# CauseCraft: selected incident evidence Checked 2026-11-05. Five selected reported incidents or business closures. Retired source records remain in the authored collection. This is a purposive teaching collection, not a sample from which to estimate an AI failure rate. Generative AI, predictive models and automated decision systems are labelled separately. These notes distinguish reports from assigned explanations. The cause deck contains hypotheses; none is an official verdict. ## F01 · CBA: the calls kept coming August 2025 · Voice bot and staffing workflow · Union account; causal evidence incomplete The Finance Sector Union said CBA reversed plans to remove 45 customer-service roles. It disputed claimed voice-bot savings, reporting rising calls and overtime. The union said management had underestimated how long higher demand would continue. ### Reported organisational mechanism The union's explanation concerns demand forecasting and staffing decisions, not simply the bot's answer quality. It said call volumes contradicted the redundancy rationale. Keep that attribution: this packet is not an independent causal investigation. ### What remains unknown We lack the complete call dataset, bot resolution rate, repeat contacts, work mix, approval criteria and the bank's full account. Governance and HR explanations remain competing hypotheses. ### Observation to seek Did bot adoption reduce total human work, after repeat calls and complex cases? ### Sources - [Finance Sector Union, 21 August 2025](https://www.fsunion.org.au/win-cba-backflips-on-customer-service-job-cuts-admits-they-got-it-wrong/) · Interested party; primary statement ## F02 · Zillow: buying the forecast November 2021 · Predictive pricing and property operations · Company wind-down explanation Zillow wound down its home-buying business. Its shareholder letter said home-price forecasts had missed in both directions by more than it had modelled as possible, making further expansion too volatile for earnings and operations. ### Reported organisational mechanism Management connected forecasting error, inventory exposure, operating constraints and return on equity. The business bought assets using forecasts; errors therefore accumulated on a balance sheet. This was a business-system failure, not proof that every Zestimate was inaccurate. ### What remains unknown The letter does not isolate model design from acquisition targets, overrides or execution. The decision to stop can itself be sensible risk management. ### Observation to seek How much inventory could the business survive if forecast errors clustered? ### Sources - [Zillow Q3 2021 shareholder letter, 2 November 2021](https://www.sec.gov/Archives/edgar/data/1617640/000161764021000085/exhibit993.htm) · Company filing; primary ## F03 · Replit: production was reachable July 2025 · Coding agent · User report and company acknowledgement A user reported that Replit's agent deleted a production database despite a code freeze. Replit's CEO called the incident unacceptable. The company announced separation of development and production databases to prevent the agent changing production data. ### Reported organisational mechanism The response addressed access boundaries, not just more emphatic prompting. A written freeze did not function as an enforced permission boundary in this incident. ### What remains unknown The user's transcript is not a complete forensic record. Do not infer permanent irrecoverable loss, a model's human motives or the prevalence of this failure across projects. ### Observation to seek Could the agent execute the destructive action even if it ignored the instruction? ### Sources - [SFGATE reporting and Replit spokesperson, 22 July 2025](https://www.sfgate.com/tech/article/bay-area-tech-product-rogue-ceo-apology-20780833.php) · Reported incident and company response ## F04 · Deloitte: assurance needed checking 2025 · Generative AI in report preparation · Official acknowledgement and published correspondence An Australian government assurance report prepared by Deloitte required corrected footnotes and references. The department acknowledged the errors. Published correspondence described generative AI used for publishing-related tasks, including summarising material and completing citations. ### Reported organisational mechanism Official correspondence and the department's statement support a workflow and verification failure. The report's subject was itself assurance of an IT system. Wrong references do not by themselves prove that every substantive recommendation was wrong. ### What remains unknown The sources here do not establish who checked each reference, the precise model contribution to each error or a unique organisational root cause. Do not conflate the reviewed welfare system with the report-writing tool. ### Observation to seek What evidence showed the cited sources existed and supported the report? ### Sources - [DEWR Secretary statement, 3 October 2025](https://www.dewr.gov.au/assuring-integrity-targeted-compliance-framework/announcements/statement-secretary-progress-under-targeted-compliance-framework-integrity-assurance-program) · Government acknowledgement; primary - [Published Deloitte–DEWR correspondence](https://www.aph.gov.au/-/media/Estimates/eet/supp2526/TCF_Assurance_Review_-_correspondence_between_Deloitte_and_DEWR.pdf) · Parliament-published correspondence; primary ## F05 · Earnest: the school became a proxy July 2025 · Algorithmic student-loan underwriting · Attorney General allegations resolved by settlement Massachusetts settled allegations against student-loan lender Earnest for US$2.5 million. The regulator alleged that school default rates in its lending model disadvantaged Black and Hispanic applicants for one product, and that the models lacked tests for unfair impact. ### Reported organisational mechanism The Attorney General alleged that models were trained on arbitrary human decisions and were not tested for disparate impact. The settlement required changes to governance and lending practices, including removing the school-default-rate variable. These are allegations resolved by settlement, not a trial finding. ### What remains unknown The announcement does not provide a complete model audit or establish a single cause for every lending decision. The settlement also addresses practices beyond the AI model; its total payment is not a measured model-only loss. ### Observation to seek Which groups received different decisions or terms, and what evidence showed that the chosen variables were appropriate? ### Sources - [Massachusetts Attorney General announcement, 10 July 2025](https://www.mass.gov/news/ag-campbell-announces-25-million-settlement-with-student-loan-lender-for-unlawful-practices-through-ai-use-other-consumer-protection-violations) · Regulator announcement; primary, search-index text verified ## Selection limits The selected cases do not establish a population failure rate. Retired cases retain their source records outside the active deck. ## Commercial-loss hypothesis C06 permits a rational risk-acceptance argument. A company would need total savings, error losses, customer retention, external costs, obligations and alternatives to defend it. C25 asks who bears the cost.