Copilot Is a Terminal

01What the 68% rules out

Nearly every published account of Copilot’s commercial trouble is an adoption story. People need training. Prompting is hard. Change management was underfunded. Governance delayed the rollout. Users do not know what it can do. Microsoft has spent three years and a great deal of money answering exactly those objections — adoption kits, a learning centre, a daily prompt guide, an entire partner motion built on enablement.

The 68% says none of that is the binding constraint. Hand the same worker the same product in the same apps with nowhere else to go and two in three make it their primary tool. They find it, they learn it, they use it. Then put a second option on the desk and it goes to 18%. A third and it goes to 8%.

Nothing about the product changed between those conditions. The training did not get worse; the integration did not get shallower. Copilot runs on the same family of OpenAI models that ChatGPT does, which is what makes this unusually clean — raw capability is roughly held constant. The only variable is whether the worker had somewhere else to go.

Figure 01

Primary tool chosen, by what was available

• Employer offers Copilot only
  • Copilot68%
• Employer offers Copilot and ChatGPT
  • Copilot18%
  • ChatGPT76%
• Employer offers all three
  • Copilot8%
  • ChatGPT70%
  • Gemini18%

02The last time this vote was held

Corporate America has run this experiment before, with the same incumbent in the room.

Through the 1970s the way an office worker reached computing was a terminal: a screen and a keyboard, wired to a machine elsewhere, with no processor of its own worth the name. It displayed what the mainframe sent and relayed what you typed. It was, by construction, perfectly integrated with the central system — that is what a terminal is — and it could do precisely nothing on its own.

Then IBM shipped the model 5150 on 12 August 1981, built from off-the-shelf parts around a non-exclusive licence to Microsoft’s operating system, and the market did something nobody in Armonk had planned for.

It had. The formal corporate response of the period was the “information centre,” a function invented to re-absorb computing that departments had started doing without asking. Rockart and Flannery, surveying around 250 end users across seven large organisations for Communications of the ACM in 1983, found end-user computing growing at 50 to 90 percent a year. The machines were not winning a procurement bake-off. They were arriving around the side of one.

And they were worse. A PC in 1984 was slower, less reliable, less secure and less capable than the mainframe down the hall, and it held none of the corporate data. It won anyway. By the US Census Bureau’s own series, the share of employed American adults using a computer at work went from 24.6% in 1984 to 36.8% in 1989 to 49.8% in 1997. IBM’s share of the market it had created fell from roughly 80 percent in 1982–83 to about 20 percent a decade later, on the historian James Cortada’s accounting; Compaq passed it in the US in 1994; IBM sold the business to Lenovo in a deal announced 8 December 2004 and closed the following May.

03What actually died

Here is the part that is usually told wrong, and getting it right is what makes the analogy load-bearing rather than decorative.

The mainframe did not lose. IBM’s Infrastructure segment turned over $15.7 billion in 2025, up 12.1%, with IBM Z revenue up 51.7% — its strongest year in roughly two decades. The 3270 data stream is a current, documented, supported protocol; TN3270 clients and the CICS Transaction Gateway are products you can buy this afternoon. Vast amounts of the world’s transaction processing still run exactly where they ran in 1979.

What died was the terminal as an object on a desk. The compute stayed central. The endpoint changed — and what it gained was the ability to do something while nobody was sitting at it.

The vote was never centralised against local. It was whether the thing in front of you could act when you were not there.

Which reframes the whole enterprise AI consensus. The consensus says the winner will be whoever is already inside the workflow, because the model will commoditise and the context will not. Copilot has all of that context: it is in Word at the cursor, in Outlook above the reply, holding your SharePoint, your mail, your calendar, your org chart, under governance, already on the invoice. Against it, ChatGPT offers a blank box in a separate tab with no access to anything unless you paste it in, and a second bill. Seventy percent.

A terminal was maximally integrated too. Integration was never the moat. Integration is what a thing has instead of independence.

04Why proximity costs rather than pays

The mechanism, in one paragraph. An assistant that lives inside the work must be supervised inside the work. It produces at your cursor, mid-sentence, which means verification happens there too — continuously, interleaved with the thinking it was supposed to relieve. You cannot batch it or defer it. Every suggestion is an interruption to be adjudicated on the spot, and adjudicating is the expensive part, because you do not know where a machine’s errors live and so must check everywhere.

An assistant across a boundary is a different transaction. You leave, ask, receive a finished object, and make one decision: use, discard, ask again. The boundary converts a thousand in-line judgements into a single judgement at a seam you chose, and it puts the output somewhere that costs nothing to reject, because it is not already sitting in your document under your name.

The corroborating numbers point the same way. Copilot’s accuracy net promoter score stayed negative across the whole measurement window — −3.5 in July, −24.1 in September, −19.8 in January — and 44.2% of workers who tried it and stopped cited distrust of the answers, the highest of the three platforms. Distrust is what accumulates when you must check a thing in place, every time, inside your own work. The Australian government trial found the same texture from the other side: participants said they would not use Copilot for important documents, and one department’s internal survey found 60% had to make moderate-to-significant edits.

