You banned ChatGPT and gave everyone Copilot. They switched to their phones.

Illustration: AI-generated, art-directed by the author.

In February 2023, JPMorgan restricted employee ChatGPT use. Goldman and Citi followed within weeks, then Apple, Verizon and Amazon. By that August, a survey of two thousand IT leaders found 75% of organizations were banning, or considering banning, ChatGPT and tools like it on work devices. The drawbridge went up across most of the economy inside six months.

The fear behind the lockdown was not irrational. Security firms that watch what employees paste into these tools measured a real, non-trivial slice of it as confidential company data. The exposure was genuine, and leadership was right to worry about it. Samsung's leak that May hardened an existing trend rather than starting it, and it happened in the window after Samsung lifted its own ChatGPT ban, while the usage was sanctioned and visible, which is why it argues against banning rather than for it. I made that case in shadow AI.

This piece is about what companies bought instead.

The fear was legitimate. Banning the tool instead of governing the use is where the money started to leak.

The safe default nobody chose

Here is where the shape of the whole problem gets set. Almost every large company already lives inside Microsoft: Windows, Office, Teams, the entire tenant sitting in Microsoft's cloud. So when the mandate came down to "give people AI, but safely," there was an answer sitting right there in a contract they already signed. Microsoft 365 Copilot.

Copilot's pitch to a nervous enterprise is genuinely strong, and notice that it is not a pitch about being the best assistant. It is that your data stays in your tenant, is not used to train the model, and inherits the compliance controls you already run. It is procurement-easy, security-blessed, and one more line on a bill you already pay. For a CIO who just watched the Samsung story make the rounds, that is the safe choice, and safe was exactly what got asked for.

And so it deployed, at scale. Microsoft says nearly 70% of the Fortune 500 now use Microsoft 365 Copilot, part of more than 100 million monthly active users across its AI products. On a board slide that reads like a win: we deployed AI across the company.

Deployment is not adoption

Then you measure, and the win evaporates. In Gartner's 2025 survey of 187 IT and security leaders, about 40% of all respondents had Copilot in pilot. Of the organizations that had actually finished a pilot, 5% said they were moving to a larger deployment that year. Those two numbers do not share a denominator, and the second is a stated intention rather than a completed rollout. Even read generously, almost nobody has crossed from trying to running, and the survey was reported as finding the return on investment still unproven.

The seat math is worse. Independent survey research across more than 150,000 US workers found that when people are handed Copilot at work, only about 36% actually adopt it. Where ChatGPT is the tool on offer, 83% do. Less than half the uptake, for the tool companies were most likely to have sanctioned.

Because when the choice is theirs, they are not on the fence about it. Given a free pick among the big three, 70% name ChatGPT as their primary assistant, 18% Gemini, and 8% Copilot.

This holds even where the company already paid. Bloomberg reported that Amgen bought Copilot for around 20,000 employees and, roughly a year later, watched them drift back to ChatGPT anyway. A top complaint reaching Microsoft's own AI team, according to Business Insider, was blunt: Copilot isn't as good as ChatGPT.

A license is a receipt. Usage is the product, and nobody ships you that in the box.

The workaround you didn't authorize

Blocking the good tools did not make anyone want the sanctioned one. It made them go around it. Most people who use AI at work supply their own tools and are careful not to say so; those numbers, and what employees actually paste into the tools they smuggled in, are in shadow AI. A separate survey of six thousand knowledge workers found 46% would keep using AI even if their employer banned it outright.

Read that against the ban and the trap is obvious. You are now paying for the safe tool that sits idle while the behavior you banned continues where you cannot see it, which is the shadow AI problem the ban manufactured. The procurement failure and the governance failure are the same failure, seen from two ends.

The tool was part of the problem

There is a comfortable story going around that the technology is fine and only the rollout failed. It is half true, and it is the comfortable half. The other half is blunter: people have used Copilot, held it up against ChatGPT, and decided it is not good enough to do their real work with. They made that judgment themselves, one task at a time, exercising the scarce skill the company never staffed for, from the wrong side of the rollout. That is what an 8% free-choice share and a steady drumbeat of "it isn't as good" complaints actually mean. When the tool you sanctioned is the weaker tool, no amount of rollout discipline rescues it.

But it is not only the tool, and this is where most post-mortems stop too early. Notice what never happened in this story: nobody chose an assistant because it was the best one for the work. The tools people wanted were removed by fear. The one they got was chosen by procurement, for safety, not for fit. And the single thing that turns any assistant into a business result, getting people trained to use it well, was assigned to nobody. BCG surveyed more than ten thousand employees and found only about a third feel adequately trained on AI, while those with even five hours of training are far likelier to be regular users. You cannot buy that on a per-seat license.

So the fix is not another rollout of the same tool. It is a different set of moves: govern the use instead of banning the tools people actually want, so they can safely reach for what genuinely helps; put in front of them the assistant they would pick on merit, not the one already on the invoice; train past the threshold that turns curiosity into habit; and measure weekly active use per workflow, because seats purchased has never once predicted a return.

Here is the uncomfortable question for your next leadership meeting: how many of the AI licenses your company pays for were used for real work this week? If nobody in the room knows the number, that is the number, and it is worth asking whether you bought the tool your people wanted or the one that was easy to buy.