TL;DR
The model licence is the cheap part of AI in travel. The real bill arrives in the plumbing: fragmented inventory and booking data, APIs that language models struggle to read, consent and automated-decision rules in the UAE, Singapore and Australia, and the monitoring needed to keep an AI assistant from telling a customer something your company then has to honour. Treat integration as the project and the model as a component, and the economics look very different.
Why the invoice is never the demo
Most travel AI pilots look excellent in a demo. A chatbot answers a fare-rules question, an assistant drafts an itinerary, a pricing model suggests a better rate. Then someone asks it to run against the live reservation system, the loyalty platform, three supplier connections and a payment gateway, and the project changes shape.
We see the same pattern in how the industry describes itself. In Phocuswright's study "Budgets, Barriers and the Race to Agentic AI", summarised in an ITB 360° piece, more than 60% of travel businesses are experimenting with or scaling agentic AI, but only 6% are scaling it across multiple domains and functions. The same write-up says most deployments remain tactical, with AI layered onto existing infrastructure, and that the leading obstacles are integration complexity, talent gaps and data security concerns rather than lack of interest or funding. (The piece is sponsored content and the full study is not published on the page, so treat the figures as indicative.)
Airlines tell a similar story. SITA's 2025 Air Transport IT Insights, reported by PhocusWire, found that 79% of airline IT professionals rank AI as an investment priority for the next 12 months, yet data challenges hamper AI scaling for 49% of airlines. Data integration and data quality were each named as a top IT challenge by 49%, and process and culture transformation by 62%.
Read those numbers together and the argument writes itself: appetite is not the constraint. Integration is.
Where the hidden costs actually sit
1. Data that was never designed to be read by a model
A travel tech stack is rarely one system. It is a property management system here, a channel manager there, a GDS or NDC connection, a CRM, a loyalty engine, a payments layer, and a handful of spreadsheets that someone's operations team depends on. Each has its own idea of what a "customer", a "booking" or a "rate plan" is.
AI exposes that inconsistency immediately. A model asked to explain a cancellation policy will happily combine a fare rule from one system with a promotion from another. The failure is not intelligence; it is that no single source of truth exists for it to read. The unglamorous work of reconciling identifiers, normalising rate and policy structures and deciding which system wins in a conflict is usually the largest line of effort in an AI programme, and it rarely appears in the original business case.
If you are weighing this for a multi-supplier product, our piece on multi-supplier travel aggregation architecture covers the normalisation layer in more depth.
2. APIs that are fine for developers and awkward for models
Amadeus' senior vice president of engineering, Gaëlle Bristiel, put the problem plainly in a sponsored Skift Q&A (April 2026): legacy infrastructure, many APIs that are hard for large language models to interpret, and strict regulation make scaling difficult, and bolting AI onto existing systems often fails at scale. Because the interview is sponsored by a vendor with something to sell, take the commercial framing with a pinch of salt. The technical observation, though, matches what we see in delivery.
Travel APIs tend to be verbose, stateful and full of codes. A human integrator reads the documentation once and writes a mapping. A model-driven agent has to infer meaning on every call, and an ambiguous field becomes an ambiguous answer. The practical cost is an interface layer: clean, narrowly scoped tools with predictable inputs and outputs that an AI component can call safely. This is ordinary API development and integration work, but it has to be funded and scheduled, and it is the reason "we will just connect the model to our booking engine" is almost never a two-week task.
3. Orchestration and fallbacks
Bristiel's second point in the same interview is worth repeating: the industry is moving toward specialised agents with narrow scopes, which she argues are more reliable and produce fewer hallucinations, and AI should handle language and intent while conventional software handles precise optimisation and fulfilment.
We would put it more bluntly. Do not let a language model decide a price, a refund amount or a seat allocation. Let it understand the request, then hand the decision to deterministic code with rules you can audit. That split sounds obvious, but it forces you to build the deterministic layer, the hand-off contract between the two, and a fallback path to a human when confidence is low. All of that is engineering and operations cost.
