Four frontier AI models launched in three weeks this June. Prices kept falling. Capability kept climbing. And for most businesses, none of it changed what happened on Monday morning.
The story wasn't the models. It was how slowly most organisations were putting them to use. A few things stood out this month.
The bottleneck moved from the lab to the office
Capability stopped being the constraint a while back, and June made that plain. With prices falling and performance rising across every major lab, the hard problem shifted from what the AI can do to how quickly people can absorb it.
There's a pattern showing up in a lot of businesses right now: the conscientious objector. Capable, thoughtful team members with real reservations about AI, about quality, about what it means for their craft. The instinct in a lot of organisations is to mandate past them. The better move is to sit with what actually worries them. Adoption runs on honesty, not compliance. The teams making real progress aren't the ones with the best tools or the biggest budgets. They're the ones treating adoption as a culture change project: one team, one real use case, measured properly, with the results doing the persuading.
Judgment became the whole point
Three separate stories this month landed on the same idea from different directions.
Marco Argenti, writing in Harvard Business Review, described an experienced banker asking him what ten percent of his job AI would never touch. His answer: let that ten percent go, and be ready to be rebuilt around a new hundred. Judgment, instinct and values stay. Step by step execution doesn't. He compares it to a horse rider learning to drive. The reflexes carry over. The habits don't.
Adobe's Forest Key made a similar point about creative work. His filter for what to automate: is this task about execution, or about intent? Repetitive production and established workflows, AI handles well. Taste and direction still need a person in the loop. The scarce skill going forward, in his view, is the ability to look at a set of outputs, know which one is right, and say why. Swap "creative direction" for "strategic direction" and it describes every leader working out what to keep for themselves as AI takes on more of the execution.
LinkedIn's hiring data backs this up from another angle. Hiring is down across the board, but the roles still open are asking for a different mix. Technical fluency with AI matters. Judgment, adaptability and communication matter more. The people staying in the room combine domain expertise with the ability to work alongside increasingly capable machines.
Measuring the wrong thing breaks things
Amazon built an internal leaderboard ranking employees by how much they used the company's AI developer platform. Employees gamed it, building pointless agents that burned through tokens just to climb the rankings. Amazon shut it down and replaced it with a measure of whether the AI-written code was actually useful.
The lesson travels well beyond Amazon. Rank people on usage and they optimise for usage, not outcomes. That's a design failure in the incentive, not a failure of the people responding to it. The wider industry landed on the same point this month: token prices have fallen 98 percent since 2022, yet enterprise AI bills have tripled, and Gartner found only 28 percent of AI infrastructure projects fully deliver against their business case. The question worth asking isn't how much AI a team is using. It's what has actually changed because of it.
A simple way to test that inside your own team: can people name three decisions in their role that require judgment AI can't handle? Can they delegate routine work without overthinking it? Are they comfortable with ambiguity in AI outputs, or do they default to deferring to it? Are they asking better questions because they have more time to think? Build goals around answers like these rather than usage numbers, and adoption becomes something you can actually measure.
The frontier isn't always the answer
Anthropic's Fable 5 landed this month as genuinely the most capable model deployed publicly. It also exposed something else: there's now roughly a thousand times price difference between the cheapest capable AI and the most expensive frontier model. Two distinct markets have formed, and plenty of organisations are paying frontier prices for problems a mid-tier model would solve just as well.
Worth twenty minutes with a team: map current AI use against what genuinely needs judgment versus what's routine execution. Most businesses haven't answered honestly which of their tasks need the most capable model money can buy, and which would work just as well on something far cheaper.
Agents need trust before they get access
Anthropic launched Claude Tag this month, embedding Claude into Slack as a digital teammate that can be assigned a task, work through it in stages, and report back. Separately, 1Password and Cursor are both building oversight architecture for AI agents: continuous monitoring, approval gates rather than blanket blocks, and credentials kept away from the agent itself.
The "teammate" framing is useful and risky in equal measure. Useful, because teams genuinely need to think of AI as something to delegate to. Risky, because it can soften a real governance question: who decides what the agent is allowed to do, and what happens when it makes a call that costs money. The organisations getting this right ask those questions before turning the feature on, not after.
What June actually asked of leaders
None of this month's stories were really about the technology. They were about whether the people around the technology were ready: ready to let old habits go, ready to measure the right things, ready to decide what judgment they're keeping for themselves.
If any of this raises a question worth thinking through inside your own organisation, hum[ai]n would be glad to talk it through.