Field notes from the AI weather system.

The last few weeks have been busy ones for AI conversations — inside Travelopia and outside it. Board prep, a podcast recording, exec presentations, brand-by-brand pushbacks, model launches, policy updates, hackathon readouts, and the usual quiet drumbeat of teams just getting on with the work. Sitting with all of that, a handful of themes have crystallised more sharply than they had before. Not new ideas exactly — sharper, more useful versions of things Sree and I have been turning over for a while.

Here are seven of them.

1. The story is finally simpler than the slides

We've landed on a sentence I genuinely like: Travelopia Tech exists to make the group AI-enabled. Choose the right tools. Work in new ways. Prove it through pilots. Scale it safely. Share what works. Small bets. Fast learning. Safely scaled.

It takes about ten weeks to write that and ten seconds to read it — roughly the right ratio. Every conversation gets easier once the story collapses to something this short. Decks shrink. Meetings move faster. The "but what are you doing about AI?" question gets a one-line answer.

2. Adoption is not activity. Activity is not value.

Eight weeks of Copilot data across one of our brands made this uncomfortably clear. A small group of power users drove most of the usage. A middle band dabbled. A frankly humbling number of licensed users did nothing at all.

Handing someone a licence is not adoption. Opening the tool once is not adoption. Sustained use that changes how they work — that is adoption. Everything else is an expensive browser tab.

Handing someone a licence is not adoption. Sustained use that changes how they work — that is adoption. Everything else is an expensive browser tab.

It's pushed us to think harder about deprovisioning, and about an old-fashioned word: enablement. The bottleneck is almost never the technology.

3. Governance is shifting from "no" to "yes, in approved tools"

We updated the GenAI policy in May. Practical effect: the door has opened from "no by default" to "yes, in the right place." Not in everything. Not anywhere. But meaningfully more room to work, in the tools we trust.

I cannot overstate how much energy this unlocks. The hardest thing about AI inside a real organisation isn't the model — it's the eighteen-month tug-of-war between people who want to experiment and people paid to say no. When governance catches up with reality, things start moving.

4. Two new forums, one bet on people

We launched two things in May. An AI Power User Drop-In — a single-window-clearance forum for getting early experiments green-lit. And an AI Super User Forum — a 30-minute weekly show-and-tell across brands and functions.

The tagline for the second one was no polish required — work in progress is the point. That's the opposite of how most corporate AI happens. The early sessions have already produced more useful sharing than most steering committees I've sat in.

5. Agents are now part of the conversation

The vocabulary is settling. The use cases we keep seeing cluster into three: a companion that optimises the work you're already doing; a researcher that goes off and finds, summarises, synthesises; and a coder — or more broadly an autonomic — that actually executes.

The question has shifted from "can it draft this?" to "can it run this?" That's a different question with different answers, and it changes who you talk to first when you scope a piece of work.

6. The pace of the models is its own management problem

Opus 4.8 dropped in late May — 41 days after 4.7. Better at coding, cheaper at fast mode, more honest about its own mistakes. And then there was Fable — available for about a day before it was pulled.

I've started thinking about model releases the way you think about weather. You don't plan around the specific storm. You build something that copes with whatever blows in. Translation: pick tools that get better, not just tools that are good now. The encouraging part is that, for the most part, this is now happening for free, on the same model string, while you sleep.

7. The big-programme reflex is the enemy

The pattern across brands: someone gets excited about AI, builds a beautiful 12-month transformation plan with a heroic efficiency target, then waits six months for the budget cycle.

My answer, with affection, is almost always the same. Start in two weeks, not Q4. Pick the smallest possible slice. Anchor it to a real metric you can baseline today. Decide what to do next based on what you actually find.

It's amazing how unpopular this advice is. And how often it turns out to be right.

Where this leaves me

AI inside a real organisation looks nothing like AI in a keynote. It looks like small conversations. Unglamorous infrastructure decisions. Policy clauses. A lot of patient explaining that yes, it really is this easy to get started, and no, you do not need a steering committee to do it.

When the small things go well, the big things take care of themselves. Sree and I will keep writing about what we learn.