Sophie Hope’s mortgage application did not fail because she was a risky borrower. It failed because a number attached to her name had been quietly corrupted by people she had never met. The Brighton-based nail technician and content creator had an excellent rating, then found herself turned down by a lender and staring at a score in the lowest band. Someone had opened catalogue accounts in her name โ a games console, sportswear, assorted goods โ and when those accounts were passed to a secondary finance company, the resulting debt sat there unnoticed for three to four years. This is the defining flaw of modern credit assessment in miniature: the record is treated as fact, while the human behind it is treated as a footnote. The story matters far beyond one household budget, because it exposes a structural contradiction. Lenders are told to assess affordability, yet the first filter most applicants meet is a retrospective score that cannot distinguish between a person who mismanaged money and a person whose identity was stolen. Practical takeaway number one: your credit file is not a passive document. It is an active claim about you, made by third parties, that you have a legal right to inspect, correct, and contest โ and the only moment most people inspect it is when they have already been refused something they needed.
To understand why this happens, it helps to know who is doing the scoring. Three agencies dominate the UK market โ Experian, Equifax and TransUnion โ and they are fed from a broadly similar pool of raw material: electoral roll entries, credit account records supplied by lenders, payment histories, court judgments and insolvency data. Each agency applies its own model to that material, which is why a person can hold three different scores that seem to disagree about their character. The consumer-facing number, however, is largely a teaching device. Most lenders do not consume it directly; they run their own scorecards and affordability calculations on top of the underlying data. That distinction is lost on most borrowers, who optimise for a dashboard figure that no underwriter actually reads. The analytical point here is subtle but commercially important: the score is a summary, the file is the evidence, and the decision is a separate act of judgement entirely. If you want to influence outcomes, work on the evidence. Register on the electoral roll, correct stale addresses, challenge mis-keyed defaults, and make sure any account closed years ago is genuinely marked as closed. Chasing a number without auditing the file beneath it is like repainting a car with a failing engine.
The sharpest critique of the current system is that it is looking backwards through a windscreen while claiming to describe the road ahead. Funmi Olufunwa, a consumer finance lawyer who founded the education business Hoops Finance, puts it plainly: a score reports what has already occurred, which is not the same thing as forecasting what is affordable next month. That gap has real consequences for what the industry calls thin-file applicants โ young adults, recent arrivals to the UK, people who rent and pay utility bills dutifully but have never taken a loan, and survivors of coercive relationships whose finances were never held in their own name. Being invisible to the system can be almost as damaging as being visibly unreliable, and neither outcome reflects genuine creditworthiness. There is also a philosophical problem at the heart of the exercise: a thin file is not evidence of risk, it is evidence of limited observation. Lenders know this, which is precisely why the industry’s growing interest in richer data is not mere fashion. For consumers, the actionable insight is to manufacture visibility deliberately. A credit-builder card cleared in full each month, a mobile contract in your own name, a registered address that matches your bank records, and consistent direct debits will do more for your file than any short-term score optimisation trick circulating on social media.
Into this gap has stepped a cohort of machine-learning lenders who argue that the traditional file is an unnecessarily narrow window. Upstart, an American firm, has built its proposition on models that weigh more than two and a half thousand variables, and claims this breadth allows it to approve more loans at lower prices across demographic groups; Zest AI offers lenders a comparable alternative for credit union and bank underwriting. The mechanics are familiar to anyone who has worked with modern predictive systems: gradient-boosted ensembles trained on historical loan outcomes, combining bureau data with employment and education signals, income estimates and repayment behaviour. The economic logic is compelling, because narrower risk segmentation means better pricing and a wider funnel at the same cost of manual review. But there is a technical trap that deserves more attention than it usually gets. These models are trained on loans that were actually made, which means they learn from the decisions of the previous system as much as from real-world outcomes. Statisticians call this selective labels, and it means an algorithm can inherit the blind spots of the scorecard it was meant to replace while sounding far more sophisticated. Any lender piloting alternative underwriting should interrogate sample selection before celebrating approval-rate gains.
