Why AI adoption is an operating model, not a tool rollout

I keep having the same conversation. An executive tells me their company is doing AI. What they mean is they bought Copilot or ChatGPT Enterprise seats, announced a pilot, and sent a memo about responsible use. Six months in, a handful of people use it daily, most don’t, and nobody can point to a line in the P&L that moved.

The instinct is to blame the tool. Wrong vendor, wrong model, maybe next quarter’s release fixes it. After three years shipping production AI at AnswerAI, the AI product company I run, I don’t believe that anymore. The tools mostly work. What failed was the framing: the company made a procurement decision when the situation called for an operating-model decision. Those are different sizes of decision, and confusing them is the most expensive mistake in enterprise AI right now.

The stall is measurable

MIT’s NANDA initiative put numbers on this in 2025. Across 150 executive interviews, a 350-employee survey, and 300 public deployments, about 5% of AI pilot programs achieved rapid revenue acceleration. The rest stalled with little to no measurable P&L impact. Read the report carefully, though, because it does not say the technology failed. It says organizations failed to absorb it — the researchers call it a learning gap. Generic tools don’t adapt to company workflows, and companies don’t adapt their workflows to the tools. Meanwhile the same report documents widespread shadow AI: employees quietly using personal accounts to do real work while the official pilot goes nowhere.

The abandonment data tells the same story. S&P Global Market Intelligence found that the share of companies scrapping most of their AI initiatives jumped from 17% to 42% in a single year, and the average organization killed 46% of its AI proofs-of-concept before production.

Those are not technology numbers. Those are ownership numbers. A 42% abandonment rate means companies started things nobody was accountable for finishing.

What an operating model actually is

When I say AI adoption is an operating-model change, I mean four specific decisions that someone with executive authority has to make and keep making:

1. What data is allowed where. Which systems can see client files, financials, HR records. What never leaves your walls.
2. Which tools are approved, and what the path is for approving a new one. Not a ban list — a decision process with a name attached.
3. Which workflows come first. Not “everyone experiment,” but a ranked backlog: this process, this team, this quarter.
4. How ROI is measured. A baseline number for the workflow before AI touches it, and a date when someone compares.

Notice what’s on that list: boring executive decisions. Nothing about transformer architectures. A tool rollout answers none of these questions. It just puts software in front of people and hopes.

Pilots without an owner produce demos

Here’s the difference in practice. In one engagement delivered with Last Rev, the platform engineering firm I co-founded, a construction services company took RFP responses from 40–60 hours of work down to 8–12. The model was maybe a third of that outcome. The rest was operating-model work: deciding what belonged in the knowledge base and who kept it current, which sections got drafted from it and which a human still wrote from scratch, and what review meant before anything went out the door. Someone owned each of those calls. That’s why it became a system instead of a demo.

The failure mode I see in unowned pilots is that they measure activity instead of work. Seats activated, prompts sent, “engagement.” None of that is a business result. So I use a simple heuristic before any pilot starts: four things written down — the owner, the workflow, the baseline number, and the decision the result will inform. If a pilot can’t name its baseline, it isn’t a pilot. It’s a demo with a budget.

Governance follows the same logic. At an ad-tech company, we found 18% of AI tool interactions touched sensitive data. Policy plus architecture took that to zero — not by banning tools, which just drives usage underground into those personal accounts MIT documented, but by deciding deliberately which data goes where and giving people a governed path that was easier than the workaround.

The data on where value comes from

If this sounds like opinion, McKinsey’s numbers back it up. Their State of AI survey tested 25 organizational attributes against reported EBIT impact from gen AI. The attribute with the biggest effect: redesign of workflows. And one of the elements most correlated with self-reported bottom-line impact — the single strongest factor at larger companies — was the CEO’s oversight of AI governance. Not the model. Not the budget. Whether the top of the house owned the rules.

The same survey shows how few companies act on this: only 21% of organizations using gen AI have fundamentally redesigned even some workflows. Most companies are running the exact play the evidence says doesn’t work — deploy tools into unchanged processes, measure nothing, wonder why the P&L looks the same.

So who owns it

The common answer is to hand AI to the IT lead or the smartest engineer in the building. I understand the impulse, but tool selection isn’t the job. The job is changing how the business operates: data policy, spend, workflow priority, measurement, and the authority to say no. Those are executive functions. Your best engineer can tell you which retrieval approach fits your document mess. They cannot decide, on their own authority, that client data never enters a public model, or that the RFP process gets rebuilt before the marketing team gets its content tools.

That ownership doesn’t have to be a $400K full-time hire on day one. It does have to be a named person with real authority over those four decisions, reporting results against baselines on a regular cadence.

Every company I talk to already has the tools, or can buy them this afternoon. Model access is now the most evenly distributed asset in business. What separates the 5% from the 95% is that someone owns the operating model the tools run inside. Buy the licenses if you want. Just understand that the purchase order is the smallest decision in the program — and the only one most companies ever actually make.