05Autopilot has the PC’s shape and the terminal’s deed

So read the June announcement again. An Autopilot is an always-on agent with its own identity that acts on your behalf without being prompted each time, so that work continues while your attention is elsewhere. That is usually described as more autonomy. It is better described as the endpoint finally getting a processor: the thing can now do something while nobody is sitting at it, which is the precise property that beat the terminal.

On this reading Microsoft has read its own data correctly, and Autopilot is the right product. But the analogy stops one step short of where it is usually taken, and the gap is the whole point.

Figure 02

Three endpoints, by what the person at the desk controls

	Terminal, 1979	PC, 1984	Autopilot, 2026
Acts while you are away	No	Yes	Yes — that is the pitch
Where the compute is	Central	On the desk	Central
Who bought it	The DP department	Often the department, around the queue	Central IT, on an enterprise agreement
Who issues its identity	The mainframe operator	Nobody — it is a thing you own	The corporate directory
Who can switch it off	The operator	The person at the desk	An administrator, by policy
Can it leave with you	No	Yes — it is in the box	No

The PC’s decisive property was not that it computed locally. It was that it escaped. It was bought outside the queue, owned by the person using it, unplugged by the person using it, and it went home in a box. Every one of those is a fact about control, not about silicon.

Microsoft’s Autopilot has the independence and none of the escape. Its identity is issued by the corporate directory. Its deployment is gated by policy. It runs on central compute, metered centrally, and an administrator can revoke it on a Tuesday afternoon without asking the person whose week it has been arranging. That combination — acts on its own, owned by someone else — is not the PC. It is the thing the industry has been trying to build ever since the PC got away: Larry Ellison’s Network Computer in 1996, Citrix and terminal services, VDI, the cloud-managed device. Those attempts have a record. Thin client shipments have been contracting for years and run around four million units annually, against roughly 270 million PCs shipped worldwide in 2025. After thirty years, the re-centralised endpoint is a rounding error.

I am not predicting Autopilot fails. The independence is real and it is what people voted for, nine to one, when the only alternative was a terminal in the ribbon. I am saying the vote that decided the last one was not about independence. It was about the deed, and on the deed this is a terminal that can type by itself.

Which lands where the previous essay in this series left off. A PC had a power switch under the hand of the person using it — the same property aviation later wrote into regulation for autopilots, that the exit be reachable without letting go. An agent provisioned centrally has an off switch too. It is just in a different building, held by a different person, on a different clock. That is not nothing. It is also not yours.

06Where this stops being true

Already true

The 68/18/8 gradient, the 76% and 70% ChatGPT figures, the accuracy NPS series and the 44.2% distrust figure are from Recon Analytics’ US AI Survey, 150,000+ respondents, July 2025 to January 2026, published 4 February 2026; Recon also states Copilot accesses the same OpenAI models as ChatGPT. The IBM PC shipped 12 August 1981 with a non-exclusive DOS licence. Keen and Woodman’s HBR article is September 1984; Rockart and Flannery’s survey is CACM 26:10, October 1983. The Census Bureau computer-use-at-work series (24.6% / 36.8% / 45.8% / 49.8%, 1984–1997) is from CPR P20-522. The IBM share range is Cortada’s, in IEEE Spectrum, 2021. The Lenovo sale was announced 8 December 2004 and closed 1 May 2005. IBM’s 2025 Infrastructure revenue of $15.7 billion, up 12.1%, with IBM Z up 51.7%, is from its own filings. Thin client and PC shipment figures are IDC and Gartner respectively.

What has to happen for the argument to hold

The choice gradient has to reproduce outside Recon’s sample — ideally in telemetry rather than self-report, and outside the United States. And the deed has to turn out to matter: if Autopilot-class agents convert well despite being centrally provisioned and centrally revocable, then ownership was never what decided the PC era either, and section 05 is a nice story about a coincidence.

Where the shorthand breaks

Self-reported “primary tool” is not usage, and it forces one answer from people who plainly use several. “Same model” is a simplification doing real work here — Copilot’s retrieval, grounding and safety layers differ from ChatGPT’s. The terminal-to-PC transition also had a large and unglamorous cost component: by the mid-1980s it was frequently cheaper to run a job on a small machine than on the mainframe, and an honest account concedes that autonomy was not the only variable. There is no defensible figure for the peak installed base of text terminals or for the date PCs overtook them as the corporate endpoint, so I have not given one. And I have deliberately left out the widely quoted mainframe statistics about transaction share and lines of COBOL, all of which trace to vendor marketing without published methodology.

Where I am probably wrong

The simplest rival explanation for the 8% is habit. People used ChatGPT at home for two years before Copilot existed, arrive at work already fluent in one tool, and pick the one they know. If that is what is happening, the number measures familiarity rather than topology, Microsoft’s problem is timing rather than architecture, and the terminal analogy is an elegant frame laid over an ordinary distribution race. I lean against it because habit should decay and this did not — Copilot’s share fell across the seven months rather than recovering as the integration deepened. But that is an inference from a trend line, not a test. The second risk was that I am overreading the deed — almost nobody owns their work laptop, and people do not appear to mind. I no longer think that survives, and the follow-on essay, My Employer Is a Node, says why: the endpoint that carries most of the working day is the phone, the phone is bought by the employee, and Apple’s and Google’s own documentation puts a removal control on the user’s side of it. The laptop is where provisioning won. It is not where the argument is.