AI booking, AI trip planning,
and AI in hospitality — shipped on your GDS, NDC, and OTA stack.

An `ai in travel` engagement we ship looks like this: scope which workflow moves a P&L line (most often `ai booking` on the search-to-offer path, `ai trip planning` on long-tail conversion, `ai revenue management` on rate decisioning, `ai in hospitality` on guest messaging, IROPS draft fanout on ops desks, or a `travel chatbot` on pre-trip / in-trip / post-trip lifecycle moments), get the model-endpoint approval matrix signed off for PNR / GDPR scope, build the integration against your real Amadeus / Sabre / Travelport / NDC tenant, pick the right model per workflow (Sonnet 4.6 for revenue + IROPS rationale, Haiku 4.5 for high-volume search + drafts, GPT-5.4-mini for structured intake), bake in audit-logging on every inference, ship behind a human-in-loop gate, then operate the workflow long enough to prove cost-of-ownership before scaling. We do not sell a product; we ship one workflow at a time and report cost-per-decision monthly. First workflow live in six to eight weeks.

`Ai booking` is the highest-volume keyword in the travel cluster (8,100/mo) because every OTA, TMC, and direct-booker is rebuilding the search-to-confirm path around an AI agent. The operator-grade pattern: agent sits in front of the GDS / NDC, reads the traveler's free-text query, drafts the structured search, scores the GDS / NDC response against the query intent, assembles the offer cards. Inventory truth never leaves the GDS / NDC adapter — the agent never invents an offer. Traveler confirms before any segment is held or ticketed; autonomous ticketing is a regulated commercial decision we don't ship without explicit scoping. On Amadeus / Sabre / Travelport, write per-GDS adapters — the dialects of EDIFACT and the quirks on availability vs schedule queries are different enough that a single "GDS layer" abstraction is the wrong shape. NDC adoption is uneven — carrier NDC level (1 / 2 / 3 / 4) determines what's possible, so the orchestrator (§6) handles the routing decision per call.

`Ai revenue management` is the highest-CPC keyword in the travel cluster ($21.35) for a reason: every dollar moved on a published rate compounds across the year. The pattern that ships: agent reads the competitive-set rates, the pickup curve, the event calendar, the inventory position; drafts a rate + LOS-restriction recommendation per night per room-type; the revenue manager reviews and approves before push to the PMS / channel manager. Never autonomous publish on rates over a configurable threshold. Realistic outcomes on mid-scale chains: a 1.5–3 percentage-point RevPAR lift directional in published trials, with the bulk of the value coming from the rate-discipline consistency across portfolio rather than peak-night home runs. The audit will tell you whether your compset + pickup data is clean enough to build an honest baseline — if it isn't, that's the first phase before any model touches a rate.

PNR data is regulated personal data under GDPR with extra friction on cross-border transfers (EU↔US, EU↔UAE, EU↔SG, etc.) and on retention scope. Operator-grade specifics: (1) Model-endpoint approval matrix signed off before any PNR is in scope — your compliance team confirms which model endpoints, in which jurisdictions, are approved for which data classes. (2) PNR minimisation pre-prompt: we strip the PNR to the minimum viable payload before the model call. The agent doesn't need the full passport scan + frequent-flyer-number history to draft a re-route. (3) Region-resident inference where the data origin requires it: EU-resident endpoints for EU-originated PNRs, with the vendor's data-processing addendum (DPA) and standard contractual clauses on file. (4) Audit log on every inference — request ID, model, tokens, scrubbing actions, retrieved context, the reviewer who approved the output. What we don't claim: a turnkey "GDPR-compliant AI" product — GDPR compliance is a posture of your team, your vendors, and your model-endpoint contracts, and the audit deliverable starts by mapping that posture honestly.

