You bought the AI seats. Here's the part nobody sold you.

The purchase is the part that works. Leadership agrees the firm needs AI, procurement negotiates enterprise seats, IT wires up single sign-on, and somebody books a lunch-and-learn. Everyone in that chain has rolled out software before, and this looks exactly like rolling out software.

Ninety days later the admin console shows 15% of seats active in a given week. The sponsor stops bringing it up in leadership meetings. Nobody declares failure; the line item just gets quieter until renewal, when it disappears without a eulogy.

The scale of this pattern is documented. MIT’s NANDA initiative found that about 95% of enterprise generative AI pilots are producing zero measurable return. I don’t read that as a verdict on the technology, because I’ve spent three years shipping AI systems that pay for themselves — at AnswerAI, the AI product company I run, and through delivery work at Last Rev, the engineering firm I co-founded. I read it as a category error. Companies are buying an operating-model change and deploying it like software, because it arrives shaped like software.

Software installs. An operating model has to be adopted

When you roll out a new expense system, adoption is guaranteed, because there is no other way to file an expense. An AI seat has no such enforcement. Every task it could help with already has a way of getting done, and the existing way requires no login, no learning curve, and no chance of looking foolish in front of a new tool. Nothing breaks when nobody uses the seat. That’s what makes this failure so quiet: an idle AI deployment produces no incidents and no complaints. Just an invoice.

So usage has to be pulled, not pushed, and the pulling is a sequence rather than an event. What follows is the sequence I now run, with numbers from a composite client — an accounting firm assembled from patterns across real engagements, not a disguised real firm — that took it from 40 seats to 78 in 60 days, with weekly active use climbing from 41% to 64%.

The order is load-bearing. Each step exists to make the next one possible.

Contractual guardrails come before the first login

Three things happen before any seat goes live: a no-training commitment from the vendor, single sign-on wired to your identity system, and data-loss-prevention rules on the channels people already use.

The no-training piece is the one leadership teams treat as paranoia, and it’s actually the foundation, because it determines what your policy is allowed to permit. Anthropic’s published commitment not to train on business customers’ data is what a real answer looks like: specific, written, and referenceable in a contract. If your vendor’s answer is vaguer than that page, your policy has to treat the tool like a public place, and a tool the policy treats as a public place can never touch real client work. That caps the value of every seat you bought before anyone logs in.

This is also why the step can’t come later. Roll out first and paper it afterward, and you get one of two outcomes: a policy that forbids the work people actually do, or people doing the actual work outside the policy. The composite firm executed its no-training addendum on day one of rollout week one, before a single seat went live — with the written policy and DLP rules already in place two weeks earlier. The DLP rules earned their keep by June, catching two near-misses that were handled as coaching conversations instead of incidents — the difference between a control that teaches and a control that makes people hide.

A policy someone can follow at 4:45 on a Thursday

With the contract signed, the policy can be short. The composite firm’s fits on a page: three data tiers. Green, public or publishable, any approved tool. Yellow, internal operations, approved enterprise tools only. Red, client tax information and financials, named governed systems only, never a personal account.

The test of an AI policy is not whether counsel is satisfied. It’s whether a staff accountant with a client file open at 4:45 on a Thursday can answer “can I paste this?” in under ten seconds. If answering requires re-reading a memo, the operative policy is whatever people guess, and people guess generously.

Champions are chosen for credibility, not enthusiasm

The instinct is to appoint whoever is most excited about AI. Resist it. The office enthusiast is often the person colleagues already discount, and endorsement from a discounted source is worth nothing. Pick the people others go to when work is stuck, and give each one a concrete job instead of an evangelism mandate: one champion per department, each owning something real — edge cases in the letter workflow, cleanup of the template library, user testing for the intake build.

The composite firm’s most effective advocate turned out to be a tax partner who had been the loudest skeptic in the assessment interviews. When she said the tool saved her time, it moved the entire tax practice, precisely because everyone knew she had wanted it to fail.

The quick win has a deadline: 45 days

People decide whether a rollout is real in the first several weeks, and they decide by watching whether anything visible changes. So the sequence includes a deliberately chosen quick win — high volume, low drama, measurable — with a shipping deadline inside the first 45 days.

The composite firm shipped its win on day 36: engagement-letter drafting from precedent, a task partners had been spending 3.5 hours per letter on. The team seeded a precedent library with 60 letters and 25 proposals before training began, so the tool knew the house style on day one. Average drafting time fell to 45 minutes. That visible result did the recruiting no lunch-and-learn could do; seat requests started arriving instead of being pushed, and the ramp to 78 seats followed. Notice the placement, though. The win comes after the guardrails and the policy, because a win you can’t govern is next quarter’s incident.

Measure from the admin console, never from surveys

KPMG and the University of Melbourne found that 57% of employees hide their AI use from their employer. Ask people about adoption in a survey and you collect the number they think is safe to report. The admin console doesn’t negotiate. Weekly active users, by department, is the metric, and its refusal to flatter is the point. The composite firm’s console showed tax at 71% weekly active use and audit at 39% — not a verdict, a to-do list. Audit’s caution had specific causes, professional standards and peer review among them, that no amount of training was going to fix.

The one survey worth keeping is the anonymous shadow-AI spot check, because it measures the thing consoles can’t see. At assessment, 41 of the firm’s 96 survey respondents were using unsanctioned public AI tools for work, and 9 admitted pasting client information into free accounts. By June the count was down to 12 of 96. Shadow use didn’t fall because of enforcement. It fell because the sanctioned tool got good enough to win.

Pilot purgatory is a place, and it has an exit

There’s a name for where the unsequenced rollout lands: pilot purgatory. Seats bought, lunch-and-learn held, adoption plateaued near 15%, and the project neither expanded nor killed, just waiting for the renewal date to make the decision nobody else would. The abandonment data shows how crowded that waiting room is. S&P Global found the share of companies abandoning most of their AI initiatives jumped from 17% to 42% in a single year. Few of those were dramatic failures. Most were quiet write-offs of seats nobody sequenced.

The exit is not more training, and it is not a better model. It’s the steps above, run in order, with dates attached. I’ve published the version I wrote for the composite firm: the sample Commercial AI Enablement Playbook — the full 60-day runbook for a composite client — with the rest of the sample deliverables. Contracts, then policy, then champions, then the win, then the numbers.

The seats were never the product. The operating model is the product. The invoice just arrives first.