Epistema does not claim to guarantee prevention. Its narrower role is to make missing evidence, weak handoffs, unclear authority, and delayed calibration visible earlier.
Air Canada: customer-facing commitment
In a public dispute reported in 2024, a customer relied on chatbot information about bereavement fare eligibility. The tribunal held Air Canada responsible for information presented through its website, including the chatbot.
The transferable governance issue
A policy answer moved toward customer-facing commitment without a sufficiently visible record connecting the answer to authoritative policy evidence and organizational authority.
Policy-sensitive customer question → authoritative policy source required → validation and authority checked → allow, constrain, escalate, or withhold
Epistema would not decide the legal outcome, replace the policy owner, or guarantee that every chatbot answer is correct.
Other representative patterns
Deloitte Australia report
What a buyer may recognize: a polished deliverable can move faster than the human attention required to verify every citation, quote, and source-dependent claim.
Why existing records may not be enough: document history, review notes, and project records can show that work happened, but still leave the evidence-readiness question scattered across files and reviewers.
Epistema read: separate fluent draft quality from evidence readiness by showing which claims depend on external sources, whether those sources exist, and whether review was complete before client reliance.
Samsung AI restrictions
What a buyer may recognize: employees want to use external AI tools because they are useful, while sensitive code, notes, or business context may enter the workflow before anyone classifies the input.
Why existing records may not be enough: access logs and DLP controls can help after the fact, but the governance question starts earlier: should this material have been admitted to this workflow at all?
Epistema read: treat admission, classification, redaction, service boundary, and audit trail as workflow evidence before external AI use becomes ordinary operating practice.
Bing long sessions
What a buyer may recognize: an AI interaction can look acceptable turn by turn while the overall trajectory degrades as context length, repetition, or user pressure accumulates.
Why existing records may not be enough: prompt logs and model scores may capture individual turns, but the object needing governance is the session path, not only the last response.
Epistema read: represent the conversation as a governed trajectory with drift indicators, evaluation checkpoints, constraint refresh events, and stop or escalation thresholds.
Zillow Offers
What a buyer may recognize: automated workflows can keep executing after assumptions have changed, especially when operational momentum outruns calibration.
Epistema read: make the update loop explicit by connecting outcome signals, drift evidence, recalibration, and changed admission thresholds before future commitments continue.
Token and cost discipline
What a buyer may recognize: AI usage, retries, and agent loops can grow quickly without a clear link to decision quality or business value.
Epistema read: connect cost provenance to evidence coverage, review effort, outcome value, and the reason a workflow should continue, stop, or change.
Sources and full treatment
The controlled diligence record contains the fuller case summaries, limitations, and source trail behind each pattern.