Ask an agency what slows down selling a ticket and the answer is rarely the fare. It is the conditions attached to it: what a change costs, when a ticket becomes non-refundable, which waivers apply and how children are priced. Airlines file these rules as text, and each carrier phrases them its own way. The result is a quiet tax on every booking — penalty quotes that are wrong, customer disputes when they are, and revenue that leaks in both directions.
Why rule-based parsing hit a wall
For years the industry answered this with regular expressions and hand-written rules. They work for the formats someone has already seen. Every new airline means new development, every format change means a fix, and the long tail of unusual wording never quite goes away.
Rule-based parsing vs GenAI on fare rules
Design comparison of the two approaches from the build, not client-measured results
| Measure | Rule-based parsing | GenAI pipeline |
|---|---|---|
| New airline onboarding | 2-4 weeks of development | Zero-shot |
| Edge-case failures | 8-15% | Under 1% |
| Maintenance cost | High, per airline | Near zero |
| Format changes | Manual updates | Adapts automatically |
What a production pipeline looks like
A language model on its own is not the answer; an unvalidated model simply moves the errors somewhere harder to see. The pipeline that worked in our delivered work has five steps: parse the raw rule text, run LLM inference, extract entities into a fixed schema, validate across categories, and only then output API-ready JSON. Categories are processed in parallel, airline-specific logic covers the carriers that need it, and anything the validator cannot reconcile goes to a person.
- Parse. Read the ATPCO rule text and its category.
- Infer. Use the LLM to interpret conditions, windows and amounts.
- Extract. Map the result to a structured penalty or discount schema.
- Validate. Check consistency across categories and against known rules.
- Review. Send ambiguous cases to a human before they reach a quote.
NDC raises the stakes
IATA's New Distribution Capability lets airlines distribute richer offers — bundles, ancillaries and branded fares — through direct APIs as well as GDS channels3. That makes content more valuable and harder to compare. Agencies now reconcile offers that look different on every channel. Structured, validated conditions are the common language that lets them sell those offers without guessing.
The model reads the rule. The pipeline decides whether the answer is safe to quote. (DaasLabs view)