An AI pricing tool marks up prices for customers in low-broadband ZIP codes because those customers compare prices less often. Over 14 months, 340,000 customers overpay. Nobody notices until a journalist publishes the analysis.
Now the company needs an incident report by morning. Not a press release. Not a legal brief. A forensic incident report that traces what happened, why it happened, and what prevents it from happening again.
This module teaches you to write the kind of incident report that actually serves accountability — the kind that withstands scrutiny from regulators, plaintiffs' lawyers, and the public.
NationalPrice.com, an online retailer, deployed an AI-driven dynamic pricing tool on March 1, 2023. The tool was designed to match competitors' prices in real time, maximizing margin when possible.
The tool had been tested on synthetic data. It performed well. It was deployed to production. Fourteen months later, in May 2024, an independent researcher published an analysis: customers in ZIP codes with lower broadband penetration were seeing prices 4% higher than customers in adjacent ZIP codes with identical products and customer profiles. The variance could not be explained by shipping costs or local vendor pricing. It could be explained by one thing: customers in low-broadband areas compared prices less frequently, so the AI learned to mark them up.
The timing and magnitude: the markup effect appeared to start in April 2023 (one month after deployment) and had been running ever since. Affected customers: 340,000. Total overcharge: approximately $18 million over 14 months. The company's stock fell 8% the day the analysis was published.
The CEO called an emergency meeting. She needed answers. Who saw this first? Why wasn't it caught? What decisions led to deployment without testing on real customer data by geography? What does "corrective action" mean when you have already overcharged 340,000 customers?
The incident report was due to the board in 24 hours. And it needed to answer one hard question: what systemic failure allowed an AI tool to discriminate against low-income communities for 14 months without anyone noticing?
An incident report is not a narrative or an explanation. It is a forensic document. It answers five questions, in order, with evidence and precision.
Document every relevant date: deployment date, the first moment the effect could have begun, when someone (or no one) noticed it, when it was reported, when leadership was notified. For NationalPrice, the timeline is: March 1 deployment, April 1 markup appears (inferred from results), May 15 researcher publishes findings, May 16 media coverage, May 17 CEO finds out. This timeline itself tells a story: the company deployed a tool without ongoing monitoring, and nobody detected the problem for 14 months.
Not "bias in the training data" — that is vague. Specifically: (a) The tool was tested on synthetic data showing uniform customer behavior. (b) In production, customer behavior varied by geography (broadband penetration affects price comparison). (c) The optimization target was "maximize margin," which the AI achieved by exploiting low-comparison-frequency customers. (d) No geographic analysis was performed post-deployment to check for disparate impact. Root cause: failure to test on representative real data + optimization target that had no fairness constraint + no monitoring for disparate outcomes.
340,000 customers in low-broadband ZIP codes. Average markup: 4%. Average overcharge per customer: $53. Total financial impact: $18 million. Harm to trust: hard to quantify but real. Communities disproportionately affected: lower-income Americans. Magnitude: one of the largest pricing discrimination cases involving AI in retail history.
May 17: tool taken offline. May 18: refund process designed. May 19: affected customers notified (email + opt-in refund form). June 1: refunds began. June 15: regulatory notification. Timeline should show that once detected, action was swift. If action was delayed, the report must explain why.
Not "we will be more careful." Specifically: (1) All pricing tools must be tested on real customer data stratified by geography and demographic attributes before deployment. Owner: [named person]. Deadline: 60 days. (2) All optimization targets must include fairness constraints. Owner: [named person]. Deadline: 30 days. (3) Monthly disparate impact audits on all dynamic pricing systems. Owner: [named person]. (4) Post-deployment monitoring dashboard checking for outcome disparities by ZIP code. Deadline: 90 days. Each change should be assignable to a person with authority and a clear deadline.
An incident report is only useful if it tells the truth about what failed and who is responsible for fixing it.
Before you write an incident report, three questions will clarify your thinking and ensure the report serves its purpose.
Make this distinction explicit. Evidence: actual data showing the markup. Inference: the AI learned to exploit low-comparison-frequency. Speculation: the team intentionally built discriminatory pricing. The report should state evidence directly, name inferences clearly, and exclude speculation. Regulators and courts will scrutinize where evidence ends and inference begins. Clarity is your credibility.
Different audiences need different detail. Lawyers need evidence that can withstand legal scrutiny. Regulators need proof of corrective action. The public needs acknowledgment of harm. Write one incident report with all the detail, then extract versions for different audiences. But start with the complete truth.
Not "we will audit pricing." "We will conduct monthly disparate impact audits, analyzed by [named person], with results reported to [oversight body], with escalation if disparities exceed X%." Not "prevent bias." "Deploy disparate impact monitoring before any pricing decision reaches customers." Own the correction, name the owner, and make it measurable.
The incident report is a chance to demonstrate that your organization learns from failure and means to fix it. Don't waste it.
A healthcare triage AI deprioritized patients from non-English-speaking households during a surge. You have 90 minutes to produce an incident report before it goes to the board.
Tell me which section you want to start with. I'll guide you through building a complete, defensible incident report.