By Daniel Wishnia · Founder, Wish On Line · wishol.com June 2026 · 7 min read
Enterprise AI adoption almost never fails the way the post-mortem claims. The slide deck blames the technology, or "change resistance," or users who weren't ready. I've run these programs inside a listed real estate group, a top-five national insurer, hospitality operators across several countries, and teams small enough to fit in one room, and the real story is duller and far more fixable. The technology worked. The way it was deployed is what broke.
There's a moment in every rollout when the dashboard tells the truth. Usually week six. Licenses assigned: hundreds. Weekly active users: a thin stripe along the bottom of the chart, mostly the same fifteen enthusiasts who'd have found the tool on their own. The kickoff was loud, the demo got applause, and then the organization quietly went back to working the way it did in March.
That's only half the picture, though. And the half nobody puts on a slide is more interesting.
While the official rollout sits empty, AI has already walked in the back door
Here's what I find underneath the flat usage chart. The AI didn't fail to arrive. It arrived without anyone deciding it should.
One team is pasting client documents into a public chatbot to summarize them. Marketing is drafting in tools procurement has never heard of. Someone in operations automated a weekly report with a model and told no one. People use assistants every day without knowing which data they're allowed to enter and which they aren't. The strategy committee thinks adoption is at 4%. The real number, if you could see it all, is much higher and completely ungoverned.
This is Shadow AI, and it isn't really a story about employees breaking rules. It's a signal about the organization itself: the company is moving more slowly than its own people. When leadership gives no direction, motivated teams route around the silence. You don't get zero adoption. You get invisible adoption, which is worse, because invisible adoption carries every risk and captures none of the value.
In regulated environments, the stakes are blunt. Working with an insurer, the questions that mattered weren't about prompt quality. They were about whether sensitive data was leaving controlled systems, whether a decision made with AI assistance could be traced afterward, and whether two departments were quietly solving the same problem twice. Prohibiting everything doesn't work; the teams just hide it better. Looking the other way doesn't work either. You don't slow AI down by banning it. You bring it into the open with leadership, clear rules, and named ownership.
So the real failure has two faces at once. A top-down rollout that's all license and no behavior, and a bottom-up reality that's all behavior and no governance. Most companies have both, and they're the same root cause wearing two costumes.
What enterprise AI adoption actually is
Enterprise AI adoption is the process of turning licensed AI capabilities into changes in daily work, measured by whether specific roles complete specific tasks differently than before. It is not license distribution. It is not a training event. It is not a usage statistic, and it is definitely not the count of people who logged in once.
A rollout has succeeded when the controller closes the month using the tool, the underwriter assesses a case with it, the marketer drafts in it by default, and the organization cannot quietly revert without someone noticing the loss. That definition sounds obvious. Almost no program I've audited was actually managed against it. They were managed against deployment milestones, which is exactly how you produce the proud announcement and the empty dashboard in the same quarter.
The three failures, in the order they happen
The generic training failure comes first. Most programs teach the tool, not the job. Here's the interface, here's what a prompt is, here are ten use cases that apply to no one in particular. It works as a theater and fails as a transfer. A financial controller doesn't need "how to use Copilot." She needs to see her actual monthly close completed in a third of the time using her real data, plus the three places where the model will mislead her. When I built the AI-and-Excel workshop for a group's business controllers, the whole thing was built from their real files, even down to an awkward constraint about where the assistant could and couldn't sit inside the spreadsheet. People adopt when they recognize their own Tuesday in the training. They never adopt a feature tour.
The missing-constraints failure comes second, and it's the one with no budget plans for. Every organization has plumbing realities that silently cap what the AI can do. Files living on legacy on-premises servers that the tool can't index. Sensitivity labels half-deployed. Meeting recording switched off across most departments. Data sitting in a system with no connector. Roll out without mapping these, and your pilot users hit invisible walls in week one, decide "it doesn't work here," and that verdict travels faster than any comms plan. In the insurance program I ran, we documented those constraints up front and tested eighteen real scenarios against that honest map. The point of a pilot isn't to show off the tool. It's to find the walls before five hundred people find them for you.
