Your client used AI to design a trip itinerary and asked you to book it
How to tell them it's terrible (scripts included)
You’ve been designing trips in Patagonia for years. You know the routes, the guides, the lodges that’ll bend the rules for a returning client. You know when to slow a group down on day three because the altitude will catch up with them on day five if you don’t. Every itinerary you send out is the product of years of fieldwork based on local wisdom and customer experiences.
Then your client writes to you explaining that their new favourite AI model has suggested some changes. Changes that make little-to-no sense. They want to add a town no one would recommend in November, and compress three days of acclimatisation into one, because apparently everyone on the trip is either Reinhold Messner or a mountain goat.
I’ve heard variations of this story a lot over the last 12 months. So I thought I’d offer some carefully considered suggestions to help itinerary designers, DMCs, and operators respond in a way that protects the client, the trip, and your reputation.
What’s actually going on
The first thing we should acknowledge is that AI edits aren’t really a reflection of mistrust in your ability to design an itinerary. It might be cost-anxiety, the desire to feel involved in trip design in a way that makes them feel like it’s truly their trip, or just a compulsion to always be ‘optimising’ everything. It might even be one spouse has been overruled by the other on a key choice, and they’re using Claude or ChatGPT as an opportunity to re-litigate the whole thing. Or maybe they just want to feel heard.
Whatever it is - and you should probably take some time to figure it out - the most important thing to note is that under no circumstance should you consider this a personal or professional rebuttal. Do not respond stating that AI is dumb or can’t do logic puzzles. The right response to “ChatGPT said we should add a day in El Calafate” is “What’s drawing you to El Calafate? Let me tell you what we considered and chose against, and why.”
You don’t need to get drawn into a discussion about the capabilities of AI. You need to figure out the signals behind their use of AI.
Acknowledge what AI is good at…
This is going to sound bizarre - but there is an argument to be made for encouraging its use.
It’s probably worth acknowledging that AI is actually quite good at travel planning. It’s excellent for packing lists, vocabulary, visa requirements, vaccination basics, first-pass research on a region a client has never visited, generating sensible questions to bring into a planning call. If a client wants to use ChatGPT to read up on the history of Estancia Cristina before they go, or to figure out how to order a Malbec in Spanish, that’s a client who’s going to arrive better prepared.
Rather than debating the capabilities of AI, acknowledge what it’s good for. Perhaps even tell them how you use it – first drafts, summaries, language polishing, translation, building out comparison tables for clients deciding between two routes. The point isn’t to hide that you use these tools. It might be that AI is freeing up the hours that drafting boilerplates used to consume, so you can spend more time on the part of the work that only you can do.
You don’t want to be seen as ‘in denial’ about the capabilities and potential of AI. So when a client raises AI in the conversation, “Yes, we use AI quite a bit as part of our work.” By conceding what it can do, you’re actually in a stronger position when you’re making the case for where the line needs to be drawn on its limitations.
…so you can justify its limitations
AI treats itineraries as a problem with a correct answer: the shortest route, most sights per day, the highest-rated hotels, the best-reviewed experiences. A large language model looks at thousands of Patagonia itineraries, reviews, blog posts, and articles, and identifies a best match based on a probabilistic determination. It’s a process of optimisation.
Design works differently. It starts with a person’s aspirations, their fitness, fears, the way they travel together as a couple, and the stories they want to tell their children. It builds an arc: anticipation, exertion, recovery, experience, awe, then recovery, and decompression. It accounts for weather windows that don’t show up on forecast models. It knows the best afternoons are sometimes unscheduled stops in a wine cellar with a monk that went to school with your guide. It is the type of design that doesn’t show up in the training data for large language models.
A lot of what your team knows has never made it onto the web: the refugio that opens early if you call ahead, the guide whose father used to run the southern crossing, the kitchen that’ll cook a private asado for ten if you ask the right person.
Some of that knowledge will eventually leak into a model’s training data. Most of it won’t, because by the time it does, your team has already moved on to the next thing. In other words, your moat isn’t one hidden gem. It’s the process of continuous discovery that happens through your networks and partnerships.
Just as the journalist is writing up their next feature for the Travel Section at the New York Times, you’ve moved on to finding, and building, the next big experience. Which is what you’re selling: access to experiences before they become commodities.
The duty of care argument
It isn’t hard to get a large language model to tell me I have a natural affinity for animals, or that I’m capable of summiting Everest. That doesn’t mean I have the authority to jump out of the safari vehicle for a lion selfie, or the lung capacity to self-finance an expedition through the Khumbu Icefall. That is because large language models tend to agree with the version of me that I’d like to be.
This is part of the itinerary design problem. Good itinerary designers ask questions that can at times be awkward: how’s your level of fitness? Have you done a lot of high altitude hiking before? Does your daughter’s boyfriend actually know how to ride a horse? Are there any in the group that prefer apres-ski over actual skiing? The answers to these questions inform trip pacing, when and where to rest, and what risks might be presented based on the group’s capabilities. In fact, one of the most common occurrences I’ve seen is that AI models tend to operate at an unreasonably fast pace, skipping rest days entirely in lieu of doing or seeing as much as possible.
Which brings me to duty of care. I’ll concede that duty of care is, as a topic, about as exciting as a debate over photocopier models or drinks with the HR department at an insurance firm, but when something goes wrong on an AI-modified itinerary, it will not be OpenAI or Anthropic that carries the liability. It’ll be the operator, the lodge, or the DMC that ‘should have known better’.
Sitting at a desk, three months ahead of departure, itinerary designs and agents are a line of defence between an aspirational client (whose AI is telling them at their weekly 20 laps in the local pool will make Kilimanjaro easily manageable) and a bad outcome whose responsibility doesn’t transfer to the model when the client decides they prefer its plan to yours.
AI-edited itineraries are going to become more common, not less, so you’ll need to push back gently, name what you’re doing, and make clear that the consideration is for the client‘s well-being.
What to actually say
Principles are only useful if they convert into language you can use regularly.. A few responses, adapted from conversations I’ve watched land well:
