Most winery AI rollouts I see are failing because of five moves the leader made before the tool ever got opened, which is a different problem than the one the rollout was set up to solve. The moves look responsible on a slide. They are the moves a software vendor recommends in their kickoff deck and an HR consultant adds to the rollout plan. Each one has a documented failure mode in the research, and most leaders are running two or three at once.
The cost of doing these things is not zero. They cost real money in licenses and training hours, and they cost something harder to recover, which is the team's willingness to believe the next rollout will be different. After the all-hands day where nothing changes, the second attempt is harder than the first one would have been if you had skipped the day.
This is a "stop doing" list because the alternative for each one is smaller, cheaper, and slower-looking, and it is the alternative the 5% of pilots that produce measurable impact keep running.
Stop doing company-wide license rollouts on day one
The standard move: buy ChatGPT Plus or Claude Team seats for every employee, send the welcome email, link to the FAQ. By month three the GM is looking at the admin dashboard, sees that 6 of 28 seats logged in last week, and decides AI does not work at the winery.
The failure mode is documented in MIT NANDA's GenAI Divide research from August 2025. Of enterprise AI pilots that ran since 2023, 95% produced no measurable P&L impact. The pilots that produced no impact were, in general, the ones with licenses and no setup; licenses alone were never the predictor. Meanwhile, the same report found that 90% of workers are already using personal AI daily, and only 40% of firms have an official subscription. Most of the team you are buying licenses for already has an account. They are using it on their phone for personal email. The license you bought is the third one they have access to, and the only one tied to a folder they have not built yet.
There is also a sequencing problem. When everyone gets the license on the same day, the second employee who tries it sees no peers using it and quietly stops. Gartner reported that 37% of non-users cite "my coworkers don't use it" as the reason. A 28-seat rollout on Monday creates 27 non-users at once.
The one-line replacement: one license, one employee, one task she does every week. Build her starter folder in a 90-minute session before anyone else gets a seat. Add the second license in week five, after the first habit is set. The "company-wide" rollout happens organically, when a third coworker walks over and asks how to set up what the first two have.
Stop doing all-hands AI training days
The pitch is reasonable on paper. Get everyone in a room for a day. Bring an outside trainer or play a vendor's recorded session. Hand out a slide deck and a list of 30 starter prompts. Move on.
The failure mode is that nothing about that format matches how a habit forms. Microsoft's Work Trend Index put the habit-formation threshold at 3 uses a week for 7 to 8 weeks. A one-day event delivers zero of that. Three weeks later, the prompt sheet is in a drawer and the most-cited barrier to AI use, per HBR, is the involvement gap: 78% of employees fear job loss to AI, and only 12% feel involved in deployment decisions. An all-hands day where leadership talks at the team for six hours widens both gaps. The cellar hand who already worried about being replaced now has six more hours of evidence the decision happened above her head.
The format also breaks the peer-trust pattern that the champion research keeps pointing to. Brennan McDonald's writeup of the wrong-champion failure mode named the underlying problem: when the person who introduces AI does not have peer credibility, the team's distrust attaches to the person, then to the AI, in that order. An outside trainer at an all-hands day is the maximum-distance version of that.
I wrote more about why this format fails so consistently in why all-hands AI training day kills adoption. The one-line replacement: a 15-minute Friday show-and-tell where one peer demos one thing she did with AI this week. Atlassian's research on AI-focused retros bolted onto existing rhythms found a 34% lift in AI alignment from this format. It costs nothing and replaces the training day's job.
Stop building individual leaderboards and usage dashboards
The instinct is reasonable. The leader wants to know who is using AI and who is not, so she asks the IT admin to pull a dashboard. Prompts per user per week. Active sessions. License utilization rate. Maybe a leaderboard on the breakroom screen showing the top three users.
Carnegie Mellon's Tepper School research on workplace gamification, alongside the older Disneyland "electronic whip" reporting, found that individual leaderboards tracking output erode moral agency. The team feels watched. The slow adopter, who is often the most experienced person on the floor and has the most institutional knowledge to encode into the tool, sees the dashboard and decides she would rather stay off it than appear low on it. The leaderboard punishes the exact person you needed to win over.
Individual usage dashboards have a second problem. They measure the wrong surface. MIT NANDA's shadow-AI finding (the 90/40 gap) means most actual AI use at your winery is happening on personal phones, not on the corporate license. A dashboard that says "the tasting room lead used AI twice this week" is showing the second-least informative slice of her real use. Her real use, the one on her personal ChatGPT account where she drafts club emails between visits, does not show up at all. The leader who runs a winery on the dashboard reads it as a failure and pulls the plug on the rollout that was working.