The incumbents are not standing still, and their movements tell you more about the market’s direction than any startup pitch deck. Experian has partnered with OpenAI to let consumers check their score inside a conversational assistant, Equifax has launched Amplify AI for sharper credit assessment and fraud detection, and TransUnion has announced integration of Google’s Gemini models into its proprietary analytics. Notice what is actually being bought and sold here: distribution and interface, as much as predictive power. The bureaus already own the most valuable asset in this market โ authorised access to lender-reported data โ so their competitive advantage lies in becoming the front door through which consumers and lenders meet, not in winning a modelling contest. That has a clear implication for anyone building in this space: the moat is regulatory permission, data rights and consumer trust, not clever architecture. There is also a governance risk that grows as interfaces become conversational. The moment an assistant paraphrases why someone was declined, imprecision becomes a compliance event, and a confident-sounding explanation that is subtly wrong is worse than no explanation. Firms should treat the conversational layer as a presentation shell over an auditable decision engine, never as a substitute for one.
The counter-argument is forcefully made by Sandra Wachter, professor of technology and regulation at Oxford, who warns that machine learning cannot know about the social inequalities embedded in the data it consumes and may simply reproduce decisions that were themselves unfair. That is not a hypothetical worry. Historical lending patterns in the United States and postcode-driven proxies in the UK have both encoded disadvantage into the record, and a model trained on that record will treat the pattern as a signal. Worse, some of the most predictive variables are uncomfortably close to protected characteristics โ shopping behaviour, device type, the stability of a postcode โ so bias can travel through a legitimate-looking feature with no obvious fingerprint. The practical response is not to abandon automation but to define fairness before building. Pick a fairness metric, whether demographic parity or equal opportunity across cohorts, test it on a held-out population that includes thin-file applicants, document the model’s limitations, keep a human in the loop for borderline cases, and publish enough methodology for external challenge. Guardrails are often described as friction. In lending they are the difference between a system that widens access and a system that quietly locks in the past.
A different route is being tested closer to home. Evangeline Atkinson and Laura Mills founded Noggin HQ after Atkinson, then a student paying rent and bills like clockwork, was refused a thirty-pound-a-month phone contract because no record of her existed at any of the major agencies. Their firm, authorised by the FCA in June, takes the opposite technical path to the American disruptors: it analyses current account transactional data to support lending decisions, but it deliberately avoids generative or opaque AI, arguing that such models carry built-in bias and could easily infer protected characteristics from spending patterns. Instead the platform translates thousands of granular data points into interpretable indicators โ income regularity, essential versus discretionary spending, use of high-cost credit, buffer against shocks โ and hands lenders a legible picture of cashflow resilience. This is the same territory the regulator has been pushing the market toward, and it carries a meaningful advantage: a rules-based cashflow assessment can be explained to the borrower in a sentence. The operational caveat is consent. Open banking access is revocable by design, so any product built on it must handle re-permissioning gracefully, or the model degrades quietly the moment a customer stops sharing.
For anyone who has been on the receiving end of a wrong record, the remedy is knowable but requires persistence. Start by pulling your statutory report from all three agencies rather than relying on one, because inconsistencies between them are common and only one might contain the fraudulent account. Flag every unfamiliar entry in writing, request a notice of correction to sit alongside disputed data, and raise a formal complaint with the creditor as well as the bureau, since the creditor is usually the party that reported the account in the first place. Register with CIFAS for Protective Registration, which makes future lenders apply extra checks, and set up alerts for new searches and account openings. If a bureau fails to correct demonstrably false data, escalate to the Financial Ombudsman Service; dispute investigations are supposed to conclude within about a month, and there is no charge for any of this. Keep everything: correspondence, reference numbers, an Action Fraud report if criminal activity is suspected. The most overlooked step is timing. Audit your file every four to six months, not when a mortgage offer is on the table, because a three-year-old dormant debt is far easier to kill before a lender has already priced you out of the market.