4. Testing, monitoring and the cost of being wrong
Traditional software is tested against fixed expected outputs. A conversational or agentic system produces a range of valid outputs, so you need evaluation sets, regression checks on every model or prompt change, logging of what the assistant said and why, and someone responsible for reviewing it. Bristiel notes the same need for monitoring, testing that accepts a range of valid outputs, and explainability when moving past pilots.
The cost of skipping this is not theoretical. In 2024 the British Columbia Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation after its website chatbot gave a passenger wrong information about bereavement fares. As summarised by the American Bar Association, the airline argued the chatbot was a separate legal entity; the tribunal rejected that, held the airline responsible for all information on its website, and said customers should not have to cross-check the chatbot against another page. The award was small, roughly CAD 650 plus interest and fees. The precedent is not: what your assistant says is what your company said.
5. Compliance work that scales with every market
For a travel business selling into the Gulf, Singapore and Australia, AI features touch personal data in ways each regime treats differently. This is not legal advice, and you should confirm specifics with counsel, but the direction of travel is clear.
UAE. According to a 2026 Al Tamimi analysis of Federal Decree-Law No. 45 of 2021 (PDPL), data subjects can object to decisions resulting from automated processing, including profiling, and controllers must provide human review on request. The law's definition reaches processing with "limited human supervision", which is wider than fully automated decisions. Cross-border transfers are permitted only on specified bases, and controllers must tell data subjects about the safeguards that apply. If your model is hosted outside the UAE, that is an architecture question, not just a contract one.
Singapore. The PDPC published its Advisory Guidelines on the Use of Personal Data in AI Recommendation and Decision Systems on 1 March 2024, as summarised by Rajah & Tann. They address consent and notification across development, testing, deployment and procurement, and set out accountability expectations for both businesses deploying AI to consumers and those procuring it. If you buy a recommendation engine rather than build one, the guidance still applies to you.
Australia. From 10 December 2026, under the Privacy and Other Legislation Amendment Act 2024, organisations covered by the Australian Privacy Principles must say in their privacy policy where a computer program makes, or substantially and directly supports, a decision that could significantly affect an individual's rights or interests and uses their personal information, according to the OAIC. That means describing the kinds of personal information used and the kinds of decisions involved. Dynamic pricing, fraud screening and automated refund decisions are the travel examples that deserve a careful look.
None of these is a blocker. Each is a design input: where data is stored and processed, what is logged, how a human reviews a decision, and what the privacy notice says. Retrofitting them after launch is far more expensive than building them in.
6. The licence-versus-everything-else gap
Gartner has predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, as reported by New Electronics. Gartner analyst Anushree Verma is quoted as saying many use cases positioned as agentic today do not require agentic implementations, and the firm also warns of "agent washing", where vendors rebrand chatbots, RPA and assistants as agentic.
That is a useful filter for travel buyers. Before funding an agent, ask whether a well-built search, a rules engine or a plain retrieval assistant would solve the same problem for far less integration effort.
A simple way to size the real cost
You cannot responsibly publish a single number for "AI integration cost" across travel businesses; it depends on your estate. What you can do is stop budgeting only for the model and start listing the work around it. A useful checklist:
| Cost area | Question to answer before approving the project |
|---|---|
| Data foundation | Which system is the source of truth for bookings, rates, policies and customers, and who owns conflicts? |
| Interface layer | Do clean, narrowly scoped APIs exist for the actions the AI will perform, or must they be built? |
| Decision boundary | Which decisions are deterministic code, which are model output, and where does a human step in? |
| Evaluation and monitoring | How will we test changes, log responses and catch drift before customers do? |
| Compliance | Where does data live, what must the privacy notice disclose, and what consent applies in each market? |
| Payments and security | Does the AI touch cardholder data or payment flows, and is it inside your audit scope? |
| Change management | Who in operations and support will own, correct and retrain the system? |
Payments deserve a flag of their own. If an assistant can initiate or influence a transaction, your PCI scope may grow; our guide to PCI DSS 4.0.1 for travel platforms explains what custom payment architecture needs to address.
