Pitch · April 2026
AI Ethics Challenge

Transparent AI
for Fair Credit
Scoring

Neo Bank's path to inclusive, explainable, and legally compliant AI-driven lending — built on a four-layer framework, sequenced over twelve months.

Presented byBassel Samir · Samir Mohammed · Lucas Quadri · Blesson Prepared forGeneva Business School DeadlineEU AI Act · 2 August 2026
01The business problem
02 / 12
The problem

Neo Bank is leaving creditworthy applicants on the table because its model cannot read them.

Approval rates on applicants of emerging-market origin are materially lower than on applicants with local Spanish footprints — even after controlling for observable creditworthiness.

The model is not wrong on average. It is systematically wrong on one specific cohort, and that cohort is both growing and commercially valuable.

Every rejected-but-creditworthy applicant is a double loss: a missed revenue line today, and a future customer acquired by a competitor that can read alternative data.

A sizable, accelerating segment.
Foreign-born entrepreneurs, freelancers, and self-employed professionals across Spain.
Cohort-level mispricing compounds quickly. The first lender to score it accurately keeps it.
02Where the bias comes from
03 / 12
Spain's credit system → structural blind spots

Spain has no universal score. It has four inputs — and two outcomes.

The bias is not a model flaw in isolation — it is a structural consequence of the inputs Spanish credit assessment is built on.

CIRBE Central bank credit register Banco de España · exposures ≥ €1,000 ASNEF · BADEXCUG Private default registries Binary flags · clean or flagged Documentation Spanish payslip · tax · social security Evidence requires a local footprint Neo Bank internal AI model Proprietary risk-scoring engine Where the bias is produced Approved Spanish-resident salaried profile Complete registries, stable nómina, local footprint Rejected · penalised Emerging-market entrepreneur profile Zero CIRBE · Zero ASNEF · Zero social security — read as risk
The same four inputs produce two different outcomes for two different profiles. The bias isn't intentional — it's structural.
03Cost of inaction
04 / 12
Doing nothing is the most expensive option

Each dimension is modest in isolation. Together, they compound.

The exponential cost is not any single line — it is that all four deteriorate together, and that they reinforce each other on a 36-month horizon.

DimensionYear 1 exposure3-year compounded exposure
RegulatoryConformity-assessment gap on the Aug 2, 2026 deadline.Up to €35M or 7% of global turnover in penalties; AESIA supervisory scrutiny in Spain.
CommercialApproval gap on the thin-file cohort; revenue foregone.First-mover advantage captured by competitors; permanent loss of the cohort.
ReputationalIsolated customer complaints, social-media exposure.Brand association with discriminatory lending; CAC inflation across segments.
Litigation / GDPRSCHUFA-style complaints under GDPR Art. 22.Class-style exposure on automated decisions made without meaningful explanations.
04The framework
05 / 12
The intervention, end to end

Four integrated layers. Every decision explained.

Not a single model — a compliance architecture. Each layer addresses a specific friction from slide 3.

LAYER 1 Data enrichment Solving the "thin-file" problem via PSD2 open banking Cash-flow data Cross-border transactions, consistency, volume Alternative signals SaaS, invoices, rent, utility consistency Behavioural data App usage patterns, financial discipline Remittance history Volume, frequency, stability as income proxy LAYER 2 Explainable AI engine Every score arrives with a rationale — not just a number SHAP values Per-feature contribution to each decision LIME explanations Local interpretability as a human-readable cross-check Challenger model Offline uplift testing — never decides on live customers LAYER 3 Fairness constraints Four metrics, continuous. Rolls back on drift. Demographic parity Equal approval rates across cohorts Equalised odds Equal true + false positive rates per group Disparate impact Four-fifths benchmark — target ratio ≥ 0.85 Proxy detection Continuous scanning for nationality leakage LAYER 4 Human-in-the-loop governance EU AI Act Art. 14 compliance · AESIA oversight Human review Borderline referrals to underwriters with XAI dashboard Audit trail Full decision logging — tamper-evident, per model version Appeal mechanism Adverse-action rights — GDPR Art. 22 compliant
05How it works in practice
06 / 12
Augmented, not replaced

Underwriters are upgraded, not removed.

