Back

July 2026 AI Insights

There’s a conversation I keep having, in different rooms, with different companies, roughly once a month. Should we do this one first, or that one? Do we go here, or over there? We just had a reorg. We need to hit these numbers. Good people, real budget, real intent. Thirty days later, same questions.

In those same thirty days, Sam Altman said we are in the singularity, this is the moment. His point is that we’re too early inside it to see it, but the takeoff has already started.

I mostly agree with him. Which makes that monthly conversation harder to sit in, not easier.

July was more of the same. New models, more capability, a few new features. The drumbeat continued. That’s why it’s worth writing about. The compounding is the story now, and this month it pointed somewhere specific.

There is a bigger version of this argument in our latest AI campaign, The Seat AI Cannot Take.

The shift isn’t that humans are protected from AI. It’s that judgment becomes more important as intelligence becomes cheaper. The work changes, but the need for leadership does not.

The demo stopped being “write this email”

Two features shipped days apart, from the two labs that rarely agree on anything.

Anthropic added Record a Skill to Claude. You hit a button, screen-record yourself doing a piece of work, and talk through your thinking while you do it. Claude turns that into a saved skill it can run again. No prompt engineering. No script.

OpenAI came at it from the other side. ChatGPT Work takes a goal and runs multi-hour jobs across Drive, Slack and Salesforce. Last week it learned to log into password-protected sites and keep going.

Both labs are betting the unit of automation is a whole process, not a task.

Most companies are still scoping AI one task at a time. Map a step, count the minutes, build the case. That habit is two years old, and it’s how you end up with a pile of things too small to move a number anyone cares about.

What actually gets captured is the judgment

The part almost nobody noticed is that the recording captures you explaining why. Not the clicks. The reasoning.

Automation has always been hard in big companies for one reason. The person who knows the work can’t write the instructions, and the person who can write instructions doesn’t know the work. Demonstration closes that gap. What lived in a senior person’s head becomes something you can hold.

Think about why coding raced ahead. GitHub. Millions of worked examples, public, well documented, with the reasoning attached. The models had somewhere to learn. Knowledge work has no GitHub. Your firm’s judgment was never written down, which is exactly why the models can’t do it yet, and exactly why it’s the scarce thing.

Most leaders assume their edge is how the work gets done today. It usually isn’t, and most of it is inefficient anyway. The edge is the decisions, the taste and the direction your best people add on top. Write that down in a form a machine can carry, and the business changes shape.

I heard versions of this all month. One manufacturer wants its own engineering agent rather than its software vendor’s, with decades of design judgment inside it. A bank described its two-year goal as moving people off building spreadsheets and slides and onto judgment. Inside our own firm, we took one person’s process for pricing, turned it into a tool, and got three days of a good manager’s work in about five minutes.

So here’s the payoff. When people stop moving data between systems and writing reports, and start interpreting, directing and deciding at the speed the machine moves, you have a different business. Not a cheaper one. A different one.

The data panic, and the answer nobody’s talking about

Every time I make that point, someone pushes back. Careful, you’re handing your context to OpenAI and Anthropic, and they’ll build the next best version of you. Satya Nadella gave the argument a name this month.

I think it’s overblown. Not baseless, but overblown.

I don’t believe the labs are stealing enterprise data. I do think they’re mining what they see to work out what to build next. And they’re pulling companies in a single industry together, healthcare loudest of all, to pool data and build models for that sector. Those models then get sold to everyone, your competitors included. The terms and conditions say what they say, and proving a breach would cost you more than you’d get back.

The reason it’s overblown is that the other half of the answer landed this month. Open models caught the frontier. Kimi K3 shipped as the first open-weight frontier model, and the gap to the closed labs is now about ten days. Xi made open source China’s official AI doctrine. Open stopped being the cheap option.

So the answer is a hybrid, and I’d start building it this quarter.

Use frontier models for frontier work. Hard strategic thinking, genuinely complex reasoning and orchestrating everything else. That’s what the premium buys. Then run open models you control for the rest, tuned on your own information, at a fraction of the cost. Cursor found a 15x swing in cost from routing alone, and Microsoft is swapping frontier models for in-house ones at up to 89% less.

Two things follow that most people haven’t connected.

You need a team you don’t have today. Someone has to choose, tune, host, route and monitor those models. That’s real headcount, and a useful reminder that AI is currently creating more jobs than it’s removing. Tech CEOs have dropped the replacement story they were telling last year. One study found the heaviest AI investors grew headcount 10.2% in the two years after adopting it, with entry-level roles up 12%.

And if you’re regulated, this month was a gift. I work with banks and insurers stuck for two years because the data can’t leave the building and the regulator wants explainability. Open models at frontier quality mean they can keep that control and finally move. They’ll pay for it in expertise instead of licenses. For most of them, that’s a good trade.

What it looks like when the CEO decides

Against all that, one story from our own work this month.

It started as an executive briefing. Ninety minutes, no commitment. The CEO left the room having decided something, which is rarer than it should be.

That turned into a six-week sprint to reimagine two parts of the business from a blank sheet. Not automate them. Reimagine them. Those two are now heading for live AI-native platforms in under ten weeks. On the back of it, the team has a remit to do the same across the rest of the company and find the next set of bets worth making, the ones with 10x to 100x in them rather than 10%.

A few months from a briefing to a working platform. Nothing about that required new technology. It required a CEO who understood what was possible and was willing to make a call, and a team that was ready to operate differently.

I love stories like these!

Which brings me back to that room

If judgment is the asset, a company that can’t make a decision doesn’t have one to capture.

Across my client conversations this month, the problem is almost never ambition. It’s that the same foundational questions keep getting re-asked instead of answered. What do we do first? Who owns the number? What do we mean by operating model? Nobody refuses to decide. The decision just never gets surfaced, routed and closed inside the window the market gives you.

One more thing I’ve noticed since about April. The first half of the year had real energy behind big swings and moonshots. Lately I’ve felt a flip back to “let’s just pick one small use case and get started.” I can’t tell whether people got overwhelmed or whether it’s the wider economy. Either way, small things cost about the same effort as big ones, and the worst outcome is a pile of small things that don’t land, followed by everyone deciding AI doesn’t work.

What I’m thinking about

Two years ago the question was whether to take AI seriously. Last year, where to start. In April, what it takes to redesign the work. In May, whether you could describe the AI-native version of your business. In June, whether you could stop planning the whole program and build one prototype.

This month I’d ask something smaller and harder. Can you write down how your best people actually decide?

That’s now a thing a machine can hold and repeat, and it’s the only part of your business not heading toward commodity. Most companies have spent two years protecting data they never wrote down and automating work they never valued.

If you want help finding the first decision worth capturing, hit reply. Tell me the one you’ve re-evaluated three months running, and I’ll tell you how I’d force it.

Adam

Unlock monthly insights

Subscribe to the newsletter