AI flags and explains - a human decides and owns it
Read this frame before you touch a prompt, because it governs everything tonight. AI can rank a batch of transactions by how unusual they look, tell you which ones smell like duplicate payments or structuring, and score a credit applicant against a pattern. What it cannot do - not here, not ever in this workflow - is decide. A flag is a candidate for a human to investigate. A credit score is an input, not a verdict. Someone licensed and accountable makes the call and owns it. And credit scoring, specifically, is classified high-risk by the EU AI Act, which means explainability, human oversight, no prohibited bias, and adverse-action reasons are not optional extras - they are the law.
Where AI helps - and the bright red line 6 min live
Three places AI genuinely earns its keep in risk work: spotting anomalies, surfacing fraud and money-laundering patterns, and assisting credit scoring. All three are real. But the third one sits behind a legal line, and the whole session bends around it: AI surfaces and explains, a human investigates and decides.
LiveThe three real uses - and why credit is different3 min▶
- Anomaly detection. Point AI at a batch of transactions, expense claims, or journal entries and ask it to rank the ones that do not fit the pattern - round-number amounts, over-threshold approvals, duplicates, off-hours postings. It is a fast first pass over a pile no one could eyeball.
- Fraud & AML pattern surfacing. Structuring just under a reporting limit, a new payee that suddenly receives a large sum, split invoices to dodge an approval tier - AI can describe these patterns in plain language so a human knows where to look first.
- Credit-scoring assist. AI can summarise an applicant's file and highlight what drives risk. But here the line is hard and legal: it assists, it never decides.
Anomaly triage workflow 5 min live
Triage is the everyday shape of this work. AI does not close cases; it builds you a shortlist. It surfaces the candidates that look off, explains why each one was flagged, and then a human adjudicates - accept as clean, reject as a real issue, or escalate. The explanation is what makes the shortlist usable: a flag you cannot explain is a flag you cannot act on.
LiveThe patterns AI flags - and what a human then checks3 min▶
| Pattern AI flags | Plain reason | What the human checks |
|---|---|---|
| Duplicate | Same vendor, amount, and near-date as another entry | Is it a true double payment, or a legitimate instalment? |
| Round-number | Suspiciously clean amount (5,000.00) unlike normal spend | Real invoice, or an estimate or manual override? |
| Over-threshold | Just above or just below an approval or reporting limit | Split to dodge sign-off, or a genuine one-off? |
| Off-pattern | New payee, unusual timing, amount far from this account's norm | Known counterparty? Does the business context explain it? |
The workflow is always the same three moves: AI surfaces the candidates and ranks them, AI explains each flag in one plain sentence, and a human adjudicates. Nothing is closed or acted on until a person has looked.
Self-studyRail one, for risk data2 min read▶
Risk and fraud data is some of the most sensitive you will ever handle, so the privacy rail is at its strictest here.
- Synthetic or public data only in consumer tools like a personal ChatGPT or Claude account. Every dataset in tonight's demos is fabricated for practice.
- Never paste real customer transactions, account numbers, PII, or material non-public information into a consumer tool. Real casework belongs in an enterprise or local system that does not train on your data and sits inside your controls.
- Credit files carry legal protections on top of confidentiality. Treat an applicant's data as regulated from the first keystroke.
Credit & fraud governance 5 min live
If AI touches a decision about someone's money, five things have to be true, or the decision is not defensible. This is not paperwork - it is what lets you explain the outcome to the person, the auditor, and the regulator. AI ranks and flags; a person owns the decision and can account for it.
LiveFive duties on any credit or fraud decision3 min▶
- Explainability. You must be able to say why something was flagged or scored, in words the affected person could understand. A black box that no one can explain cannot drive a high-risk decision.
- Fairness. The system must not use prohibited attributes or their proxies, and it must not produce biased outcomes. Test for it; do not assume it.
- Adverse-action reason codes. When someone is declined, they are owed clear, specific reasons - not "the model said no". These reasons are a legal requirement, and drafting them well is a real skill (Demo 2).
- Model risk (SR 11-7). Any model driving these decisions is validated before use, monitored in production, and governed with an owner. A prompt-based workflow is still a process that needs controls.
- Audit trail. Record what was flagged, the reason, who reviewed it, and what they decided. The trail is how you prove a human was in the loop when it is questioned later.
