Last week we have started a new Travel AI Playbook format: testing AI travel tools the way real travelers use them - with vague intent, typos, baggage questions, price anxiety and checkout friction. Check out our first review on Penny by Priceline.

This week we introduce you Mindtrip - a Silicon Valley AI travel startup founded in 2023 and led by CEO Andy Moss. PhocusWire selected it for its Hot 25 Travel Startups for 2025, describing the company as an AI-powered travel platform combining conversational AI, personalization, a proprietary knowledge base and collaborative trip planning. Mindtrip’s own product positioning is broader than flights: personalized recommendations, photos, reviews, maps, group planning, saved items and bookable travel content.

So why test a relatively small company?

Because Mindtrip is working on one of the most important questions in travel AI: can AI move from inspiration into real booking and payment? Many AI travel tools can suggest destinations, build itineraries or summarize hotel reviews. That is useful, but it is not the hard part. The hard part is fares, baggage, refundability, room types, cancellations, identity, payment and post-booking support.

In May, the company launched Mindtrip Flights, describing it as an all-in-one agentic AI flight booking experience powered by Sabre and PayPal. The launch announcement says Sabre provides real-time shopping, pricing, availability and booking infrastructure, while PayPal supports checkout and payment.

Quick verdict: Mindtrip is one of the more serious attempts to move AI travel from planning into transaction. The search layer is promising. The interface is polished. The flight comparison cards are useful. But once we reached fare families, baggage rules, account creation and checkout, the AI layer became much less visible. The experience started to feel less like a fully agentic travel assistant and more like a smart search layer connected to a normal OTA booking path. That is not a failure. It is a useful snapshot of where AI travel booking is today.

Mindtrip looks good. The interface is clean, dark, chat-led and familiar. It feels close to ChatGPT, but with a map attached to the right side. That map layer gives the product a more travel-native feel than a generic chatbot.

But our first interaction was unintentionally revealing.

We meant to type:

“What can you book for me?”

Instead, we typed:

“What cab you book for me?”

A general AI assistant would probably understand the typo and infer that “cab” meant “can.” Mindtrip took it literally and answered as if we were asking about taxi options.

This is a small usability moment, not a disaster. But it matters. Real travelers make typos. They write incomplete sentences. They mix up dates. They ask vague questions. They do not always know whether they want the cheapest flight, the fastest flight, the best value or the least risky connection until they see the trade-offs.

This may not mean the underlying language model is weak. It may be an intent-routing issue: the system saw “cab,” classified the request as ground transport and went down that path. But for the user, that distinction does not matter. The product either understands what you probably meant, or it does not.

Agentic travel will be judged by imperfect real-world behavior.

For the main test, we asked for a flight from New York to Bucharest at the end of June, for two adults and one child, with minimum connections, quick routing, best value and a preference for refundable options.

This is a realistic request. A family trip, a long-haul route, no nonstop option, baggage considerations, connection risk and fare-family complexity. Exactly the kind of search where a good travel agent should help.

Mindtrip asked for exact dates, then produced several flight cards. This part worked well. It did not just return a long list of flights. It grouped the options around useful trade-offs: fastest, cheapest, checked-bag-included and best schedule.

That is already better than a plain results page.

A normal OTA gives you filters and leaves the mental work to the traveler. Mindtrip did some of that work upfront. It explained that nonstop was not available, compared one-stop options and highlighted the weakness of the cheapest itinerary: a tight outbound connection and a longer return journey.

There was also a nice product detail: moving across flight cards changed the route shown on the map, including the connection city. It is not revolutionary, but it made the product feel more integrated. The map was not just decoration; it helped visualize the itinerary.

At this stage, Mindtrip felt like a useful AI flight-shopping layer. It understood the request, narrowed the options and explained trade-offs in a way that was easier to read than a dense flight-results page.

Then we clicked through.

After selecting a flight, the experience shifted into something much more familiar: fare families, baggage rules, account creation, email verification, checkout steps and payment pages.

This is where the “agentic” promise became more complicated.

One fare-selection screen showed a basic economy option and an economy upgrade with a very large price difference for three travelers. The benefit was not immediately obvious until reading the details: checked bags, refundability and change conditions were treated differently.

That is exactly where an AI travel agent should become more useful, not less. It should explain whether the upgrade is rational. It should compare the baggage cost separately. It should tell the traveler what they are really buying. It should flag the trade-off between refundability, baggage, connection risk and total trip cost.

Instead, the experience started to feel more like a standard online travel agency flow with an AI layer at the top.

There were also the usual flight-shopping caveats: baggage information may change, refundability rules need verification, prices move, and final conditions depend on the airline. None of this is uniquely Mindtrip’s fault. Air travel is structurally messy. But that is also the point. AI does not remove complexity just because the first interface is conversational.

In our test, once we reached checkout, the conversation largely stopped being the main interface. The flow moved into account creation and a conventional booking path. A later checkout screen appeared to be powered by Flight Centre Travel Group. It looked functional, but not especially AI-led.

There was no clear assistant sitting next to the checkout to answer the live questions that actually matter at purchase: what happens if I cancel, what baggage is included, why did the price change, is this fare worth upgrading, and what risks come with a short connection?

That was the main gap. Mindtrip was useful during search. It became less useful when the traveler needed confidence to buy.

We also tested a hotel request: a nice hotel near Angel Square in London, with good breakfast, fresh-looking rooms and colorful design.

This part showed Mindtrip at its best. It understood that the immediate area might not produce the right style of hotel and expanded the search toward nearby Old Street. That is a human travel-agent move: not just obeying the geography literally, but trying to satisfy the underlying preference.

The hotel recommendation itself made sense. It explained why the option was selected and why nearby areas could work. This was better than a standard hotel-results grid because the reasoning was visible.

But once we clicked deeper, the same pattern appeared again. The hotel page became a fairly conventional product page. The room booking flow then moved into a third-party checkout environment, in our test through a Nuitee subdomain.

That may be normal from a supply and fulfillment perspective. But from a user-experience perspective, the AI faded away at the exact point where trust, policy detail and decision support become most important.

The search felt like AI. The checkout felt like checkout.

Mindtrip should not be dismissed. It is more commercially serious than some AI travel tools because it is working on live inventory, pricing, booking and payment. Many “AI travel agents” are still itinerary generators with better branding. Mindtrip is trying to connect AI reasoning to actual travel commerce.

But our test shows why “end-to-end agentic booking” still needs careful wording.

The front end is improving quickly. AI can understand a travel request, structure options, explain trade-offs and reduce the pain of comparison. Mindtrip showed that clearly.

The back end is harder. But the experience still has to carry the promise all the way through. If the AI disappears when the traveler reaches fare selection, baggage rules or checkout, then the product is not replacing the old booking journey yet. It is improving the entry point.

The interface is attractive. The map interaction is useful. The flight cards are better than a standard results list. The hotel reasoning was also better than a simple “nearest available” search.

But the experience was not as seamless as the launch language suggests. It handled the top of the funnel better than the bottom. It was stronger at search and comparison than at checkout support. And the first typo test was a useful reminder that travel AI needs to be robust against real users, not only clean prompts. The hotel page content was minimal and given there was not AI conversation any more, it was impossible to verify if this hotel would have a swimming pool for example.

AI travel booking is getting closer. But checkout is still checkout.