The strongest test of augmentation: a member of staff who understood their job before the framework should recognise it as a better version of the same job — not as a different job.

Before the framework

A black-box score, or a 30-minute file read

Either an opaque decision the underwriter cannot challenge, or a raw application file that takes 20–30 minutes to interpret from scratch.

Decision quality depends heavily on individual experience. Cases that should be reviewed get auto-rejected.

→
With the framework — XAI dashboard
XAI · Underwriter app·esp-2026-0847 Routed · Review required
M. K.
34 · self-employed designer · Madrid · 22 mo. with bank
Risk score
642
Review required
Top contributors · SHAP
+Cash-flow stability (PSD2, 18 mo.)
+Invoice-payment cadence
+Utility & rent consistency
−NIE recency < 36 mo.
−No CIRBE history
Fairness · cohort vs ref
0.87within target
Audit trail
model v3.2 · 14·04·26 · logged
No customer rejected by a model in isolation. Authority stays human; the inputs get better.
06ROI & valuation
07 / 12
What it costs. Why it earns it back.

A build-heavy Year 1, then steady-state — and four valuation inputs that move simultaneously.

Year 1 — build
≈ €2.7M – 3.7M
Tech build, MLOps + data pipeline, conformity assessment, training and change management. One-off.
Years 2–5 — steady state
≈ €600K – 950K / yr
Maintenance, retraining, monitoring, ongoing governance. Folded into BAU risk and tech budgets.
Where this shows up in the valuation
Revenue growth
DCF · numerator

Approved volume on the thin-file cohort adds incremental revenue. Compounding, not one-off.

Operating margin
DCF · numerator

Lower CAC on the cohort and reduced manual-override load. Operating leverage as it scales.

Cost of capital
DCF · denominator

Demonstrable AI-Act compliance lowers the regulatory risk premium. WACC falls.

Exit / multiple
Multiple expansion

Acquirers pay a premium for compliance-by-design. Underwrites the terminal value.

07The three KPIs
08 / 12
The three numbers the business is judged on

Three KPIs. No others.

The framework stands or falls on these three. Reported monthly to the risk committee, quarterly to the board. No fourth vanity metric gets added to make the picture nicer.

I
Demographic parity ratio
Approval rate on the target cohort divided by approval rate on the reference book, after controlling for observable risk. The ethics KPI that is also a commercial KPI.
12-month target
Ratio ≥ 0.85 over rolling 90-day windows — above the widely used four-fifths disparate-impact benchmark.
II
Customer acquisition cost
CAC on the underserved cohort vs. the general book. The framework should materially reduce CAC on the cohort by unlocking inbound demand and lowering rejection-driven churn.
12-month target
CAC on the cohort converges on the general-book baseline.
III
Loss-adjusted NIM
Net interest margin on the cohort, after impairment. The commercial test. An approval-rate lift that degrades loss-adjusted NIM is not a win — it is subsidised volume.
12-month target
Held flat or better on the cohort vs. pre-framework baseline.
08Ethics, regulation, human factor
09 / 12
One synthesis: compliance, ethics, geography

Compliance is built in. Ethics is resolved in the architecture. Spain is the right jurisdiction.

EU AI Act · high-risk obligations

Every article, answered.

Art. 9Risk management — bias & fairness audits as continuous process.
Art. 10Data governance — PSD2 enrichment, proxy-leakage detection, lineage.
Art. 13Transparency — model fact sheet, plain-language adverse-action statements.
Art. 14Human oversight — Layer 4: routing, dashboard, appeal route, revocable authority.
Art. 15Accuracy & robustness — champion-challenger, drift monitoring, rollback.
Art. 26-27Deployer duties — FRIA completed; AESIA-aligned incident channel.
The human factor · three tensions

Resolved in the design.

