DevRev
Case study · Live in production

Autonomous payment operations at enterprise scale

How One of the world's largest airlines transformed payment exception handling with AI-driven resolution across 4 legacy systems.

Computer
Case study · Live in production

One of the world's largest airlines: AI Payment Resolution Agent

Autonomous resolution across 4 enterprise systems. The AI investigates, validates, and resolves payment exceptions - the way an experienced specialist would.

4

Enterprise systems orchestrated in real-time

75%

Reduction in resolution time

90%

Lean L2 Operations - 4 agents replacing a team of 42

38

Resolution workflows powered by a single AI agent

Key insight

The agent uses Computer's memory as session state, skills orchestrate live across legacy systems, and the LLM manages the flow end-to-end.

DevRev · ConfidentialCase study
Architecture · Context engineering

AI Resolution Engine: autonomous operations across enterprise systems

TRIGGER Ticket Created payment issue arrives COMPUTER AGENT Prompt-engineered state machine · 16+ states Case Lookup Flight Search Sell + Save Ancillary Services Fare Override Commit + Pay Hold Release Refund Verify Computer memory: context · PNR · OrderID · flags SKILLS → LIVE API CALLS Booking Engine search, book, modify, commit, ancillaries Payment Gateway links, polling, refunds, verification Booking Session History PNR / transaction details Ticketing Tool ticket lifecycle, status updates, close
Context engineering

Computer memory holds session context (PNR, order IDs, state) so the agent never loses track or asks humans to repeat.

Live orchestration

Skills call legacy systems in real-time. Context assembled per interaction, on demand.

Deterministic + AI

Skills handle deterministic logic (fare math, ancillary codes); the LLM handles flow orchestration and UX.

DevRev · ConfidentialArchitecture
How we built it · State Engineering

The State Loop: one transition per turn

The Ledger Never Lies Reader (State In) Decider (LLM) Guard (Validation) Committer (Write Back) Observer (Trace) Terminals (Exit Check)

Core principle

The entire state machine is engineered into the prompt. The AI follows a deterministic decision graph - no guessing, no skipping steps. Ground truth externalized to the ledger.

Five layers

Ledger · Transition · Guards · Terminals · Observability - each layer eliminates a class of production failure.

Production results

AI automation reduced resolution from 30-45 min to 5-10 min, enabling 4 L2 agents to handle workload previously managed by 42.

Failures eliminated

Hallucinated state drift · Infinite transition loops · Skill-call drift - all gone when state is externalized.

DevRev · ConfidentialState Engineering
What was built · 8 weeks

High-value payment scenarios: 38 resolution workflows

Failed Payment Recovery & Rebooking

Search, sell itinerary, save passengers, replicate ancillary services and seats - all automated across booking engine APIs.

Automated Refund Processing

Generate payment links, poll status, handle refund verification, fare overrides with tax calculation.

Customer Case Resolution

Ticket intake, status updates, email notifications, and auto-close - full loop from open to resolved.

Duplicate Charge & Hold Resolution

Release held PNRs, modify travel dates, verify refund status by PNR or transaction, manifest lookup for lost PNRs.

How it's different
AI investigates the issue across all systems autonomously
Full resolution context maintained - no repeated contacts
Resolution completed, customer notified, case closed automatically
Deterministic business logic - zero hallucination on financials
Expert review for high-complexity exceptions only
Production-ready in 8 weeks - no training data required
DevRev · ConfidentialWhat was built
DevRev

Context engineering works.

Your data stays where it lives. Computer orchestrates across systems in real-time, fetching only what's needed per interaction.

DevRev Team
Solutions Engineering
Computer
devrev.ai
01 / 05