Yes — and the failure modes are different per stack, called out in §6 above. Quick recap. Amadeus: EDIFACT-based, MQ transport, deep schemas; the quirks live in availability vs schedule queries and in PNR command structure. Sabre: similar EDIFACT shape with different dialect choices and a different command grammar on Sabre Red — per-host adapter is the right shape, not a generic "GDS layer". Travelport (Galileo / Apollo / Worldspan): three legacy hosts under one brand with three different command grammars; capability varies by host. NDC: modern XML-over-HTTPS standard but adoption is uneven — carrier NDC level (1 / 2 / 3 / 4) determines what's possible, and ancillary support is the laggard. The pattern we ship: one orchestrator at the center (§6 octopus), per-channel adapters as spokes, the AI agent owns interpretation + draft + scoring, the channel adapter owns the real call. Channel managers (SiteMinder, RateGain, Cloudbeds) abstract a lot of the OTA-side mess but introduce their own retry-semantics quirks — design idempotent on the orchestrator side.

Four places we'll say no — covered in §8 above and worth repeating. (1) Dumping a full PNR into a non-scoped model endpoint is GDPR exposure, not a feature. Scope the model-endpoint approval matrix before any inference. (2) Autonomous customer comms on DOT-regulated decisions (refund denial, Tarmac Delay disclosures, mandatory baggage / cancellation notifications) is a CFR-violation waiting to happen — bot drafts, human approves, customer hears. (3) Marketing an NDC level the carrier list doesn't actually support is how engagements die at month 4 — the audit checks carrier capability honestly and specs the GDS fallback. (4) Shipping an AI-built booking flow that breaks accessibility (keyboard traps, missing labels, contrast regressions, screen-reader-invisible dynamic content) is ADA exposure under DOJ guidance — every AI-driven UI ships with an accessibility eval suite in CI or it doesn't ship. Beyond those four, we'll also say no on any workflow where the data isn't clean enough for a baseline, where the regulatory posture is unclear, or where the metric won't move within the pilot window.

Both buy `ai in travel` services and they share an orchestrator topology (§6), but the workflows that pay back are different. `Ai in hospitality` ($16.82 CPC) tilts heavily toward guest-message reply drafting, revenue management ($21.35 CPC), arrival-day exception management, housekeeping-route inputs, and pre-arrival nudges — bimodal load on flu season + event weeks rather than discrete IROPS events. `Hotel ai` is brand-voice-intensive: a 40-property portfolio with one tone-eval suite per brand keeps the guest experience consistent. `Airline ai` and `aviation ai` tilt toward IROPS draft fanout (§5 is the exact shape), gate-agent assist, crew-swap drafts against FAR Part 117 duty limits, IATA NDC-level-aware booking and servicing, and ancillary attach on confirmation. The disruption-event shape is the design constraint: airlines design around a 9-times-a-year cascade event, hotels design around a 90-day surge season. Both ship behind a human-in-loop gate.

Three tiers, pricing-locked across the cluster. (1) Travel AI audit: $3K fixed, 1–2 weeks. We shadow commercial + distribution + revenue + ops, score candidate workflows, deliver a 90-day roadmap with per-workflow cost bands, a model-endpoint approval matrix for PNR / GDPR scope, and an honest "these won't pay back yet" list. (2) Pilot to production: $10–25K fixed, 4–8 weeks. One workflow shipped end-to-end on your Amadeus / Sabre / Travelport / NDC stack, human-in-loop gate, accessibility eval in CI, surge-mode config tested, walk-away point at week 6 — if the metric won't move, we stop before production hardening and you don't pay phase 2. (3) Continuous travel AI team: from $5K/month, no annual contract. Embedded PM + AI engineer + travel-ops analyst shipping the next workflow on your roadmap, with per-workflow monthly cost-of-ownership reporting and a surge-readiness review before peak booking + holiday-IROPS windows. Median model cost at scale is ~$0.003 per pre-trip / post-trip message and ~$0.04 per revenue-management rate decision — verify on your own usage before locking a pick.