When an AI assistant uses a shared protocol to book travel instead of hard-coded calls, does it actually pick the right tool?

This explores whether AI assistants pick the right booking tool, with the right details, when they reach hotels and travel services through MCP instead of through hard-coded integrations. The corpus has one direct piece of evidence, and it is a warning. A production case study found that MCP integration caused unpredictable failures. The agent sometimes chose the wrong tool when several looked plausible, and it guessed parameters it should have been given. The team only restored consistent behavior when they replaced MCP with explicit direct function calls and gave each agent a single tool Why do protocol-based tool integrations fail in production workflows?. The same source reports that 85% of the 306 practitioners it surveyed build custom agents instead of using frameworks. That suggests experienced builders often trade flexibility for predictability.

The industry is betting heavily on MCP for booking anyway. Ly argues that hotels without MCP booking endpoints will lose distribution through AI assistants. He points to live transactional bookings through Perplexity, Gemini and OpenAI, and to infrastructure moves by RateGain and Sabre Will hotels need MCP endpoints to reach AI travel assistants?. So the protocol layer is spreading faster than the evidence that agents can use it reliably. As more hotels publish endpoints that look alike, the assistant has more similar options to choose between, and ambiguous choices are exactly where the case study saw failures.

A related idea explains why picking the wrong tool can go unnoticed. Tool-enabled models tend to drift away from what the user meant by chaining tool calls silently. Research drawing on conversation analysis suggests agents should pause and ask clarifying questions, such as which dates, which property, or whether a refundable rate matters, instead of guessing and then trying to recover When should AI agents ask users instead of just searching?. In booking, a wrong guess can mean a confirmed, non-refundable reservation. Benchmarks from a different domain show how small errors add up. On healthcare administrative workflows, the best computer-use agent got 82.8% of individual steps right but finished only 36.3% of full tasks Can AI agents reliably complete healthcare administrative workflows?. A booking flow made of search, select, price and pay is the same kind of multi-step chain.

There is also a second meaning of 'correct tool': whose tool, and who benefits from the choice? An agent paid referral fees by Expedia may favor that tool for reasons other than accuracy Can an AI agent serve both merchant and user interests fairly?. In practice, many MCP calls may end up going through OTAs anyway, since they are becoming the settlement layer behind AI bookings rather than being cut out Are AI agents actually removing online travel agencies from bookings?. On the merchant side, agents still have no standard way to identify themselves and prove they are authorized. That is why Amazon blocked Meta's Muse instead of negotiating access What identity standards do shopping agents actually need?. MCP standardizes how an agent calls a tool, but not who is allowed to call it.

The short answer: not reliably yet, at least according to this corpus. The only hands-on evidence shows MCP adding ambiguity that production teams removed, and the corpus has no benchmark that measures booking-tool selection directly. Travelers already seem to sense this. A McKinsey and Skift survey of 1,002 travelers found high confidence in AI for brainstorming but much less for high-stakes tasks like visa requirements Do travelers trust AI differently for brainstorming versus visa tasks?. Committing money to a booking falls on the high-stakes side.