The one-line replacement: a team-versus-task metric the team picks together. "We are trying to cut the weekly club-cancellation reply turnaround from three days to one." Track the outcome at the team level. The research on team-vs-task framing, also from CMU, is that it produces engagement without the moral-agency erosion. Every member contributes; the dashboard does not name names. I went deep on why individual usage rate is the wrong number to track at all in stop measuring AI usage rate.
Stop running individual performance dashboards
This is a cousin of the leaderboard, and it deserves its own line because it slips into rollouts under a different name. The IT team or the HR head pitches a "personal AI insights" dashboard where each employee sees her own usage stats. Prompts per week. Saved-time estimate. Comparison to a benchmark.
The pitch is that the individual dashboard is supportive rather than surveillance: "we are giving each employee her own data." In practice, the dashboard activates the same anxiety. The employee opens it, sees she is below the benchmark, infers the company is watching, and goes quieter. Slack and Salesforce's Workforce Index from April-May 2025, with a sample of 5,156 workers, found that daily AI users reported 64% higher productivity and 81% higher job satisfaction; the dashboard converts that voluntary use into something measured, which is a different cognitive load. The use rate of an employee on a dashboard goes up briefly and then settles below where it would have been without the dashboard.
The second problem is that an individual dashboard implies AI use is a personal performance dimension on which the employee will be evaluated. The HBR involvement-gap data (78% fear job loss, 12% feel involved) means most employees walk into the dashboard already worried they are being scored on AI fluency to decide who keeps her job. A dashboard confirms the worry whether or not the leader intended it to.
The one-line replacement: a shared prompt library the team contributes to. Klarna's published rollout retrospective put the impact at 3.2x faster expertise development from a shared library. The library is the dashboard. The depth of the library tells the leader how the rollout is going, without naming any individual. Attribution by name in the library is the form of recognition that does not backfire (quiet recognition for the prompt's author; nobody outside the contributors sees ranking).
Stop forming an AI strategy committee
The fifth move is the one I see most often at wineries above 50 employees. The owner reads three pieces about AI, decides she should "get serious," and forms a cross-functional committee. The committee meets every other Wednesday. It has a charter document. It produces a strategy memo by the end of the quarter. The memo recommends piloting AI in three departments.
There are two problems. First, the committee output is a memo, and a memo is not a habit. Microsoft's 3x-a-week-for-8-weeks threshold tells you what you need to produce by the end of the quarter, and it is not a strategy document. The committee's deliverable does not pass through the threshold; it ends at the recommendation stage.
Second, the committee composition usually pulls the most senior people in the building, which is the opposite of who should be running the experiment. The champion research (Hartz AI, Brennan McDonald) is consistent: peer-trusted curious staff one or two levels down outperform the most senior or technical person on adoption. The committee structure routes the decision around them. By the time the memo recommends "piloting AI in three departments," the tasting room manager has been ignored for four months.
BCG's 10/20/70 finding is the broader version of the problem. Top-performing AI implementations split effort 10% algorithms, 20% tech and data, 70% people and process. A committee meets in the 10% bucket. The 70% (the part where the GM sits down with the DTC coordinator and rebuilds how she answers club cancellation emails) happens nowhere on the committee's agenda.
The one-line replacement: one named champion with a recurring 90-minute calendar slot. McKinsey's 2025 State of AI named the practices of firms that reach the scaling stage as a dedicated adoption owner alongside embedded use cases inside existing processes. A committee handles neither. A champion with a calendar slot handles both.
What to do instead
The pattern across the five stops is the same. Each thing-to-stop is a move that looks responsible because it is visible and scalable. Each thing-to-start is smaller and harder to put on a slide. The research is on the side of the smaller move every time, and the wineries I have watched run the smaller move first are the ones where, six months in, the second and third employee organically asked to be included.
If you are planning a rollout in the next quarter, the one-line replacements above give you a working draft of the play. The fuller version, with the timing and the sequencing for a 30-person winery, is what I am walking through in the Employee AI Engagement webinar. It is free and aimed at winery leaders specifically, and it includes the 60-day version of the play I keep handing to clients after the audit. If the moves above sound familiar from a rollout already underway, the webinar will give you a way to redirect what is already running without scrapping the budget.
Part of a 20-post series on employee AI adoption for wineries — see the full series under AI Adoption.