The measurement failure comes third and finishes the job. "Adoption" gets reported as logins, which is like rating a gym by counting door swipes. What works instead is a per-role scenario matrix: for each function, the five tasks the tool should transform, each with a pass/fail test and an owner. Now the rollout has a scoreboard that managers actually understand, the gaps have names, and the conversation moves from "people aren't engaged" to "scenario twelve fails because the data source isn't connected," which is something a human can go fix on Thursday.
The six gaps Shadow AI leaves behind
When adoption stays invisible, the cost shows up as a predictable set of holes. I see the same six in nearly every organization that skipped the governance step:
- Teams using AI with no oversight.
- Sensitive data leaving controlled systems.
- The same process was rebuilt in parallel two or three times.
- Decisions no one can trace back later.
- Productivity gains that never get measured or scaled.
- Governance that exists on paper and nowhere else.
None of these gets solved by a better model. They get solved by a decision to govern, which is a leadership act, not an IT ticket.
The shape of a rollout that sticks
Strip away the vendor methodology decks, and the working pattern is short. Start narrower than feels ambitious: two or three roles, their highest-friction recurring tasks, real data in every exercise. Build role-specific playbooks instead of tool documentation, including the failure modes, because teaching people where the model is wrong earns more trust than pretending it never is. Run the constraints audit before the pilot, not as its autopsy. Assign a named owner to the scenario matrix and review it monthly, like any other operating metric. And bring Shadow AI into the light early, not with a ban but with a simple, usable policy that tells people what they can put where, so the energy already in the building gets pointed somewhere safe.
The honest timeline is quarters, not weeks. The honest budget is mostly people-time, not licenses. Any plan whose cost is 90% software is a plan to fail with excellent procurement.
Why is this now a board topic, not an IT one
For a while, a stalled rollout was an embarrassing line in the IT review. Two things changed that. The first is compounding skill. The organizations where AI use has become the default work are accumulating prompt fluency, workflow redesign, and hard-won judgment about the tools at a rate latecomers can't buy back later, because it lives in people, not contracts. The second is regulation. The EU AI Act and its relatives quietly reward exactly the disciplined pattern described here, documented use cases, named owners, tested constraints, traceable decisions, and they punish the laissez-faire version where nobody can say how, where, or with what data the company is using AI.
Which is the real question for any executive now? Not whether your company is using AI; it is, with or without you. The harder one is whether you actually know how, where, for what, and with whose data. Adoption was never the soft part of the AI story. It was always the whole story. The technology was the easy bit.
FAQ
What is Shadow AI? Shadow AI is the unsanctioned use of AI tools by employees without organizational oversight, governance, or data controls. It usually signals that the company is moving more slowly than its own teams, and it carries the risks of AI use while capturing little of the measurable value because nobody is tracking or scaling what works.
Why do most enterprise AI rollouts fail? They fail on deployment, not technology. The three recurring causes are generic training that teaches the tool rather than the job, unmapped technical constraints that block pilot users in week one, and measuring adoption by logins rather than by whether real tasks have changed. All three are fixable without changing the underlying tool.
How do you properly measure enterprise AI adoption? With a per-role scenario matrix: for each function, list the specific tasks the tool should transform, give each a concrete pass/fail test, a status, and an owner. This replaces vanity login counts with a scoreboard that turns vague "low engagement" into named, fixable blockers.
Daniel Wishnia is the founder of Wish On Line (https://www.wishol.com). He has led GenAI adoption and Copilot 365 enablement across a listed real estate group, a top-five national insurer, and hospitality operators, following his tenure as Chief Digital Transformation Officer. Contact: daniel.wishnia@wishol.com
Top comments (0)