Privacyvs. accuracy. Enriched data only under explicit PSD2 consent, with withdrawal path. The applicant controls the data used to score them.
Biasvs. performance. Conscious trade — small accuracy cost for legal, ethical, reputational gain. Cost closes as better data enters.
Workforcevs. automation. Underwriters keep authority. Time shifts from raw file-reading to structured judgment on borderline cases.
Tech-Spain · the geography edge

Built to Spain's standard. Ready for the EU's.

AESIA — the Agencia Española de Supervisión de Inteligencia Artificial — is the EU's first national AI supervisory authority.

Conformity assessment, CE marking, and EU high-risk AI database registration run through the EU process. AESIA supervises deployers in-country, and is where Spanish institutions evidence operational compliance.

09Execution & the principal roadblock
10 / 12
12-month rollout · cultural mitigation

Twelve months, three exits. The hard part is cultural, not technical.

Phase One
Pilot
Months 1–4

Layers 1 & 2 in shadow-mode on a ring-fenced segment. Outputs logged, not used in live decisions. Fairness audit and proxy-leakage scans run side-by-side with the legacy model. Conformity-assessment workstream initiated; technical documentation begun in parallel. Exit: stable fairness metrics; no material proxy leakage.

Phase Two
Scale
Months 5–8

Live on a defined lending segment with legacy fallback. Layer 3 fairness constraints active. Layer 4 governance and appeal route operational. First conformity-assessment draft delivered to external auditor. Risk committee retains authority to revert at any point. Exit: loss-adjusted NIM flat or better vs. baseline.

Phase Three
Sign-off & expansion
Months 9–12

Phased rollout to additional lending segments toward bank-wide coverage. Conformity assessment completed and signed off; CE marking applied; EU high-risk AI database registration filed. Monthly fairness reporting institutionalised. Exit: conformity sign-off and EU registration completed ahead of Aug 2, 2026.

The principal roadblock
Technology. Cultural resistance.

Risk functions trained on traditional scoring stall quietly at exit criteria, not openly. Mitigations are engineered into the phase design, not bolted on after.

Risk-committee sponsorship from day zero.The committee owns the exit criteria, not the technology team.
Continuous comparative evidence.Resistance has to be argued against the bank's own data, not against ideology.
Underwriter co-design, not notification.The dashboard is built with the people closest to rejection — not handed to them.
10Marketing & trust
11 / 12
The trust rebuild — radical transparency

"The Clarity Corridor" — every decision, explained, on the record.

A fairer credit framework is only half the work. The other half is communicating it. The Clarity Corridor turns the framework's transparency from a back-office property into a public-facing brand asset — to the cohort, to the market, and to the regulator.

I
The personal touch

The "Why" Receipt

Every applicant — approved, declined, or escalated — receives a personalised, plain-language explanation of the decision and the data that drove it. The reasoning is no longer opaque, even to the rejected. Doubles as our GDPR Art. 22 and CCD 2023/2225 answer.

II
The cohort, in their own words

Founder Stories

A video series featuring entrepreneurs the legacy system would have rejected and the framework approved — Madrid-grounded, with founders whose origins span emerging markets. Owned distribution and earned media. Builds the proof the regulator and the market both want to see.

III
The receipt for the institution

The Ethics Ledger

A quarterly public report — demographic parity ratio, drift incidents, model versions in production, fairness-audit outcomes — independently audited by a third-party ethics firm. Turns a reporting obligation into a competitive signal.

ClosingThe thesis
12 / 12
The thesis, distilled

The solution is not "be nicer to emerging-market entrepreneurs."

The solution is "stop confusing unfamiliar data with bad risk."

Fairer lending and better credit-risk measurement are the same technical problem — solved by the same four-layer framework. Neo Bank, building it first, gains a structural competitive advantage that compounds through every quarter it runs.
01 / 12
Navigate←→·EscOverview

All slides