Identity fraud in the credit system is a supply chain problem, which is why it lingers so stubbornly. Catalogue credit is the softest point of entry because retailers are incentivised to approve at the point of sale with minimal verification, and once an account is opened fraudulently the liability question cascades downstream. The debt can be sold to a secondary finance company, then to a collections operation, and each new owner re-reports the account, refreshing the damage on the file. That is exactly the pattern Hope describes: the balance was not a single event but a slow accumulation in the background, invisible until a lender looked. Knowing this changes the advice. Disputing with one agency is insufficient; you need to tell the original creditor that the account was never yours, ask for it to be recorded as fraudulent rather than merely settled or partially satisfied, and pursue removal of the linked collection activity as a separate matter. Settlement is a trap worth naming explicitly, because paying a debt you do not owe can make the record look like an admission and complicate later removal. Be firm, be documented, and be unembarrassed โ the burden of proof sits with the organisation that profited from the mistake.
Explanation is becoming a legal expectation, not a courtesy. UK data protection law gives people meaningful protections around decisions made solely by automated means that have significant effects, and the regulator’s guidance on explaining such decisions has pushed firms toward disclosing the logic involved and the relative weight of the data used. Layer on top of that the Consumer Duty regime, which requires firms to deliver fair outcomes and to ensure customers understand what they are getting, and the practical bar for an AI-driven lender rises sharply. This is where many alternative underwriting projects quietly stall, because the models that perform best are often the least self-describing. The workaround is well understood by mature teams: generate local attribution for every adverse decision, map those attributions onto human-meaningful categories rather than raw feature names, record the model version and the scoring date, and retain immutable decision logs for as long as a complaint or ombudsman case can run. Explainability is also a competitive asset, because a borrower who understands the reason can change the behaviour that caused it. A system that only delivers verdicts produces resignation; a system that delivers reasons produces better customers.
Looking forward, the direction of travel is fairly clear even if the timing is not. Open banking has already normalised the sharing of transaction data, and the same consent architecture is being extended to savings, investments and pensions, while rental payment reporting and tax-authority-verified income are gradually thickening the files of people who were previously ghosts. The likely end state is hybrid underwriting: the bureau score retained as a stable baseline feature, cashflow and alternative signals layered on top to lift genuinely creditworthy thin-file applicants, and human review reserved for the controversial middle. This is a much better description of where the market is heading than any story about an algorithm replacing underwriters outright. The generational dimension matters too, as Olufunwa notes: younger consumers raised inside data-sharing platforms tend to be relaxed about their transactions being analysed, while older cohorts may find the same proposition intrusive, so trust will fracture along demographic lines rather than converge. Lenders should pilot challenger models in shadow mode against existing scorecards for six to twelve months, comparing approval rates and defaults by cohort before switching anything in production. That discipline costs patience and buys credibility.
What should you actually do with all this? If you are a consumer, run a three-agency audit twice a year, correct every inaccuracy in writing, register for protective monitoring if you have been targeted, and build a visible file deliberately if you have none. Never pay a debt you do not recognise without a written response first, and remember that a lender’s refusal is a decision you can question rather than a verdict you must accept. If you are a lender or product lead, define your fairness metric before you train, measure outcomes across thin-file and protected cohorts, keep a human override, log every decision with its model version, and design a complaints path that produces a real explanation rather than a script. If you are building in this space, treat consent, explainability and authorisation as product features rather than compliance overhead, because the firms that win will be the ones whose decisions survive scrutiny. The deeper lesson from Hope’s four lost years is that automation does not make lending fairer by itself โ accountability does. Data can widen access or entrench exclusion depending entirely on who is willing to answer for the output, and that question is still the one the industry has not properly settled.