White papers / White paper 09

AI Governance Retrofits: Making Existing Models Explainable and Auditable

Audit the fraud and credit models you already run, and give your risk officers a live view of how they decide.

White paper 09 · November 2026 · Updated 17 November 2026Download PDF

Summary

Banks and fintechs moved quickly to use machine learning for fraud detection and lending decisions. Many of those models now run in production with little documentation, and the people who built them have sometimes moved on. Regulators in the UK, EU and US are asking for the same things: evidence of how models decide, whether they treat people fairly, and who is accountable.

Fintech Den™ already builds explainable risk engines from scratch. This paper is about the models you already have. We audit existing models, including vendor models where you can get access, explain how they reach their decisions, test them for bias and drift, and build a governance dashboard so that risk officers, not only engineers, can see how decisions are being made.

We're a new company, and we'd rather be straight with you about that. This paper sets out the problem, how we'd approach it, the rules in the UK, EU and US, how we handle your data, and a first project small enough to judge us on.

Data protection is usually the first thing UK and EU buyers check, so we'll say it up front: we comply with the UK GDPR and the EU GDPR, and we're registered with the UK Information Commissioner's Office. The GDPR section below explains what that means for your project, from where your data is hosted to how we'd tell you about a breach.

The problem

Many production models are hard to explain. They were trained quickly, on data that has since changed, and their documentation describes what was intended rather than what the model does today. When an auditor or examiner asks why a particular customer was declined or flagged, the honest answer can take weeks to produce.

The rules are tightening. Under the EU AI Act, credit scoring of individuals is classed as high-risk, with obligations for risk management, logging, human oversight and documentation, and a right for people to an explanation of decisions that affect them. In the UK, the PRA's model risk principles set expectations for how banks govern their models. In the US, lenders must still give specific reasons when they decline credit, whatever kind of model made the decision.

Fixing this usually falls to data teams who are already busy, and the output is often a document for engineers rather than something a risk committee can use. Risk officers end up relying on others to tell them what their own models are doing.

How we would approach it

We'd start with an inventory of every model that affects customers: what it decides, what data it uses, who owns it and what evidence exists. Then we'd audit the highest-risk models first and build the reporting and monitoring that lets your risk function oversee them day to day.

In most cases the build has three parts:

PartWhat it doesWhat we'd aim for
AI compliance auditReverse-engineers and tests existing models, including black-box and vendor models where access allows, for explainability, bias and driftA clear picture of how each model decides, with evidence for regulators
Explanations per decisionReason codes and plain-language explanations added to each decision, without retraining where possibleAnswers for customers, complaints teams and examiners
Governance dashboardA live view of decisions, approval rates, fairness measures and drift for risk officers, with alertsOversight by the people accountable for the models
Existing models are audited, then governed in one place. Your current models; Fraud scores; Credit decisions; In-house and vendor models; →; Audit; Model inventory; Explainability tests; Bias and drift checks; →; Governance dashboard; Live view for risk officers; Reasons for each decision; Alerts on drift; Evidence pack: documentation, test results and decision logs ready for regulators and auditors

The diagram shows the flow. The audit establishes what each model does and how well it can be explained. The dashboard then keeps watching it in production, so a change in behaviour shows up as an alert rather than as a finding in next year's audit.

Not every model can be made explainable enough. Where a model can't meet the standard its use requires, we'll say so plainly and help you decide whether to constrain it, add human review, or replace it with something simpler. Our explainable risk engine is one option, not the default answer.

AI rules in the UK, EU and US

The rules differ by market, but they ask for the same core things: know your models, explain their decisions, test them for fairness and keep records. This is a summary for discussion, not legal advice.

MarketRuleWhere it stands (November 2026)What it means for the build
EUAI Act: credit scoring is classed as high-risk AIAfter the Digital Omnibus, approved by Parliament on 16 June and the Council on 29 June 2026, high-risk obligations apply from 2 December 2027Risk management, logging, human oversight and documentation, built in now
EUAI Act Article 86: right to an explanation of individual decisionsApplies to decisions based on high-risk AI systemsAn explanation for each decision that a person can understand
UKPRA model risk management principles (SS1/23)In effect since 17 May 2024A model inventory, validation and governance that cover AI models
UKUK GDPR rules on automated decisions, as amended by the Data (Use and Access) Act 2025Reforms coming into force in stages during 2026Information for customers, human review and a way to contest decisions
USECOA and Regulation B adverse action noticesSpecific reasons for declines are still required; the CFPB withdrew its AI-specific circulars in May 2025Reason codes that hold up for every declined applicant

UK and EU GDPR compliance

Auditing models means looking at the personal data they were trained on and the decisions they make about people, so we keep access tightly controlled. We comply with the UK GDPR, alongside the Data Protection Act 2018, and with the EU GDPR. On client projects where we handle personal data, you're the controller and we act as your processor. Here's what that means in practice.

AreaWhat we do
Data processing agreementWe sign one with every client, covering everything Article 28 of the UK and EU GDPR requires. We act only on your documented instructions, and everyone on your project is bound by confidentiality. Our standard agreement is ready for your legal team to review.
Collecting lessSystems collect only the personal data they need. By default we build and test with synthetic or masked data, not live customer records.
Where data livesProduction systems run in your own cloud account, in the UK or EU region you choose.
Automated decisionsWhen a fraud or credit decision made by automated means alone significantly affects someone, UK and EU rules set conditions and require safeguards: people must be told, and be able to challenge the decision and have a person review it. Every decision in our designs carries a readable reason and a route to analyst review.
AreaWhat we do
Access and transfersAny access to real personal data is agreed with you first, limited to named people, logged, and removed when the task is done. If access ever involves a transfer outside the UK or EU, it's covered by the EU standard contractual clauses or the UK international data transfer agreement. Both are ready to sign.
SecurityEncryption in transit and at rest, least-privilege access, logging, and code review on every change.
Sub-processorsWe tell you which sub-processors we use, and add new ones only with your prior authorisation.
People's rightsWe help you respond when people ask to see, correct or delete their data.
BreachesIf we become aware of a personal data breach affecting your data, we tell you without undue delay, so you can meet the 72-hour deadline for reporting it to the ICO or your EU supervisory authority.
Assessments and auditsWe support your data protection impact assessments, give you the information you need to show compliance, and allow audits.
End of the workWe return or delete your personal data, whichever you choose.