"The model said no" is not a reason. A lender let an automated score drive declines without human-readable reasons attached. When a declined applicant challenged it, the team could not explain the specific factors behind the outcome - and under adverse-action rules, that is the finding, not a detail. The fix was not a better model; it was reintroducing the human gate and generating clear reason codes for every decline. AI ranked the risk. A person had to own - and explain - the no.
Flag anomalies in a synthetic batch - then adjudicate ★ 13 min · everyone builds
Hand AI a small, fabricated set of transactions, ask it to flag the anomalies and explain each one, then do the part that is actually yours: adjudicate. Accept the clean ones, reject the real issues, escalate the unclear. Human in the loop, from the first flag.
Use the synthetic data below - copy it straight into the prompt. It is fabricated for practice, so a consumer ChatGPT or Claude account is fine (rail one: synthetic only).
Ask AI to rank the rows by how unusual they are and, for each flag, give one plain reason and the exact fields that triggered it. Tell it to describe, not accuse.
Read every flag critically. For each one, decide: accept (looks clean), reject (real issue to pursue), or escalate (need more context). Write your call next to each - that is the adjudication.
Look for a flag you disagree with. Where did the tool over-call or miss? That instinct is exactly why a human sits at the gate - note it.
Draft plain-language reason codes for a flagged item ★ 8 min · build your own
When a flag becomes a decision that affects someone - a declined application, a held payment - the reason has to be written in words the person and an auditor can understand. Practise drafting explainable reason codes for one synthetic case. This is the adverse-action and audit-trail skill from Part 3, made concrete.
Take one flagged synthetic case - use the profile below, or a fabricated one of your own. Never a real applicant.
Ask AI to draft the reasons in plain language: specific factors, no jargon, no protected attributes, phrased so the affected person could understand them.
Review each reason for fairness and clarity. Strip anything vague ("did not meet criteria") and anything that looks like a proxy for a prohibited attribute. Keep it factual and specific.
Remember the frame: these are a draft for a licensed human to review, approve, and own. The final decision and its reasons belong to a person, not the model.
From "declined" to a reason you can defend. A credit team used a drafting step exactly like this to turn terse internal flags into clear, specific adverse-action reasons - each tied to a factual factor, none touching a protected attribute. The AI produced a fast first draft; a human reviewed every one, corrected two, and signed off. When a decline was later challenged, the reasons held up because a person had written and owned them. The model assisted the wording; the accountability stayed human.
Try it yourself - this week ◐ 30-45 min total
- Finish both demos if you did not complete them live. The adjudication step in Demo 1 is the one that matters - do not let the tool "decide".
- Build your own synthetic batch (10-15 fabricated rows) for a data type you actually handle - expense claims, journal entries, or payments. Plant two or three anomalies on purpose, then see if AI catches them and explains them well.
- Draft reason codes for a second synthetic flagged case. Test them against one question: could the affected person understand and check every reason? Rewrite any that fail.
- Write, in one paragraph, where the human gate sits in a risk workflow you own today. Name who investigates, who decides, and where the audit trail lives. If any of those is "the tool", that is the gap to fix.
- Optional: read the EU AI Act's high-risk provisions on creditworthiness, and skim SR 11-7's model-risk lifecycle, so you can name the duties, not just sense them.
Official sources covered
This session teaches AI-assisted risk, fraud, and credit work as a hands-on practitioner, anchored to the credit-scoring and default-risk material in the mapped programs and to the regulatory canon. Building or fine-tuning risk models from scratch stays with the dedicated ML courses - said honestly. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · How does the EU AI Act treat credit scoring and creditworthiness assessment?
The EU AI Act classifies credit scoring as high-risk. AI may assist, but the system must be explainable, kept under meaningful human oversight, free of prohibited bias, and able to produce adverse-action reasons.
2 · Why is a human gate mandatory on a credit or fraud decision, not just nice to have?
These are regulated decisions about people's money. AI surfaces and explains; a licensed, accountable human must investigate, decide, and own the outcome - including being able to explain it to the person and an auditor.
3 · What are explainability and adverse-action reason codes for?
Explainability and reason codes give the specific, plain-language factors behind a flag or a decline. They make the outcome checkable and defensible - and they are a legal requirement, never a substitute for the human gate.