People ask me all the time where AI is headed, so I’m putting a date on it. It’s October 2026 as I write this. This is what I think AI looks like in September 2027, and you can hold me to it.
The big change won’t be smarter chat. We’ll stop chatting with AI and start delegating to it. Today most people still type a question, read the answer, and type another one. By next fall, handing an agent a task and walking away for a few hours, or a few days, will be normal for knowledge work. It’ll do the research, check its own work, and hand back something you can actually use.
The economics will split too. Mid-tier models will be close to free and good enough for most of what we do every day. Frontier models will go the other way. They’ll cost more, compute for them will be harder to get, and governments will have more say over who gets access to what. That’s already starting.
How we interact with AI changes as well. Voice and assistants that can see your screen will be the default for most people. And prompting, the skill everyone rushed to learn, will matter a lot less. What will matter is knowing what to hand off, and being able to judge whether what comes back is any good.
I could be wrong in two ways.
First, capability keeps climbing but reliability doesn’t. Then 2027 looks like today with better demos. An agent that’s right 90% of the time sounds impressive until you realize you still can’t trust it alone with anything important. Ninety percent means you’re checking everything, so you haven’t really delegated anything.
Second, and this is the one I’d put money on, the technology won’t be what holds us back at all. We will. Next September most companies will still have a small group doing impressive work with AI and a big majority using it as a fancier search box. Models improve every few months. Org habits move much slower.
I see this every week with Copilot rollouts. The license isn’t the hard part. The habit is.
That leads to the part I find most interesting. If agents do more of the work inside our apps, fewer people will be clicking through those apps at all. The adoption challenge doesn’t go away. It moves. It shifts from teaching people how to use the tool to teaching them how to delegate to it, supervise it, and decide when to trust it. That’s a management skill, and most organizations haven’t started building it.
So my prediction for September 2027 is simple. We’ll have more autonomy, cheaper intelligence, and the same human bottleneck we have today.
Ask me again next September. I’ll tell you how I did.