Part two of a three-part series on the agentic bank written by Oper’s CEO and Co-Founder Geert van Kerckhoven.

The front office gets the headlines. But if you ask me where AI delivers the most immediate, measurable return in mortgage lending right now, the answer is the mid-office and back-office. And if you ask where the most transformative long-term opportunity lies, the answer surprises most people: it's in what happens after the loan is booked.

Document Intelligence at scale

The core bottleneck in mortgage origination hasn't changed in decades: documents. Collecting them, reading them, validating them, cross-referencing them against policy, sending them back when something's missing, waiting for resubmission, repeating the cycle.

Agentic AI fundamentally breaks this bottleneck. Agents parse PDFs, bank statements, ID proofs, property documents, and income records. They classify files, detect missing pages, and extract entities - names, employer details, cashflows, account balances - across multilingual, multiformat inputs. Not the brittle OCR of five years ago. Real comprehension: understanding what a document means, not just what it says.

The numbers are striking. AI-enabled underwriting can process 70–85% of credit applications without human intervention. Agentic workflows reduce per-loan processing costs by 35–50%. Freddie Mac estimates that lenders fully utilising AI automation can save up to $1,500 per loan and shorten production cycles by five days on average. Automation Anywhere reports that agentic process automation slashes processing time by 88%.

And this isn't pilot-stage. According to Celent, 83% of lenders plan to increase their generative AI budgets in 2026. Two-thirds have already completed or will implement GenAI strategies this year. The market has moved decisively past experimentation.

For both conventional and non-conventional lending - where document complexity and edge cases multiply - we'll see the most dramatic progress. Tool use makes it trivial for agents to query external databases, verify employment records, cross-reference tax filings, and check regulatory registries. The mid-office and back-office analyst who previously spent hours assembling a dossier will spend minutes reviewing one.

The AI credit committee

Here's a concept that sounds radical until you think about it for five minutes: multi-agent credit deliberation.

Today, a credit committee is a group of experienced professionals who review a dossier and debate whether a borrower is creditworthy. They bring different perspectives - risk appetite, regulatory awareness, portfolio balance, relationship context. The process works, but it's slow, expensive, and constrained by human availability.

Now imagine multiple agents - potentially running on different models, with different specialisations - conducting that same deliberation. One agent analyses financial risk. Another evaluates the business model. A third checks regulatory compliance. They surface their findings, flag disagreements, and produce a structured recommendation. McKinsey describes these as "multiagentic squads" that facilitate full credit review workflows, and reports that one financial institution is already using a multiagent system to draft financial-risk assessments for corporate clients.

The human doesn't disappear. The human becomes the final reviewer - ensuring guardrails have been respected, checking for discrimination, exercising judgment on edge cases. But the preparation that previously took days now happens in near real time. McKinsey estimates 40–80% productivity uplift per use case, with approximately 30% faster decision-making.

The compliance dimension is critical. AI systems now analyse hundreds or thousands of variables compared to traditional models that use fewer than 20 data points, achieving 25–50% uplift in loan approvals without additional risk. But with that power comes responsibility. The $89 million in penalties levied against Apple and Goldman Sachs for algorithmic discrimination concerns in 2024 underscore why the human-in-the-loop guardrail isn't optional - it's existential. Best practice: establish a Model Risk Committee with authority over approvals, performance reviews, exception handling, and sunset decisions.

Collateral through a new lens

For collateralised lending - mortgages above all - multimodal AI also transforms how you evaluate the collateral itself.

Computer vision can now assess property conditions from imagery at scale. Restb.ai's partnership with HomeVision, announced March 2025, uses computer vision to detect inconsistencies between appraisal data and property imagery, automatically analysing property conditions and unlocking visual insights. Over 70% of real estate firms now use AI-driven models for valuation. Fannie Mae and Freddie Mac approve automated valuation models for specific mortgage products, with Fannie Mae reporting that standardised data collection saves consumers $350-$400 over traditional appraisals.

