01 — The premise

Two questions lenders keep conflating

Lending asks two different questions, and most stacks answer them with one blunt instrument.

Can we accept this customer? — a question about risk: their track record, their conduct, the negative events on file. And: what can they sustainably repay, and on what terms? — a question about liquidity: the actual money moving through the account each month once everything else is paid.

These are not the same question. They don’t move together, and they are not best answered by the same data. Risk is historical and relative. Liquidity is present-tense and absolute. Collapse them into one score and you accept people you shouldn’t, decline people you should be serving, and set repayment terms that break the moment life wobbles.

Credit Canary decisions them separately: risk for acceptance, liquidity for repayment and terms. The rest of this paper is the evidence for why, and the architecture that delivers it.

02 — The evidence

The bureau lottery

If you decision risk off one bureau — and 91% of applications use exactly one1 — you’ve taken a single, noisy read of a moving target.

The FCA’s Credit Information Market Study put this beyond argument. It matched the same ~48,000 people across all three large CRAs, ranked each person within each bureau, and asked a simple question: does the same customer land in the same place? Mostly, they don’t. Only 16–29% sit in the same risk decile across two bureaus, and 16–36% are three or more deciles apart.1

Figure 1 · Same customer, two bureaus
5 2 1 2 2 2 1 1 2 2 2 1 1 2 2 2 1 1 2 2 2 1 1 2 2 2 1 1 2 2 2 1 1 2 2 2 1 1 2 2 2 1 2 3 12345678910 12345678910 Risk decile · Bureau A Risk decile · Bureau B 1 = highest risk 10 = lowest risk
Read the diagonal. If the bureaus agreed, every customer would sit on the dark diagonal. They don’t — the mass spreads off it: the same person, a “good” risk at one agency and a marginal one at the next. Agreement is strongest only at the bottom-left, where everyone concurs on who is unambiguously high-risk. In the middle deciles — where most decisions are actually made — it’s a lottery. Heatmap is illustrative of the published pattern; the 16–29% / 16–36% figures are the FCA’s reported values.

This is not a reason to abandon the bureau. Negative events, conduct history and relative ranking are exactly what a CRA is for, and we use it for precisely that: acceptance. It’s a reason never to mistake one bureau’s read for the whole truth — and never to read affordability off it.

03 — The weakest field

Income is the field that lies — and forms make it worse

If risk is noisy across bureaus, income is worse. The same study found more than 54% of people had an income that differed by two or more deciles depending on the agency1 — and income is the input your affordability assessment leans on hardest.

Figure 2 · Bureau income disagreement
54% — differ by 2+ deciles
46% — broadly aligned
Share of matched customers whose income estimate lands two or more deciles apart between two bureaus. Affordability built on this alone inherits the disagreement.1

Now layer on the form. The industry likes to say customers don’t fill in forms honestly. That’s the wrong diagnosis. Customers aren’t lying — they don’t know. An application form assumes a level of recall and financial self-knowledge almost nobody has: monthly outgoings to the pound, every subscription, the real cost of childcare this month, exactly what leaves the account on payday. People answer with a hopeful estimate, because that’s the best the format allows.

A form doesn’t catch liars. It catches everyone — because it asks people to recall a financial life they’ve never had to hold in their head.

So we stop asking the question the form can’t answer. We verify instead. Open banking transactions show real income hitting the account. Direct debit data shows the committed outgoings a customer would never list from memory — the credit agreements, utilities, BNPL, the subscriptions. Then we do the thing forms never do: we play it back. “We can see around £X going to these commitments each month — does that look right, and is anything missing?”

Validation beats interrogation. The customer corrects and confirms a picture we’ve already built, instead of constructing it cold from memory. That’s the split in one line: CRA data answers risk; open banking and direct debit reality answer affordability, repayment amount and terms.

04 — Why completeness is a control

One missing field can decide the whole case

There’s a second reason the data has to be complete, not just accurate: a single fact can flip the entire decision. Lenders told the FCA they would automatically decline on one data point alone.

Figure 3 · Share of lenders that auto-decline on a single factor
Currently in an IVA94%
IVA in the last 6 years70%
Bankruptcy in the last 6 years67%
Unsatisfied CCJ in the last year45%
Satisfied CCJ in the last year39%
Default in the last year27%
Unsatisfied CCJ in the last 6 years15%
If one of these facts is missing or divergent at the bureau you happened to pull, the decision is wrong — harshly or generously, but wrong.1

When one variable can flip the outcome, a partial picture isn’t a smaller version of the right answer. It’s a different answer. And the bureau you happen to pull is, as Figure 1 showed, partly a matter of chance.

