Why Your Team Ignores the AI Tool You Paid For

I ran an agency's AI rollout that took two years to get half the staff using ChatGPT for anything beyond LinkedIn captions.

We went all in early, back in the o3 days. We built custom GPTs for house style: tone, formatting, structure, all spelled out in JSON so there'd be no ambiguity. The plan was simple. Give people a tool that writes and designs in the agency's voice, and adoption follows.

It didn't. The custom GPTs couldn't reliably follow their own instructions. Ask for a specific structure and you'd get something close, then something else close, then something that ignored the JSON schema entirely. The image generation GPT, built specifically to hold visual consistency and keep AI-slop out of client work, got ignored outright. People used ChatGPT to draft social posts and generate throwaway images. That's where it stopped. Average age in the building was 43. This had nothing to do with generational resistance to new tech. The tool didn't hold up its end of the deal often enough to earn trust, so people used it for the one job it couldn't mess up: a caption.

Two years later, adoption cleared 50%. The model hadn't gotten dramatically better in that window. What changed is we stopped treating the rollout as a procurement decision and started treating it as a change management problem.

Most companies buying Copilot right now are about to relearn this, faster and more expensively than we did.

The tell: employees already know what works

Recon Analytics tracked usage across companies that deployed both ChatGPT and Copilot. Given a free choice, employees pick ChatGPT 82.8% of the time. Copilot gets picked 35.8% of the time.

Take the choice away and the picture flips. When Copilot is the only tool available, 68% of employees use it. The tool isn't the obstacle. Coexistence is. Companies buy the enterprise license, leave ChatGPT reachable in a browser tab, and treat the resulting silence as a mystery.

This is what WalkMe and SAP researchers found when they surveyed the workforce directly: 78% of employees admit to using AI tools their company never approved. That's not rebellion. It's a diagnostic. When the sanctioned tool feels slower or more restricted than the alternative sitting one tab over, people use the alternative and don't mention it.

It's the same shape as what we ran into with the custom GPTs, just inverted. There, the tool that was supposed to be the answer got quietly avoided in favor of the path of least resistance. Here, the expensive tool gets avoided for the same reason. Something else does the job with less friction, and friction is what people route around rather than push through out of loyalty to a purchase order.

The scale of this problem shows up everywhere researchers have looked. MIT's research on generative AI deployments found 95% show zero measurable impact on the P&L. PwC found only 12% of companies get both revenue and cost benefits from AI; most get one, or neither. Prosci surveyed over 1,100 people running these rollouts and found technical issues accounted for 16% of the reported difficulty. User proficiency accounted for 38%. The bottleneck was never the model.

What actually closes the gap

A handful of things separate the companies where adoption sticks from the ones still wondering why the licenses sit unused. They're the same things that eventually worked for us, just applied on purpose instead of by accident.

Narrow the use case before you widen access. Our mistake wasn't building custom GPTs. It was handing them to everyone at once with no single workflow they were required to replace. Microsoft's own Copilot adoption guidance says the same thing: start with one focused application, not a blanket rollout. A tool with no specific job to do becomes a tool people route around.

Train for the actual job, not the demo. Generic training moves adoption from roughly 25% to 76%, according to Bright Horizons research, a lift most companies skip entirely by assuming the tool sells itself. Microsoft's internal Copilot rollout hit 90% monthly active usage through role-specific training and peer learning, not a company-wide announcement. We didn't do this until year two. Once someone sat down with the design team specifically and showed them how to get the image GPT to actually hold house style, that GPT started getting used.

Frame it as backup, not replacement. Harvard Business Review research from Narayandas and Zhang found employees avoid AI tools when the tool threatens their expertise and standing, independent of whether the tool actually works. This was the quieter reason our writers defaulted to captions. Writing client strategy with a tool that might get the brief wrong felt like a bigger risk than writing a caption with one that might. No one said this out loud in a meeting. It just showed up in what people actually did with the tool.

Measure adoption, not procurement. Licenses purchased tells you nothing. Microsoft got to 90% monthly active usage on its internal Copilot rollout because it tracked actual use as the metric that mattered, not seats provisioned. We made this mistake ourselves in year one: we counted GPTs built and assumed that counted as progress. It didn't. The number that mattered the whole time was how many people opened the tool that week, and nobody was watching it.

None of this is a technology problem. It's the same problem change management has always solved. Pick the workflow, train for it specifically, and don't ask people to bet their credibility on a tool that hasn't earned it yet.

What this costs you if you skip it

We spent two years finding this out by trial and error, with a team that wanted the tools to work and still avoided them for eighteen months. Most companies deploying Copilot today don't have two years of patience built into the budget, and the tool improved considerably in the time it took us to fix our approach rather than the model.

If your Copilot licenses are sitting at 25% usage, the fix isn't a better announcement or a more impressive demo. It's picking one workflow, training the people who'll actually run it, and giving them a reason to trust the tool before you ask them to depend on it.