Where AI does earn its keep in travel
The sceptical case is not a case against AI. Done on solid foundations, the use cases are real. SITA's survey reports that 75% of airlines already use AI for customer service and 63% in operations control for disruption, aircraft assignment and crew availability. Disruption handling is a good example of a place where the integration cost pays back, because the data already exists and the decisions are time-critical; we cover the patterns in disruption management systems for travel platforms.
Other candidates with a sensible risk profile include internal tools for contact-centre agents, itinerary drafting that a human reviews before it reaches a customer, content and description generation for suppliers, and demand and pricing analysis that feeds a deterministic engine. Customer-facing agents that can commit money or policy are a later step, not a first one.
Distribution is shifting too. If your inventory may soon be booked by someone else's AI assistant, your systems need to be readable and trustworthy to machines; see our analysis of agentic AI and OTA booking.
What to do this quarter
- Inventory your data and your APIs. Map every system a customer-facing AI feature would touch, name the source of truth for each entity, and list the integrations that do not yet exist.
- Pick one bounded use case. Choose something with existing data, a human in the loop and a measurable outcome such as handling time or first-contact resolution. Avoid autonomous refunds and pricing for now.
- Draw the decision boundary on paper. Write down which decisions the model may suggest and which only deterministic code may make. Review it with operations and legal.
- Build the evaluation set before the feature. Collect real, anonymised customer questions, define what a good answer looks like, and re-run the set on every change.
- Check market obligations early. For the UAE, Singapore and Australia, have counsel confirm consent, cross-border hosting, human-review rights and the December 2026 privacy-policy disclosure where relevant.
- Ask vendors hard questions. What does the product do that a rules engine would not? Where is data processed? What logging and human-override controls ship by default?
- Set a stop condition. Agree in advance what result would end the pilot, so the project can be cancelled on evidence rather than sunk cost.
How Syslabs approaches this
We build and integrate the layers around the model for travel businesses: API and integration work, AI chatbots and assistants, and platforms across our travel industry practice. If you are deciding whether to build or buy ground-operator or booking software for the UAE, Singapore or Australia, our DMC and ground operator build-versus-buy guide is a good starting point. Scope and pricing depend on your systems and are set out on a consultation, not on a rate card.
Next step
If you want a second opinion on the real integration scope of an AI project in your travel stack, book a meeting with Syslabs.
References
- Gartner prediction on agentic AI cancellations, as reported by New Electronics: https://www.newelectronics.co.uk/content/news/over-40-of-agentic-ai-projects-could-be-cancelled-by-2027
- ITB 360° / Phocuswright, "2026: Travel's race to scale agentic AI" (sponsored): https://www.itb.com/en/itb-360°/news-insights/2026-travel’s-race-to-scale-agentic-ai.html
- PhocusWire, SITA Air Transport IT Insights 2025: https://www.phocuswire.com/news/technology/sita-airline-it-sights-2025
- Skift, "Why AI Integration Will Define Travel's Next Competitive Era" (sponsored, 8 April 2026): https://skift.com/2026/04/08/ai-infrastructure-travel-industry-orchestration-scale/
- American Bar Association, BC tribunal and AI chatbot liability: https://www.americanbar.org/groups/business_law/resources/business-law-today/2024-february/bc-tribunal-confirms-companies-remain-liable-information-provided-ai-chatbot/
- Al Tamimi, data regulation in the AI era: https://www.tamimi.com/law-update/the-technology-issue-2026/articles/data-regulation-in-the-ai-era-a-comparison-of-korean-and-uae-data-protection-laws/
- Rajah & Tann, PDPC AI advisory guidelines: https://www.rajahtannasia.com/viewpoints/pdpc-issues-advisory-guidelines-on-use-of-personal-data-in-ai-recommendation-and-decision-systems/
- OAIC, transparency for AI and automated decision-making: https://www.oaic.gov.au/news/media-centre/new-resources-on-transparency-for-use-of-ai-and-automated-decision-making
