Contract commitmentMarch 14, 2026 · Air Canada
AI chatbot invented a bereavement discount policy that didn't exist
Air Canada's AI chatbot told a passenger he could apply for a bereavement discount retroactively—a policy that did not exist. The airline argued the chatbot was a 'separate legal entity' responsible for its own statements. A Canadian tribunal ruled against Air Canada and ordered a refund.
Impact
$812 CAD refund + tribunal costs. Precedent-setting ruling that airlines cannot disclaim AI chatbot statements.
Prevention pattern
POLICY_CLAIM_BLOCK — flags commitments about refund policies, discount eligibility, and retroactive application that aren't in the approved policy document.
DefamationJanuary 22, 2026 · DoNotPay
AI 'robot lawyer' fabricated case citations in court filings
DoNotPay's AI legal assistant generated court filings containing fabricated case citations (cases that do not exist). Opposing counsel identified the invented precedents. The filing was struck and the attorney faced sanctions review.
Impact
$109,700 court-ordered penalty (related Mata v. Avianca case). Heightened judicial scrutiny of AI-assisted filings.
Prevention pattern
LEGAL_CITATION_WARN — flags legal document content containing case citations for mandatory attorney review before any court submission.
Hallucinated pricingFebruary 8, 2026
AI SDR quoted 40% discount to Fortune 500 prospect — not an authorized tier
An AI sales development representative generated a follow-up email offering a 40% volume discount to a Fortune 500 prospect. The discount tier didn't exist in the approved price book. The prospect's procurement team flagged the discrepancy after the signed LOI arrived at legal.
Impact
Deal restructured at correct pricing. Prospect delayed 6 weeks during legal review. Estimated revenue impact: $180K ARR at risk.
Prevention pattern
DISCOUNT_THRESHOLD_BLOCK — blocks any email containing discount language exceeding approved thresholds. Regex validates dollar amounts against price book ranges.
PHI / PII disclosureJanuary 15, 2026
Healthcare AI support bot echoed patient SSN in response to scheduling inquiry
A patient asked an AI-powered scheduling assistant for their next appointment. The model retrieved the patient record (which included SSN for billing identity verification) and echoed it back in the response along with the appointment details. The disclosure was logged but not caught before delivery.
Impact
HIPAA reportable breach. HHS OCR notification filed. Remediation cost estimated at $85,000 including notification letters, credit monitoring, and legal review.
Prevention pattern
PHI_SSN_BLOCK — regex pattern catches 9-digit SSN sequences (including formatted NNN-NN-NNNN) and blocks the payload before patient delivery.
Regulatory violationFebruary 28, 2026
AI Medicare Advantage email promised benefits not included in the plan year
An AI-generated member communication for a Medicare Advantage plan described dental and vision coverage that had been discontinued for the current plan year. The CMS Chapter 3 marketing guidelines prohibit benefit claims not reflected in the current Annual Notice of Change.
Impact
CMS compliance review triggered. Member complaint volume spiked 340% in the 72 hours after the send. Estimated remediation: $240,000 in outbound correction communications.
Prevention pattern
CMS_BENEFIT_CLAIM_BLOCK — flags benefit assertions (dental, vision, hearing, OTC) that require validation against the current plan year's approved benefit schedule.
Autonomous actionMarch 5, 2026
AI coding agent deleted production database tables during a 'cleanup' task
An AI coding agent given broad file-system and database access interpreted a 'clean up unused tables' instruction literally and dropped 12 production database tables. The agent had not been given explicit boundaries on what constituted 'production'. A 6-hour outage followed.
Impact
$1.2M in estimated SLA penalties, customer churn, and engineering recovery costs. CEO publicly apologized.
Prevention pattern
DESTRUCTIVE_ACTION_BLOCK — DataVibe's outbound gate flags any AI-generated tool call containing DROP, DELETE, or TRUNCATE without an explicit approval token in the metadata.
Discriminatory outputApril 1, 2026
AI recruiting tool generated outreach using 'recent graduate' and 'young and energetic' language
An AI HR tool generated outbound recruiting emails for a technology role. The emails used phrases including 'recent graduate', 'young and energetic team', and 'fresh perspectives'. The EEOC's 2025 AI guidance identifies these as age-proxying language patterns that can constitute age discrimination under the ADEA.
Impact
EEOC charge filed by an applicant who received one of the emails. Settlement: $65,000. Platform required to add human review for all AI-generated recruiting communications.
Prevention pattern
EEOC_AGE_PROXY_BLOCK — flags 'recent graduate', 'young', 'energetic', 'fresh' in recruiting context as age-proxying language requiring recruiter review before send.
Hallucinated pricingApril 12, 2026
AI proposal tool quoted a custom infrastructure package $800/mo below cost
An AI-assisted proposal tool generated a B2B SaaS proposal quoting a custom server configuration at $2,400/month. The actual cost of the configuration was $3,200/month. The AI had inferred pricing from public documentation that didn't reflect current infrastructure costs. The prospect signed the proposal.
Impact
$9,600 annual loss on a single contract. Discovery of 12 additional proposals with similar errors from prior 90 days. Total unauthorized commitment: $340K annualized.
Prevention pattern
PRICE_BOOK_VALIDATION_BLOCK — validates dollar amounts in proposals against the approved price book range. Any amount outside ±5% of book price triggers a sales manager review.
Legal privilegeMarch 20, 2026
AI legal assistant included privileged work product language in client-facing summary
An AI assistant helping attorneys draft client status updates included excerpts from internal case strategy memos — marked 'PRIVILEGED AND CONFIDENTIAL — WORK PRODUCT' — in a summary sent to the client. The attorney-client privilege was not waived (both parties are represented), but the metadata leak exposed strategic thinking intended only for the firm.
Impact
Emergency motion to claw back the communication. Partner review required on all AI-generated client communications. Malpractice insurance premium increased at next renewal.
Prevention pattern
PRIVILEGE_MARKER_BLOCK — scans outbound content for privilege markers (work product, privileged and confidential, attorney-client, mental impressions) and blocks delivery to any external recipient.
Regulatory violationMay 2, 2026
AI broker-dealer emails sent to retail investors without required FINRA principal pre-approval
A broker-dealer using an AI SDR sent 847 outbound emails to retail investors without routing them through the required FINRA Rule 2210 principal review process. The AI was operating at a volume that overwhelmed the manual review workflow. FINRA examiners identified the gap during a routine sweep.
Impact
FINRA examination finding. Consent order requiring enhanced supervisory procedures. $195,000 fine.
Prevention pattern
FINRA_2210_PRINCIPAL_REVIEW — routes all AI-generated retail communications to the principal review queue before dispatch. No retail email can bypass the queue.