Back

June 2026 AI Insights

I had some version of the same conversation a dozen times this month, with companies that have nothing in common, and it wasn’t the one I expected to be having.

We get in a room and it pours out. They need governance. A strategy. A new operating model. A way to choose tools. A rollout plan. Training. A way to measure any of it. Every one is real. Then they add it up and land in the same place: “we don’t have the bandwidth for this right now.”

So they do one of two things. They wait. Or they carve off a slice small enough to feel safe, so small it can’t move a number anyone cares about, and when it doesn’t, the whole program loses the right to exist.

That’s the trap I kept seeing all month. Capability now compounds faster than most companies can make a decision, and once that gap gets wide enough, they stop making decisions at all.

And in the same thirty days, I watched a handful of organisations do the exact opposite.

Rebuild beats retrofit

Shout-out to the team at Allianz UK, who did the thing most incumbents still tell me is impossible.

They launched Slick Cover, a brand-new digital motor business, and went from an idea to selling the first live policy in under six months. Their own CEO called it “genuinely remarkable,” and he’s not wrong.

Set the technology aside and look at what happened. A global insurer didn’t just run a workshop about the future of insurance. They stood up a real business, in market, taking real premiums, in two quarters. That’s the early data point I’ve been waiting a year for. You can now rebuild a piece of an organisation from scratch and have it generating revenue before your annual planning cycle would have finished debating whether to try.

Most people hear a story like that as “go build a whole new insurer,” and think it doesn’t apply to them. What I want you to take away is that you don’t have to reinvent the company, you rebuild one function, or stand up one new line, designed from the ground up around what’s now possible.

That’s some of what we started in June. One insurer working out how to turn decades of underwriting judgment, claims history and data into an advantage a digital-native competitor can’t copy. An investment manager exploring what AI-native wealth management might look like. A physical services company asking what happens when robotics gets good, and whether they’d rather build that future or have it built for them. Different industries, same move: find the place where your existing advantage meets a newly-possible capability, and come at it the way a disruptor would.

Quick plug, if you haven’t been out to San Francisco lately, it’s the single best place to see and experience this happening. We’d be honoured to host you and your team.

The rest of the market is starting to shift too. Morgan Stanley opened its own platforms to client AI agents this month, and Notion moved to selling “work, not software”. The incumbents who get it have stopped asking how to add AI to what they already sell, and started asking what they’d build if they were the ones trying to put themselves out of business.

You don’t need the whole program, you need a prototype

There’s a way through the overwhelm, and almost nobody is taking it.

The models got good enough that the thing you used to scope as a three-month custom build, you can now prototype in a few days.

With Claude or ChatGPT, one motivated person, or a small team, can stand up a working version of a redesigned workflow in hours or days. Not a slide about it, a thing that runs. The time saved is the smaller half of the story. What you really get is a look at what the reimagined function does before you commit a budget to it. A prototype turns an unknowable program into a decision you can actually make.

Most organisations are underusing this by a mile, still running procurement cycles for something their own people could rough out in a week and learn ten times more from.

The labs are betting everything on it. OpenAI now says Codex generates 99.8% of its own weekly output tokens, and Anthropic has put its own recursive self-improvement on the record. The agents are moving into the tools where work already lives, from Claude as a teammate you @mention in Slack to Codex folded into ChatGPT for close to a billion people. The build-it-yourself speed that used to belong to engineering teams is now in reach of anyone willing to try.

How to run it, and how to pay for it

Two things we’re doing that are working right now.

The first is people. We’ve started dropping small AI pods into client teams, lean groups that do training and building at the same time. Most transformation programs run enablement first and results later, and the space between them is where the momentum dies. Do both at once and the team learns by shipping. You get the capability lift and the return in the same motion, which also happens to be the only version finance keeps funding.

The second is cost, and this is where the month’s uglier headlines point somewhere useful. Flat-rate AI pricing broke this month as real consumption ran past what anyone priced for, and OpenAI’s leaked numbers showed $34B spent against a $38.5B loss. Running everything on frontier tools at frontier prices does not survive contact with a CFO.

The pattern that does work looks like a factory with two floors. Upstairs, your power users prototype on the best off-the-shelf tools, Claude and ChatGPT, fast and messy, proving what’s worth doing. Downstairs, the things that prove out get rebuilt on infrastructure you own, cheaper models you control, with the monitoring and evaluation to drive performance. Open models crossed into serious production use this month, which makes that second floor real in a way it wasn’t a year ago. The frontier is your R&D lab. Your own stack is the factory floor. Confusing the two is why the bills look insane.

The slow part was always the people

One more pattern…

A lot of companies that cut jobs on the promise of AI spent this month admitting it’s taking longer than they said. The clean “AI replaced them” story got a lot messier, and at least one economist argued there’s still zero hard evidence AI is doing the cutting. Meanwhile Oracle cut around 21,000 roles and bet the savings on AI infrastructure, so the pressure is real even where the results lag.

What do I think? it takes longer because the hardest part is human. Getting a team to operate in genuinely new ways, and helping them find those ways in the first place, runs on human time, not compute time.

And I’ll be honest, the people part is the piece I’m least sure how to speed up. You can prototype a new workflow in an afternoon. You cannot re-wire how a team sees its own work in an afternoon, and I still haven’t found the shortcut. Anyone who tells you they have is selling something.

Which is the same lesson as everything above. The giant program overwhelms you because it’s mostly people work, not technology work. You prototype because it gives people something real to change toward instead of a memo. You embed the pod because capability and behaviour have to move together or neither one sticks.

What I’m thinking about

Two years ago the question was whether to take AI seriously. Last year it was where to start. In April I said it had become what it takes to redesign the work. In May, whether you could even describe the AI-native version of your business.

This month the honest update is harder. You can probably describe it now. You can see the whole program you’d need to build to get there. And that clarity is exactly what’s freezing you in place.

The organisations breaking through don’t have a bigger plan. They built one working prototype of one reimagined function, learned from it, and had the nerve to take it all the way, the way Allianz just did.

If you’re staring at the size of it and can’t find the first move, that’s the conversation we’re built for. We put the AI OS and a pod on that exact problem, turning the overwhelming everything into the one thing worth doing first. Hit reply and tell me where you’re stuck, and I’ll tell you what I’d prototype.

Adam

Unlock monthly insights

Subscribe to the newsletter