# Reading a vendor's AI claims like an analyst
A vendor slide claims "99.2% fraud detection accuracy." A bank's risk team approves a six-figure pilot. Eight months later, the model flags 40% of legitimate transactions as fraud in production. The 99.2% was real, measured on a test set the vendor built, on data that looked nothing like the bank's actual transaction mix. This happens constantly in fintech procurement, and it's avoidable if you know what to ask.
This lesson gives you a framework for interrogating AI vendor claims: benchmarks, demos, and accuracy stats, before you sign anything.
Fintech sits at the intersection of two conditions that make AI marketing risky: high regulatory stakes and thin technical literacy among buyers.
A lending platform claiming an AI underwriting model "reduces defaults by 30%" is making a claim about a live population of borrowers, under specific economic conditions, using specific data. Change any of those, and the number can collapse. Unlike a productivity tool, a bad fraud or credit model creates direct financial loss and potential regulatory exposure (in the US, under the Equal Credit Opportunity Act, ECOA, and Regulation B; in the EU, under the AI Act, which classifies credit scoring as "high-risk").
Buyers often don't have in-house data scientists to challenge the numbers. That asymmetry is exactly what a vendor pitch deck is designed to exploit, usually not maliciously, but through selective framing.
Accuracy alone is almost meaningless in fraud and credit use cases because the underlying classes are imbalanced. If only 0.5% of transactions are fraudulent, a model that predicts "not fraud" every single time is 99.5% accurate and catches zero fraud.
What to ask instead:
A vendor who leads with "accuracy" and won't volunteer precision/recall on request is a flag.
Simple worked example: Suppose a bank processes 100,000 transactions, 500 (0.5%) are fraud. A model flags 1,000 transactions as fraud, and 400 of those are true fraud.
That gap between a 99%+ accuracy headline and a 40% precision reality is the single most common fintech AI vendor trick.
Live demos often run on curated data: clean, balanced, sometimes synthetic, and always sized to make the model look decisive. Ask directly: "Was this demo run on synthetic data, historical data, or a live production feed?"
For document-processing or KYC (Know Your Customer, the identity verification process required under anti-money-laundering rules) tools, ask to see failure cases, not just successes. A vendor confident in their product will show you where the model struggles (blurry scans, non-Latin scripts, unusual document formats).
"Outperforms GPT-4 on financial sentiment analysis" sounds impressive until you ask: which dataset, what's the sample size, is it public, and is it the vendor's own benchmark? Public leaderboards like Papers With Code or open benchmark sets (e.g., FiQA for financial sentiment) let you check if a claim is independently reproducible. A benchmark built and scored solely by the vendor is not evidence, it's marketing collateral.
Ask how often the model is retrained, and on what data. A credit model trained on 2019 to 2022 data may not generalize well to a different rate environment or a new customer segment. This is called data drift or model drift, when the statistical properties of live data diverge from training data over time. Any vendor unable to describe their monitoring and retraining cadence hasn't operationalized the model responsibly.
When evaluating a fintech AI vendor, request:
1. Confusion matrix or precision/recall/F1 score on a holdout set, not just accuracy
2. Dataset description: size, source, time period, whether it resembles your own customer base
3. Independent validation: has the tool been reviewed by a third party, regulator sandbox (like the UK's FCA Regulatory Sandbox), or academic benchmark?
4. Explainability method: for credit and underwriting, can the vendor produce reason codes for adverse action notices (required under ECOA in the US)? Tools like SHAP (SHapley Additive exPlanations) values are a common technical approach.
5. Drift monitoring plan: retraining frequency, performance alerts, rollback procedure
6. Reference customer with comparable scale, not just a logo on a slide
Here's a simplified snippet showing what a precision/recall check looks like in practice, the kind of output you should ask a vendor to produce on *your* data during a pilot, not theirs:
from sklearn.metrics import precision_score, recall_score, confusion_matrix
# y_true: actual labels (1 = fraud), y_pred: model predictions
precision = precision_score(y_true, y_pred)
recall = recall_score(y_true, y_pred)
print(f"Precision: {precision:.2f}, Recall: {recall:.2f}")
print(confusion_matrix(y_true, y_pred))If a vendor resists running this kind of evaluation on a sample of your real, anonymized data before contract signing, that resistance is itself information.
Knowledge check
1. A vendor claims their fraud model is '99.5% accurate.' Why can this specific claim be misleading in a fraud detection context?
2. Why did the bank's fraud model perform far worse in production than the vendor's claimed 99.2% accuracy suggested?
3. A lending platform claims its AI model 'reduces defaults by 30%.' What is the most important follow-up question for evaluating this claim?
4. Select ALL correct answers about why fintech AI vendor claims carry unusually high risk compared to claims for a typical productivity tool.
Select all the correct answers.
5. Select ALL correct answers about precision and recall as alternatives to a single accuracy metric in fraud detection.
Select all the correct answers.
Genuine, well-documented fintech AI applications tend to cluster in a few areas: transaction fraud scoring, document extraction/KYC automation, credit risk scoring augmentation (not full replacement), and customer service triage via chatbots or copilots.
Realistic, publicly discussed ranges (treat as estimates, figures vary widely by institution and are not standardized across vendors):
Treat any vendor ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → number the same way you treat their accuracy claim: ask for the baseline it's measured against, the time period, and whether it was measured on your peer institutions or a hand-picked case study.
🎬 [VIDEO: "How Banks Detect Fraud Using Machine Learning" - https://www.youtube.com/results?search_query=how+banks+detect+fraud+machine+learning - Search for recent explainer content from a reputable fintech or data science channel covering real-world fraud model architecture and evaluation metrics.]