Figure 4 · How many bureaus inform a decision
91% one bureau
One bureau — 91%
Two bureaus — 9%
All three — ~0%
Waterfalling to a second bureau is far rarer than the industry assumes — even on thin files, the vast majority of applications never look past the first agency.1
05 — The single view

One corroborated view of the customer

No single source is trusted on its own. We overlap them — and confidence lives in the overlap.

Multiple bureaus, open banking transactions, direct debit payments, income-statement uploads and the customer’s own validation all feed the same picture. Where two or more independent sources agree, we trust it. Where they diverge, that’s the signal — and the agents go to work. The output is one clean view of the customer that no single bureau, no form and no lone open-banking feed produces by itself.

Figure 5 · Sources converge into a single validated view
CORROBORATED Single customer view Multiplebureaus Openbanking Directdebits Incomeuploads Customervalidation
Each source is partial and, on its own, fallible. Their intersection is where a decision-grade picture forms — the more sources corroborate a fact, the higher our confidence in it.

From that corroborated base we resolve the four things that actually decide whether a loan is affordable:

Validated income — not what was declared, but income we can see arriving and the customer confirms. Credit exposure — the real repayment commitments, from bureau lines and the direct debits servicing them. Essential spending — utilities, rent, council tax: the non-negotiables. And discretionary spending — what’s left, read from transactions and balanced against ONS household-spend benchmarks, so we can judge whether the picture is plausible and how much genuine headroom exists.

Figure 6 · What the single view resolves — and what corroborates it
Bureaus
Open
banking
Direct
debits
Income
uploads
ONS
benchmark
Customer
validation
Validated income
Seen arriving & confirmed — not declared
Credit exposure
Live repayment commitments
Essential spending
Utilities, rent, council tax
Discretionary spending
Headroom, sense-checked
Primary source Corroborates ONS benchmark
Every dimension is proven by more than one independent source. Discretionary and essential spend are benchmarked against ONS household-spend data, so we can tell whether a customer’s outgoings are plausible for their household — and where the real headroom sits.

That single view is what the rest of the engine runs on — and the next question is simply how complete it is.

06 — The mechanism

The Data Quality Score

This is why every Credit Canary decision begins with a question about the data itself, before any question about the customer: how complete and corroborated is the picture we’re standing on?

The Data Quality Score runs across the variables that determine whether a decision is safe to make — among them identity and bureau coverage (thin-file and no-hit flags), bureau depth, whether declared income is corroborated by open banking, transaction-history depth, whether direct-debit commitments reconcile with what the customer told us, and the confidence of the resulting affordability read.

A high score means we have enough to decide well. A low score doesn’t trigger a decline — it triggers more questioning. Our AI agents take the gaps the score exposes and act on them: ask the customer a targeted follow-up, request an additional connection, reconcile a discrepancy, and play the findings back to confirm. The score drives the conversation; the agents run it; the customer ends up validating a complete picture rather than seeding a thin one.

Figure 7 · Two-axis decisioning architecture
Bureau data
conduct & negative events
Open banking
real income & spending
Direct debits
committed outgoings
Income verification
corroborated, not declared
Identity & coverage
thin-file / no-hit flags
Data Quality Score
scored across every input — completeness, corroboration, confidence
Low score → AI agents ask targeted questions, request data & play findings back to validate ↻
Risk model · explainable ML
Acceptance
Accept / decline with reason codes, grounded in bureau conduct & history.
Liquidity model
Repayment & terms
Affordable amount, term & headroom from verified cash-flow reality.
Risk and liquidity are decisioned on different inputs and answer different questions — gated by a single measure of whether the data is good enough to act on.
07 — The guarantee

Explainable by construction

Two engines sit on top of that validated picture.

The risk model decides acceptance, grounded in explainable machine learning — glass-box by design, every decision carrying the reason codes behind it. In a Consumer Duty and SM&CR world, a risk model you can’t explain isn’t a model you can deploy: customers are owed a reason, regulators an audit trail, and your own credit committee the logic. Accuracy you can’t account for is a liability, not an edge.

The liquidity model decides repayment amount and terms — built on the verified affordability surface, not the self-reported one. It’s what lets us say not just yes, but yes, this much, on these terms, and here’s the headroom if income dips.

Risk for acceptance. Liquidity for repayment. Both explainable. Neither guessed.

08 — The takeaway

What this means

The data the whole industry relies on is noisier than it admits. Bureau risk reads diverge, income data diverges further, and the form was never going to fix it — because the problem was never honesty, it was recall.

The answer isn’t a better form or a bigger model on worse inputs. It’s to ask the right question of the right data, verify instead of ask, fill the gaps deliberately with a data-quality score and the agents to act on it, play the findings back to the customer, and keep the whole chain explainable end to end.

Two questions. Two data sources. One picture you can stand behind — and explain.