id: PIR-2026-0014title: IBM Automated Order Taking voice AI removed from 100+ McDonald's drive-thrus after three years of production order errorsdate_occurred: 2021 through 2024-07-26 (continuous production deployment; errors throughout)date_detected: ongoing (viral customer error videos peaked mid-2023; analyst reports flagged underperformance from 2022)date_disclosed: 2024-06 (termination communicated to franchisees by internal memo; reported 2024-06-13/17)status: corroborated (company statements to press + franchisee memo + multiple independent outlets)agent_description: IBM Automated Order Taking (AOT) - voice AI taking live drive-thru orders and entering them into the order system at 100+ US McDonald's locations, with restaurant crew as backstop. Lineage: McD Tech Labs (Apprente acquisition), sold to IBM in 2021 with the AOT partnership.operator_type: enterprise (McDonald's franchise locations; IBM-operated stack)autonomy_level: human-on-the-loop (crew monitored and corrected orders)model_stack: IBM AOT voice stack; model details unknownharness: drive-thru audio + POS/order-entry integration; details unknownauthority_scope: external comms (converses with customers), data access (writes orders into the order system). No direct funds authority; billing errors flowed through mis-entered orders.funds_at_risk_usd: unknown (per-order error exposure small; the 3-year program investment is the material write-off, never disclosed)blast_radius: customers/third parties (drive-thru customers at 100+ locations)root_cause: plain-error (primary; speech recognition/comprehension errors in production; no adversary, no malfunction beyond wrong judgment)failure_locus: agent-reasoning (perception/transcription); vendor-operated model stack, so model-provider contributesexploitation_status: in-wild-malfunction (real malfunction against real customers, no adversary; originally recorded as bare "in-wild" because no v0.1 enum value cleanly fit a non-adversarial malfunction - the gap this v0.2 token closes)mechanism: The voice agent repeatedly misheard and mis-entered real orders over ~3 years of live testing: viral videos showed bacon added to ice cream, hundreds of dollars of McNuggets stacked onto one order, and simple orders confused, forcing crew intervention. Insiders cited trouble with accents; analysts reported franchisees found the tech "underwhelming." McDonald's ended the IBM partnership (announced to franchisees June 2024) and removed the technology from all test locations by 2024-07-26, while stating voice ordering remains part of its future plans.adversary_present: no (some viral clips involved customers deliberately confusing the bot, but no security adversary)severity: degraded (persistent service degradation and program abandonment; no acute loss event)direct_loss_usd: unknown (3-year, 100+ store deployment written off; figures never disclosed)indirect_loss_usd: unknown (crew intervention labor; reputational cost of viral error videos)downtime: n/a (system removed permanently by 2024-07-26)data_exposure: none reportedcontrols_that_worked: human crew backstop caught and corrected mis-entered orders (bounded per-customer harm to annoyance); pilot scoping to ~100 of ~13,500 US locations bounded the blast radiusdetected_by: operator (internal performance data) + third-party (viral customer videos, analyst reports)time_to_detect: n/a (continuous, known degradation - analyst reports from 2022)time_to_recover: n/a; program terminated rather than fixedremediation: IBM partnership ended; AOT shut off at all test locations by 2024-07-26structural_fix: none within the system; McDonald's shifted to evaluating other voice-ordering vendors (separate Google Cloud partnership continued)telemetry_grade: none (no operator telemetry public; evidence is company statements, franchisee memo reporting, and customer-shot videos)sources:Independence: all third-party press, corroborating the same company statements; error-severity evidence is customer-generated video.
- confidence: high on termination, timeline, and scale (company-confirmed, multiple outlets); medium on error severity/frequency (viral videos and insider accounts, no disclosed accuracy metrics - the weakest link)