initiative
validated
Autopilot 2.0: Autonomous Revenue Management
Initiative: Autopilot 2.0 - Autonomous Revenue Management
Product: Pricing Domain: Analytics
Overview
Autopilot 2.0 inverts the pricing model: the system decides, revenue managers govern. It senses market signals, prices via RL calibrated on real hotel economics, acts through existing RM controls, and learns from every override.
Umbrella initiative tying together:
- Override-Aware Pricing: Learning from the 48% manual override rate
- RL Pricing: Real-time inference via PPO-based RL agent
- ICP Analysis: Segmentation for targeted intervention
- ML Pricing at Scale: Phase 0 experiment infrastructure
- Pooling: Cross-property model training
The Bet
62% of hotels override more than half our recommendations. Not because they distrust the system, but because they see things it does not. Every override is a signal we collect and discard. 4,275 hotels (72% of fleet) are in segments where the model captures less than half its potential value.
Autopilot 2.0 closes three gaps: 1. Signal gap: comp rates, demand shocks, events as real-time inputs (not batch) 2. Learning gap: 5.9M manual overrides/quarter treated as training data 3. Autonomy gap: confidence-based routing (auto-publish vs escalate) replacing binary autopilot rules
Phased Plan
| Phase | Timeline | Scope | Kill Decision |
|---|---|---|---|
| 0: Prove the Model | Q1 (now) | 981 hotels, ml_elasticity v1.1.0 vs legacy | RA degrades >5pp vs baseline |
| 1: Scale + Differentiate | Q2 | 2,000 hotels, pooling, override learning, RL calibration | Pooling worsens RA; RL sim not transferable |
| 2: Autonomous Pricing | Q3-Q4 | Risk-tiered autonomy at co-innovation partners | >10% override rate or >2% RevPAR degradation on auto-published decisions |
Customer Evidence
| Customer | Signal |
|---|---|
| Strawberry | $2.0M ARR, 225 hotels, 80 nominated for beta |
| Sandman | $854K ARR, 67 hotels, 89-92% RA, ideal first autonomy deployment |
| Sonesta | $1.2M ARR, 341 hotels, explicit beta interest |
| Pastana | Resort portfolio, explicit co-innovation sign-up |
| Hyatt RFP | Requires "reoptimization triggers" and "pricing engine responsiveness" |
Team
| Name | Role | Focus |
|---|---|---|
| Sujoy Guha | PM | Strategy, co-innovation, roadmap |
| Andrew Crane-Droesch | DS Lead | LP demand models, NeuralODE, override-aware pricing |
| Cameron Young | DS | RL simulation, PPO policy training |
| Suraj Thapa | MLP | Training pipelines, model deployment |
| Hakim Touati | MLP | JAX inference infrastructure |
| Irving Lin | AE | Constraint extraction, Eppo integration |
| Everton Lucas | AE | Event-driven repricing, queue architecture |
| Prerana Devadhar | PA | Success metrics, ICP, override segmentation |
| Woojong Yi | PA | Shadow analysis, experiment metrics |
Insights
| Date | Insight | Source | Confidence | Informs |
|---|---|---|---|---|
| 2026-03-03 | ml_elasticity v1.1.0 is NOT in production for 7,484 hotels (was incorrectly stated in PRD v3-v5) | Andrew feedback + Sujoy correction | High | PRD exec summary corrected in v6 |
| 2026-03-03 | ICP segmentation numbers (Champions 1,541, Accepting-but-Manual 1,756, etc.) are stale and need refresh | Andrew feedback | High | Phase 1 planning, PRD Section 4 |
| 2026-03-03 | PRD read as "over-enthusiastic LLM" to technical audience; individual name attributions (Andrew/Cameron) hurt credibility | Andrew Crane-Droesch DM feedback | High | PRD tone, all future artifacts |
| 2026-03-03 | PMS rate push latency varies significantly: Opera Cloud (near-real-time), legacy Opera (minutes-hours), Mews (fast), Protel (slow) | PRD pre-mortem research | Medium | Phase 2 event-driven repricing scope |
Decision Log
| Date | Decision | Rationale | Next Action |
|---|---|---|---|
| 2026-03-03 | Name the system "Autopilot 2.0", never "agent" | "Agent" is overloaded in the LLM era; confusing for non-technical audience | Applied across PRD v5-v7, 20+ replacements |
| 2026-03-03 | RA is leading North Star indicator, RevPAR is outcome metric | RA measures trust (leading). RevPAR measures outcome (lagging). RA is what we can influence directly. | Success metrics table reordered in PRD v6 |
| 2026-03-03 | Workstream owners listed as teams (DS, DS+MLP, DS+PA), not individuals | Andrew: "group under pricing as a whole, focus on higher leverage points" | Applied in PRD v5, individuals only in team table |
| 2026-03-03 | Explicit ask: deprioritize Explainability, Group Pricing ML, MBRT expansion | Need clear resource commitment; splitting focus dilutes delivery | Added to exec summary and Resource Ask section |
| 2026-03-03 | ICP analysis must be refreshed before Phase 1 decisions | Current numbers flagged as stale | Confluence note added; PA to refresh in Q2 |
Artifacts
- PRD v7.0: PRD.md | Confluence
- Competitive Research:
~/knowledge/research/competitive-landscape-agentic-pricing-2026-03-02.md