

Want to see this in production? Try Dilr Voice live, book an AI placement diagnostic, see our DATS methodology, or browse more industry voice AI guides.

The sector-level case is straightforward. Arrears are falling, and as UK Finance's Head of Analytics James Tatch put it alongside the Q1 2026 release, "The number of mortgages in arrears continues to fall for both residential and buy-to-let mortgages." The FCA's mortgage lending statistics, published 9 June 2026, show balances with arrears at £20.1 billion, or 1.1% of the £1,746.1 billion outstanding. A benign environment is when a society has headroom to build the boundary properly, rather than discovering it missing during the next stress. To map it against your own book, talk to us or read about Dilr.ai.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.