Where we stand on registrations and certifications. We'd rather be precise than impressive.

ItemWhere it stands
ICO (UK)Registered with the Information Commissioner's Office.
EU to UK data flowsThe EU recognises UK data protection law as adequate. The European Commission renewed that decision in December 2025, and it runs until December 2031.

For EU firms under DORA. If you're an EU financial entity, our DORA information sheet gives you what you need to record us in your register of ICT third-party providers: our company details, the services we provide, where data is processed and the subcontractors we use. We'll also work through the DORA contract terms with your team.

How we work

Since we can't yet show you years of past projects, we'd like to show you our process instead, in writing, before you commit.

The people you'll work with. You deal directly with our leadership. There's no account manager in between.

RoleWhat they look after on your project
Chief Executive OfficerThe relationship, scope, security and data protection, and fit with your regulatory obligations
Chief Technology OfficerAudit method and model testing, explainability approach, technical risk, final technical sign-off, and on some projects the relationship and scope
Head of EngineeringDecision logging, monitoring pipelines, testing and releases
Head of Visual DesignThe governance dashboard, risk committee reports and analyst screens

Company and working hours. Fintech Den™ is a trading name of ROMBI PIXEL LTD, a UK company registered in England and Wales (no. 15947340) with its registered office in Bury St. Edmunds. We run on Central European Time (CET), so our working day overlaps with business hours across the UK and EU. For US clients, our afternoon overlaps with the East Coast morning, and we schedule calls in that window.

The practical side. We quote and invoice in euros, US dollars or pounds, whichever suits your finance team, and EU clients can pay by SEPA transfer to our euro account. Proposals spell out scope, milestones, payment points and VAT treatment, and payments are tied to milestones you can check. Our master services agreement is ready for your legal team, and we're open to discussing governing law and other terms.

Habits you can hold us to. We send a written summary within one working day of every call. You get a short written update every week and a demo every two weeks. Risks go in a shared log as soon as we spot them, and nothing changes scope without your written agreement.

What the first 90 days usually look like. We'd inventory your models, audit the highest-risk one in depth, and switch on the monitoring that keeps it in view.

From black box to a reviewable model in 12 weeks. Weeks 1 to 3; Inventory; List every model in use; Collect data and documentation; Agree the audit scope; Weeks 4 to 8; Audit; Explain decisions case by case; Test for bias and drift; Gaps against the rules; Weeks 9 to 12; Govern; Dashboard for risk officers; Decision logging switched on; Evidence pack drafted; Inventory signed off; Findings reviewed; Decide on fixes; Every week a written update. Every two weeks a demo. No scope change without your written OK.

You can stop at any checkpoint and keep everything we've produced so far, including the inventory and the audit findings.

Starting a conversation

The first call is mostly us listening. We'll probably ask about:

  1. Which models affect customers today, and who owns each one.
  2. Whether any of them are vendor models you can't see inside.
  3. What your auditors, supervisors or partner banks have asked about AI recently.
  4. Whether your risk committee gets regular reporting on model behaviour.
  5. Where your data has to be hosted.

If it looks like we can help, we'll suggest one of these as a first step. Timings are typical, not fixed, and we'll confirm them in a written proposal.

OptionTypical lengthWhat you'd receive
Model inventoryAbout 3 weeksA list of every model that affects customers, its risk level and the evidence that exists, with a fixed-scope audit plan
Model auditAbout 2 monthsAn in-depth audit of one high-risk model, with explanations, bias and drift tests and a gap list
Governance dashboardAbout 2 monthsA live view for risk officers with decision logging, fairness measures and alerts

If it doesn't look like a good fit, we'll tell you, and where we can, we'll point you to someone better placed.

Before the call, we're happy to send our trust pack: our data processing agreement and the EU and UK transfer clauses, plus the DORA information sheet if you're an EU financial firm. We are working towards IASME Cyber Assurance and the NCSC-backed Cyber Essentials certification.

To arrange a call, email contact@fintechden.com. We'll reply within one working day.

Common questions

Can existing fraud and credit models be made explainable without rebuilding them?

Often, yes. Reason codes and plain-language explanations can be added to each decision, without retraining where possible. Where a model can't meet the standard its use requires, we say so and help you decide whether to constrain it, add human review, or replace it with something simpler.

Can you audit vendor or black-box models?

Yes, where access allows. The audit reverse-engineers and tests existing models, including black-box and vendor models, for explainability, bias and drift.

When do the EU AI Act's high-risk rules apply to credit scoring?

After the Digital Omnibus, high-risk obligations apply from 2 December 2027. Credit scoring of individuals is classed as high-risk under the Act.

How long does an AI governance retrofit take?

A model inventory takes about 3 weeks. An in-depth audit of one high-risk model takes about 2 months, and a governance dashboard for risk officers takes about 2 months.

Sources

More white papers

Ready to Talk About Your Project?

Book a discovery call and we'll mostly listen. Not sure which service fits? Start with a fixed-price Discovery Sprint. We work on Central European Time and reply within one working day.

Discovery call Contact