Buying AI Tools Isn't Adopting AI — and the Bill Comes Due Later
Here’s a number that should stop any manager mid-sentence: more than 80% of enterprise AI projects fail to deliver what they promised — roughly twice the failure rate of ordinary IT projects, according to a 2024 RAND Corporation study. And when researchers dig into why, the answer is almost never the technology. The models work. The rollout doesn’t.
If your organization has bought AI seats for everyone and is quietly wondering why nothing changed, this is the reason. Adoption got treated as a purchase. It’s actually a behavior change — and those are two very different budgets.
Buying seats is not a strategy
Watch what companies actually do. They sign the enterprise license, flip on access for a few thousand people, send one launch email, and mark “adopt AI” as done. The spending is real and the intent is real. The follow-through isn’t.
The clearest evidence is a promise employers made and then didn’t keep. 77% of employers say they plan to reskill their workforce for AI between 2025 and 2030 (World Economic Forum’s Future of Jobs Report 2025). But when you ask the workers, only about 13% say they’ve actually received any AI training (Randstad). That gap — 77% intend, 13% delivered — is the failure mechanism. You can’t buy your way across it with a bigger license.
Untrained rollout doesn’t produce nothing — it produces mess
Here’s the part leaders miss. When you hand a powerful tool to people with no guidance, they don’t simply not use it. Some of them use it badly, at scale, and pass the results downstream to colleagues.
Researchers have a name for the result now: Low-effort, AI-generated work that looks polished but is hollow — a plausible-sounding doc or deck that quietly pushes the real thinking onto whoever receives it. Coined by BetterUp Labs and Stanford’s Social Media Lab in 2025. . In their study, 40% of workers reported receiving workslop from a colleague in the past month, and each incident cost nearly two hours to untangle. So the untrained rollout isn’t free — it has a negative return. Someone generates a confident, empty draft in thirty seconds; someone else spends an afternoon discovering it’s empty. Multiply that across a company and you’ve manufactured work, not saved it.
That’s how an AI investment can show up on the books as a cost with no matching benefit — the tool is being used, just not in a way that helps.
What the 1-in-5 that succeed actually do
The RAND research is more useful for its minority than its majority. The projects that worked shared a short list of traits, and almost none of them are about the model:
- Clear definition of what success looks like before rollout — a specific task, a measurable before-and-after, not “use AI more.”
- Real The unglamorous work of helping people actually change how they do their jobs: training, new norms, examples of good and bad output, someone to ask when you’re stuck. It’s the difference between installing software and changing behavior — and it’s the part that gets cut first. — training tied to actual workflows, not a generic webinar.
- Integration into the work people already do, instead of a shiny tool bolted on beside it.
- Sustained sponsorship — a leader who stays interested past launch week.
Notice what’s not on that list: a better model, a bigger budget, a newer vendor. The successful fifth didn’t buy something different. They did something different after buying.
The cheapest fix is the one nobody funds
The strange part is that training is comically cheap next to everything around it. It’s a rounding error against the license spend, and a bargain against the cost of a failed six-figure initiative that gets quietly shelved. Yet it’s the first line cut, because a training program is harder to point at than a signed contract.
Workers already know the tools matter — they’re not resisting. 52% of technology professionals report seeking AI training on their own (Randstad Digital), precisely because their employer’s program can’t keep up or doesn’t exist. That’s the demand signal. People want to get good at this. The organizations that win are the ones that meet that appetite with real support instead of leaving everyone to figure it out alone and then wondering why the numbers don’t move.
The takeaway
If you’re deciding where the AI budget goes: stop measuring adoption by licenses activated and start measuring it by behavior changed. A smaller rollout with real training beats a company-wide rollout with none — every time, and it’s not close. Pick one team, one repeated task, define what “better” means, train for that, and prove it before you scale.
And if you’re an individual watching your employer promise training that never arrives — you don’t have to wait for them. That’s a post of its own.
Sources: the 80%+ failure rate (twice that of non-AI projects) and the traits of the projects that succeed come from the RAND Corporation’s 2024 report on why AI projects fail. The 77% of employers planning to reskill is from the World Economic Forum’s Future of Jobs Report 2025; the 13% of workers who’ve received training and the 52% self-training figure are from Randstad’s workforce surveys. “Workslop” and the 40% who received it come from BetterUp Labs and Stanford’s Social Media Lab, via Harvard Business Review. The through-line connecting the training gap to the failure rate — that buying seats without change management is the mechanism, not just a correlate — is my own read of these findings, not a single study’s claim.
Related: How to Learn AI at Work When Nobody’s Training You is the individual’s companion to this piece — what to do when the training your employer promised never shows up. And Why Most Workplace AI Rollouts Quietly Fail covers the missing-returns problem that this training gap helps explain.