The Surprise Was Not the Adoption
Last quarter I sat with more than 30 teams. I expected to spend most of that time measuring how far along each one was with AI. That part turned out to be the least interesting thing in the room.
The surprise was who was driving it.
Not the engineers. The operators. The people in finance, in sales support, in operations, the ones who have never written a line of code and never wanted to. They were the ones with the tabs open. They were the ones who had quietly made the model part of how they got through a Tuesday.
I keep turning that over. We have spent two years assuming the technical staff would lead and everyone else would follow. What I watched was closer to the reverse.
What It Looks Like at the Desk
The pattern is consistent enough that I can describe the arc.
Someone starts nervous. They are not sure what the tool is for, or whether they are allowed to trust it. Then they get one real result, something that genuinely saved them effort, and the nervousness drops away. Within a week or two the tool is simply part of the workflow. They do not announce it. It just becomes how they work.
A lot of these people never had an assistant. They never had a chief of staff. An ordinary model, nothing custom, gives them that experience for the first time. I think that is a large part of why it takes hold. It is not the novelty. For a lot of people this is the first time they have had anything like an assistant, and that experience is already shaping what they ask for.
Their actual usage stays elementary. They are not writing elaborate instructions. But they get experimental, because those early wins lowered their resistance, and they will try almost anything as long as the setup does not confuse them.
I watched someone drop a business card into the model and ask it to write the contact into the CRM. I watched someone drop in a blog post, ask for a version focused on a different angle, then ask the model to publish the new version to the company site. Small things. Done by people who would tell you, flatly, that they are not technical.
The Wall, and the Question
Then they hit a wall. The systems are not connected. The model can draft the CRM record, but it cannot reach into the CRM. It can write the new post, but it cannot actually put it on the site.
Here is the part that stayed with me. They do not stop at the wall. They ask a different question.
"What would happen if I could just connect it?"
That is a developer's question. It is, almost word for word, the thing a developer asks. Developers go looking for APIs. They look for the connectors that plug into a terminal so they can run many actions at once, programmatically, without doing each step by hand.
Functional people are now reaching for exactly that. They are treating the chat window as a terminal. They are using natural language as the command line.
They are treating the chat window as a terminal, and using natural language as the command line. They just do not have the word for it yet.
They do not have the vocabulary for any of this yet. They could not tell you what an API is. But the instinct underneath the question is the developer's instinct, arriving in people who never trained for it.
A Tangent I Have Earned
I want to step sideways for a moment, because I have thought a version of this before, and I was wrong about the shape.
Years ago I believed a major cloud storage platform would become the terminal for everyone who does not write code. The idea was that you would speak in plain language and the platform would carry out commands across your entire cloud system. It seemed obvious to me at the time.
It did not happen. A large part of the reason is that the model required a single provider to own all of the interfaces. The real problem was the work that crosses between customers and between systems, and that problem was bigger than any one product organization could solve on its own.
So I hold this current observation loosely. I have been early and wrong about the shape of this exact thing once already. That is the tangent. Now back to it.
The Real X-Factor
What I missed then is the thing that matters now. The unlock is not any single product. It is interoperability between models.
Picture a shared place where integrations can be served through a common protocol. Anyone who builds a connector can expose it to the terminal and open a connection to it. You do not need one company to own everything. You need everyone to agree on how to plug in.
The shape rhymes with something we already lived through. It behaves like a new version of an email protocol. Email worked because nobody owned it. Any provider that spoke the protocol could talk to any other. A common protocol for integrations would do the same thing for actions instead of messages.
It behaves like a new version of an email protocol. Email worked precisely because no single company owned it.
The net effect, if it holds, is that functional employees become terminal users. They drive real systems with plain language. The connection problem that stopped them at the wall stops being their problem.
The Problems Nobody Has Solved Yet
The standard risks are real. Security. Write access into live systems. Observability into what the model actually did. These need to be worked through, and the field is working through them. I will not dwell on them, because they are the part everyone already names.
The risk I think is underdiscussed is quieter. Acceptance of AI output is high and review is low. People ship what the model gives them. And the more a person tweaks an output, the more time it costs, so they skip the tweaking. What they value is the speed of the terminal, not the minutes they spend inside it. That is a reasonable thing to want. It is also where things slip out the door unchecked.
Which opens a question I do not have a settled answer to. How do people actually want to talk to their systems?
I see two candidate shapes. One looks like messaging. You address a specific system in a channel, the way you would message a coworker, and ask it to do a specific thing. The other looks like a command, where a company keeps configuration profiles that point each kind of request at the right tool for the job.
Either way, companies will probably want guardrails on the prompting itself. A way to refine a request, and a way to deliberately slow the fast path down when slowing it down is the right call.
The analogy I keep coming back to is the way you manage salespeople. You want them out hunting leads, moving fast, talking to everyone. You also put gates in front of them, so the data that comes back into the company stays clean. The same logic applies to prompts. Call it internal prompt refinement, done on the company's side.
That points to a role inside the loop that resembles a product manager. When someone asks the company's system to do something, the system prompts back. It helps refine and standardize the request, so what gets executed lines up with the business's priorities rather than one person's hurry.
Three Cohorts and a Missing Layer
Across those 30 teams, I saw three cohorts.
The first has not really adopted AI yet. The second rolled it out free to everyone, with no standardization, hit token abuse and runaway cost, and switched it back off. The third is the more advanced hybrid. There, skills and workflows route to a central office of excellence, and humans approve agents and skills before anything ships.
Here is what struck me. Even the most advanced cohort seems to be missing the same thing. A harness that sits on top. Something that helps configure the business, and lets users filter and get results that align with company priorities, programmatically, without each person reinventing the approach.
It feels like that layer is coming. And whatever interface delivers it at the enterprise level matters a great deal, because in the end it governs a model reasoning over the company's own data. I will not tell you who builds it or when. I do not know, and the last time I was sure about a shape like this I was wrong.
So I will leave you with the thing I actually think is worth watching. It is not the tooling. It is the question.
When your operators start asking how to connect their tools, instead of what a tool can do, the shift has already started. The vocabulary will come later. The instinct is already here.
Are you hearing it yet?