The CFPB approved new rules in June 2024 on AI/algorithmic use for home appraisals, establishing quality standards for automated valuation models. The regulatory framework for AI-driven property valuation exists. The technology is mature. And for lenders processing  thousands of mortgage applications, the combination of automated document analysis and automated property assessment creates a fully agent-mediated path from application to credit decision.

Beyond RPA: Lifecycle management

Most conversations about AI in lending stop at origination. That's a mistake. The most transformative implications may be in what happens after the loan is booked.

Today, lifecycle management is a world of forms. A borrower wants to swap collateral - form. A property gets split because of a divorce - form. Someone passes away and the estate needs to restructure - form. Each form triggers a registration on the core banking system. RPA has already automated many of these procedural tasks, and credit where it's due: it helped.

But RPA is brittle. It follows scripts. It breaks when the process changes. It can't handle ambiguity, exceptions, or tasks that require understanding context.

Agentic AI is different. An agent that understands the lifecycle event, knows the procedures, has access to the tools - the core banking system, the document management platform, the regulatory databases - and can reason through the specifics of each case. A borrower sends an email requesting a modification. The agent picks up the email, interprets the request, verifies eligibility, executes the task across systems, sends confirmation, schedules follow-ups. Not a predefined workflow - an adaptive process that handles the variability of real life.

AI-driven post-close audits can now review 100% of loans versus the previous 10% sample, reducing repurchase exposure and audit preparation by approximately 50%. Fannie Mae reports a 29% average decrease in operational costs for lenders using AI, and a 50% reduction in mortgage fraud cases through ML-based detection.

The cost-of-tokens thesis

Here's the argument that I think changes everything in lending - and it has nothing to do with origination.

When the cost of a loan modification is measured in tokens, not hours, the entire product design space opens up.

Today, loan modifications are expensive. They require human analysts to review, process, and register changes. So lenders minimise them. Many tasks that would genuinely improve borrower outcomes are simply deleted from the product because the operational cost doesn'tjustify the benefit. Post-renovation rate adjustments? Too expensive to process manually. Dynamic amortisation based on life events? Operationally impractical. Continuous energy-efficiency monitoring that triggers rate improvements? Nobody has the staff for that.

But when the marginal cost of processing a modification drops to fractions of what it costs today - because agents work 24/7, scale elastically, and run at token prices that decline year over year - those deleted tasks come back. And with them, entirely new product categories.

Dynamic amortisation tables

Imagine a mortgage where the repayment schedule adjusts automatically based on life events - job loss, salary increase, parental leave, retirement. The agent monitors, verifies, and modifies. No forms. No waiting.

Energy-efficiency repricing

A borrower renovates their property, improving its energy rating. The agent detects the change through utility data, renovation permits, or updated energy certificates, verifies it, and adjusts the interest rate accordingly. A task that was impossible when it required human processing becomes a standard product feature.

Continuous post-close monitoring

After-checks that were cut for cost reasons - income verification, property condition, insurance compliance - become standard when they cost tokens instead of hours. The lender's risk management improves, and the borrower gets better terms because the lender has betterinformation.

‍

Forrester research confirms the direction: 70% of financial services respondents anticipate using agentic AI to deliver tailored experiences previously available only to high-net-worth individuals - including automatic refinancing when rates decrease and proactive portfolio management. One financial services organisation already has 60 agentic agents in production, with plans for over 200 by 2026.

The implications for affordability are significant. More flexible products, more responsive servicing, and lower operational costs all feed into a mortgage landscape where lending can be more affordable, more personalised, and more adaptive to the realities of borrowers' lives. That's not incremental. That's a different kind of product entirely.

Next in this series: What actually needs to be true for all of this to work - model selection, sovereign deployment, and the EU AI Act deadline that's closer than most banks think.

Geert is CEO of Oper Credits, a pan-European mortgage origination platform. Herman is Oper's AI agent for residential lending, already in production with 22+ financial institutions across 6 EU countries.