The Case for Predictable AI
More and more travel businesses we work with now rely on AI every day. That's hardly surprising. Travel still runs on manual effort, with fragmented systems forcing people to spend much of their day on boring, repetitive administrative work.
AI can be excellent at eliminating some of these small, repetitive tasks that are too minor to justify ever making it onto an IT development roadmap.
People are now beginning to automate these tasks using AI but the biggest frustration that we hear isn't that AI doesn't work. It's that it works most or some of the time.
"We automated 80% of the task."
"It gets it right 80% of the time."
That sounds impressive until the remaining 20% forces someone to check every result manually. Suddenly you've created another process instead of removing one.
Most AI projects don't fail because the model isn't clever enough. They fail because they're not predictable enough.
Here are the four complaints we hear most often with a new client, and what to do about each.
"The same question gets different answers"
Large language models contain an element of randomness by design. Ask it to extract information from a spreadsheet and it may produce a perfect answer today—and a slightly different one tomorrow.
If the task never changes, neither should the solution.
A useful distinction here is static versus dynamic. If the input data and the task are broadly the same each run, the optimal solution is a fixed one, not a changeable one. Ask the AI to turn its approach into code, and it becomes repeatable every time. Standardise your prompts into templates or skills, lock down format requirements, and build workflows rather than ad-hoc chats. Save AI's creativity for problems that are truly dynamic and which require it.
"It can't automate the whole job"
I drive a Tesla. An electric-first car isn't simply a petrol car with a battery bolted on.
The same applies to AI. Most complex travel processes have evolved over decades, accumulating workarounds, duplicate checks and unnecessary steps. Many are so layered that nobody in the business can tell you anymore why each step exists or is done that way. Automating those processes unthinkingly simply preserves yesterday's inefficiencies.
Instead, redesign the workflow around a human and AI partnership where each does what it does best.
Stop automating processes that were built for a different era (you know, more than 18 months ago!). Start from a clean sheet and design a human-plus-AI workflow that can improve over time as the AI underneath it continues to evolve.
"It makes things up"
AI is so capable now that it's easy to forget it's predictive text on steroids. It can and will come up with things that simply aren't true, and hallucination hits hardest when you ask for facts, figures, or sources.
Use AI for drafting, transforming, and reasoning over material you supply, not as a source of truth. Base it on your own documents and tools and verify anything factual before it leaves the building. Treat it as a capable assistant that needs review, not an oracle. And don't let imperfection put you off: used the right way, it does remarkable things.
"It gives me generic, fluffy answers"
This is almost always a prompting problem. Garbage in, garbage out is as true for AI as it ever was. Yes, sometimes a vague request can be lucky and produce the perfect answer, but that approach is not predictable. If you want predictable results, give it detail: context, constraints, and examples of what good looks like. Brief it like a new junior hire, not a wise senior colleague you can hand half a sentence to.
The Goal
The pressing goal isn't to make AI more intelligent but to make it more predictable.
Creativity is valuable when the problem changes. Most operational work doesn't. That's why the biggest gains come from making AI behave consistently, integrating it into well-designed workflows, and giving people confidence in the output.
Once AI becomes predictable, it stops being a novelty and starts becoming infrastructure.
One caveat
Everything above is said in the context of making AI dependable in your day-to-day work. A fully automated solution delivering mission critical results across your business is a different animal, with far greater scope and risk. There you need to think about trust and adoption, process integration, compliance and auditability, brand and customer experience and cost control at scale. That's the subject of another post.
