AI Implementation in Mortgage Origination: Challenges and a Practical Approach

The AI opportunity in mortgage origination today

“What are the blockers stopping AI adoption in mortgages at the moment?”

It’s a deceptively simple question, because the easy answers of “models aren’t good enough” or “documents are too messy.” aren’t correct. Instead, Oper Founder and CEO Geert van Kerckhoven argues:

“The only hurdle today is that, from a compliance and a risk point of view, there is a limited visibility on regulations… If it's not clear, the answer is: let's not do it.”

Technically, the building blocks for AI-enabled mortgages are already here:

• OCR and layout‑aware vision–language models can reliably extract fields from payslips, tax statements, and property appraisals in controlled conditions.
• Open-banking and national data rails (like the Dutch HDN and UWV income verification) can validate income and property attributes in minutes.
• Large language models (LLMs) can read underwriting manuals, explain policies, and even assemble decision rationales.

What governs the pace in Europe is not model capability, but three things:

1. Regulatory clarity – The EU AI Act explicitly classifies AI used to evaluate creditworthiness as high-risk, with heavy obligations.
2. Governance and auditability – Supervisors expect documented data lineage, model oversight, and human control, following frameworks like the EBA’s Loan Origination and Monitoring (LOM) Guidelines.
3. Data reality – In some markets, everything’s on APIs; in others, income still arrives as scanned PDFs or even faxes.

The thesis: phased, compliance‑first operationalization

The temptation is to aim straight for “one-click mortgage approval”; walking into a house and getting an offer in real time.

“I think there is no real hurdle there technically… But if you immediately want to do that, then there's a lot of compliance work that needs to come first.”

Instead, Oper argues for a different route:

• Start with low‑risk, high‑ROI, narrow use cases that do not themselves make or materially influence the credit decision (e.g., document classification, key‑field extraction with human review, policy parsing, product pre‑screening).
• Chain these automations over time, building the governance, audit trails, and comfort that will eventually support end‑to‑end underwriting.
• Design everything compliance‑first, aligned with the EU AI Act, EBA LOM, DORA, GDPR, and national supervisory expectations.

You can create meaningful efficiency gains in 12–24 months, and quietly prepare your organization for fully automated decisions when regulation and internal governance catch up.

What’s Actually Blocking AI in Mortgages

Regulatory ambiguity and model‑risk realities

Under the final text of the EU AI Act, AI systems used “to evaluate the creditworthiness of natural persons or establish their credit score” are explicitly listed as high‑risk. That classification brings with it stringent obligations:

• Risk management and control frameworks
• Data governance and quality processes
• Detailed technical documentation and logging
• Transparency to users and clear human oversight
• Accuracy, robustness, and cybersecurity standards

At the same time, Article 6(3) clarifies that narrow, procedural systems that do not materially influence a decision - for example, classifying documents or transforming a PDF into structured data with human validation - can remain outside the high‑risk category.

The challenge is that most banks don’t yet have a robust internal playbook for deciding what is, and isn’t, “material influence.” So the safest answer from compliance and risk teams has often been: “Let’s not do it yet.”

Post‑crisis risk posture in European banks

That caution didn’t emerge in a vacuum. After the 2008 financial crisis, European supervisors doubled down on governance, documentation, and model risk management. The EBA’s Loan Origination and Monitoring Guidelines codified this for credit processes: banks must show that each loan decision is based on sufficient, verified data, and that any models used are well understood and controlled.

Or, more directly put:

“If it's not clear, the answer is let's not do it.”

With DORA now in force, ICT risk, logging, and third‑party oversight requirements are even tighter. That’s another reason why AI pilots that touch input data (classification, extraction) are simply easier to defend than those that directly make or override credit decisions.

Not a tech problem: LLMs are capable, but governance lags

From a pure technology perspective, we’re much closer to automated underwriting than many assume:

• Layout‑aware document models achieve high F1 scores on public key‑information‑extraction benchmarks like SROIE and FUNSD.
• LLMs can interpret long, messy policy texts and output machine‑readable rules plus human‑friendly explanations.
• Orchestration tools can chain these steps into full workflows.

That technology capability gets us much of the way there:

“To get to a fully automated underwriting, you need to be able to create a profile of a client, read the documents… A lot of these steps can already be automated with AI.”

The blockers, then, are:

• Governance maturity (model inventory, testing, monitoring)
• Documentation and explainability at a regulator‑grade standard
• Comfort in risk and compliance that AI is controlled, not a black box

Where AI Can Create Value Now (Low‑Risk, High‑ROI Use Cases)

Document classification and key‑field extraction

“The Achilles heel today is very easy. It's the extraction out of paper documents.”

In API‑rich markets, much of the mortgage payload is machine‑readable. In others (such as Germany, parts of Benelux, and Southern Europe) income and property evidence still arrives as PDFs or scans. That’s precisely where AI can create immediate value without stepping into high‑risk territory.

What to do now:

• Automatically classify incoming documents (payslip vs. bank statement vs. appraisal).
• Extract key fields—names, addresses, gross/net income, dates, loan amounts—with confidence scores.
• Route low‑confidence or critical fields to human reviewers.

Because this is data transformation with human‑in‑the‑loop, and not an autonomous credit decision, it can be structured to remain outside the AI Act’s high‑risk category - provided humans validate before the data is used for underwriting.

Policy extraction and rules application

Most underwriting manuals are dense PDFs that live in shared drives and underwriters’ heads. LLMs can:

• Read these manuals and extract eligibility rules (e.g., LTV caps, minimum tenure, income stability conditions).
• Convert them into machine‑readable checks (if/then rules or decision tables).
• Attach plain‑language rationales to those checks.

Underwriters remain in control: AI prepares the rule set and applies it, but humans approve both the rules and any exceptions. In that form, it’s support and pre‑analysis, not automated decision‑making.

Product fit and eligibility screening

One of the easiest wins is intelligent pre‑screening:

• Given a borrower profile and extracted data, AI filters the product set to “likely eligible” options.
• It can propose required condition based on the rules above.
• An underwriter or advisor confirms the recommendation.

This doesn’t replace your credit policy; it reduces noise and rework, so specialists spend their time on real edge cases.

Data prefill from public and partner APIs

In more digital markets, AI and APIs reinforce each other:

• In the Netherlands, the HDN standard and UWV’s Inkomensbepaling Loondienst (IBL) allow lenders to verify salaried income directly from social‑insurance records—now used in hundreds of thousands of cases per year.
• Property and address data can be pulled from Kadaster/PDOK APIs to prefill collateral details.
• In Estonia and parts of the Baltics, the X‑Road infrastructure enables rapid, secure sharing of identity and registry data.

These aren’t “AI use cases” in isolation, but AI orchestration (e.g., an intake assistant that knows which APIs to call and when) can turn them into a fully guided, low‑friction origination experience.

Market Dynamics Creating Urgency

Consolidation and scale without headcount

Across Europe, banks are consolidating mortgage books and seeking scale—but they’re not planning to double headcount as volumes grow.

“I do not hear the narrative anymore of ‘if I use technology, I can slash my headcount by X.’ … It's more like, I'm not finding the people. Volumes are back, we would love to double production and we don't want to double analysts.”

External benchmarks back the opportunity: European “average” mortgage time‑to‑cash sits around 40 working days, while top performers are closer to 18 days. Reducing touches per file and cutting rework loops is not a nice-to-have; it’s the only way to hit those numbers without burning out staff.

Talent scarcity and backlog pressure

AI, in this context, is not a headcount‑reduction tool; it’s a throughput multiplier:

• Fewer manual data entry tasks for underwriters.
• Fewer back‑and‑forths with customers and brokers for missing documents.
• Cleaner files arriving at the credit desk, with pre‑assembled rationales.

That’s why we’re seeing a genuine inflection in operational seriousness: banks are finally treating their back office as a competitive differentiator, not just a cost center.

What to do next

If you’re a Head of Mortgage Operations, a Chief Risk Officer, or an AI transformation lead, the path forward is surprisingly concrete:

1. Pick two low‑risk use cases to start:
   • document classification + key‑field extraction, and
   • policy extraction + eligibility pre‑checks.
2. Bring risk and compliance in from day one. Map use cases to the EU AI Act, EBA LOM, DORA, and national guidance. Document why they are preparatory and non‑high‑risk.
3. Stand up a basic QA and lineage framework. Dual extraction for critical fields, sampling, clear thresholds, and full data lineage.
4. Measure what matters. Time‑to‑offer, touches per file, extraction accuracy, rework rates, and customer/broker satisfaction.
5. Iterate and chain. Once the first steps are stable, add policy checks, product screening, and richer decision support—always with human oversight and clear governance.

Fully automated underwriting is not science fiction anymore; it’s technically within reach. The question is not if you get there, but how—and whether you choose a path that builds trust, withstands regulatory scrutiny, and delivers real operational value at every step along the way.