Picture a winery owner who has put a plan together that sounds responsible. A private room rented at the local conference center, a trainer recommended by the AI vendor, a Wednesday in March blocked off for the whole team. Tasting room staff, cellar hands, the club concierge, the wholesale ops person, the GM, the bookkeeper, all of them in one room for six hours. Lunch catered. Slides coming. There will be a quiz.
The question the plan does not answer is what happens on Thursday.
The plan dies on Thursday. It always dies on Thursday. The training day produces a binder, a Slack announcement, a thank-you note from the trainer, and a team of fourteen people who go back to their jobs and never open the tool again. By the time the next monthly check-in lands, the GM is asking why nobody is using the AI subscription he is paying for. The team's answer involves a quiz the cellar lead lost sleep over for three nights and a slide deck about prompt engineering that meant nothing to a tasting room manager who has been hand-writing thank-you cards for a decade.
This is the second post in a short series on what to stop doing. The five things to stop doing post that precedes this one names the all-hands training day as the second item on the list. This post is the long version of why, and what the replacement looks like.
The math of a forced rollout
The pattern is well documented and the numbers are easy to reach. Harvard Business Review published research from Keith Ferrazzi's group in February 2026 that put two numbers next to each other. Seventy-eight percent of employees fear losing their job to AI. Twelve percent feel involved in the AI deployment decisions their employer is making. The gap between those two numbers is the room you walk people into when you book a training day.
Twelve percent involvement means the other eighty-eight percent are showing up to a meeting where someone is going to teach them to use the thing they are already afraid will replace them. From inside the team's experience, the training day is the announcement of the replacement, with the slide deck as proof and the mandatory attendance as the warning.
Nothing about that experience produces a habit. It produces a quieter team for a week, a flurry of resume-update activity from the most senior staff that the GM will not see until somebody resigns in May, and a tool sitting unused because using it now signals to the room that you are taking the side of the thing that came to replace you.
What a training day does not contain
Even setting the fear aside, a training day fails on the mechanics. The play that produces a habit has three pieces, and none of them are in the conference room.
The first missing piece is peer modeling. Gartner found that thirty-seven percent of employees who are not using AI cite "my coworkers don't use it" as the primary reason. A training day teaches people how to operate the tool in isolation. It does not let them see another tasting room staffer use it on a Tuesday afternoon task they recognize. The thing that flips a skeptic is watching a peer get a useful answer back. Watching a vendor consultant get one moves nothing. The room at the conference center contains a trainer the team does not know and will never see again.
The second missing piece is follow-up. Microsoft's Work Trend Index research keeps showing up in this series because it is the only credible habit-formation number anyone has published on AI use. Three uses a week for seven to eight weeks. Below that threshold, employees bounce off the tool. A training day on a Wednesday produces, at best, one use that week, by people who were forced to be there. By the second Wednesday after the training, the count for most attendees is zero. There is no scaffolding for the weeks that matter.
The third missing piece is the staffer's own work. The training day uses sample prompts: write a poem about a vineyard, or summarize a fake meeting transcript, or generate a marketing email for a wine that does not exist. The staff member spends six hours watching the tool produce things that have no relationship to anything she will be asked to do on Thursday morning. The transfer never happens, because there is nothing to transfer to.
A six-hour block of catered training has high optics and almost no operational surface area. It feels like the winery is doing the responsible thing to the owner, the vendor, and the trainer. It does not feel that way to the staff member who is back in front of her work on Thursday morning with no prompt that fits and no peer who can show her one.
What the involvement gap is measuring
The HBR involvement number is the one worth holding in your head for the longest. Twelve percent of employees feel involved in the AI deployment decisions at their company. That figure describes a structural choice every winery has already made about who gets to shape the rollout, well before it describes any attitude problem on the team.
When the GM walks in with the training day plan and a vendor relationship already in place, the staff is being told that the decisions have happened. The training day is a downstream artifact of a tool choice they did not weigh in on, a budget decision they did not see, and a workflow change they will be asked to absorb. The implicit message is that the work of figuring out what AI is for at the winery is finished, and now they are being briefed.
The work of figuring out what AI is for at a thirty-person winery is not finished. It has not started. The training day is being scheduled before the conversation that should produce the rollout has happened with a single staff member. That conversation looks like this: pull up a chair next to the tasting room manager on a Tuesday afternoon and ask her what she already uses AI for in her personal life. The answer is almost never zero. Ask her what part of her job she would hand to a tool that could do it well. She has an answer for that too.
That conversation is the involvement the HBR number is measuring. The training day is what skipping the conversation looks like at scale.
The shape of what works
The replacement is a different shape entirely from the training day. The series will get into the full play in later posts, but the contrast that matters here is between one room of fourteen people and one room of two.
The play starts with one person, in their normal work setting, with the recurring task they care about open in front of them. The session is ninety minutes. The output of the session is a folder with the role description, three or four samples of how the work usually looks, and the shape of the output the staffer wants. It is built collaboratively. The leader does not bring slides. The staffer does most of the talking. The tool is a prop, used in the last twenty minutes, after the work has been described well enough that the prop can produce something useful on the first try.
I have written about the selection criteria for that first person in posts that will arrive later in this series: why your tasting room manager is the right first AI champion, and the cautionary wrong-champion failure mode that comes from picking the most technical or most senior staffer instead of the most trusted one. The selection question matters more than any other call the GM will make, because the first user is the entire visible surface of the rollout for the next six weeks.
After ninety minutes with that one person, the play asks for three uses a week for eight weeks. Nothing else for the first two months: no Slack announcement, no second user, no quiz, no rollout email. The team will notice that the tasting room manager is doing something new and will ask her. She will show them on her phone. That is peer modeling, free, with no training day required.
If you want a measurement question to stop tracking, "AI usage rate across the team" is the one. I will go after that one in detail in the next post, stop measuring AI usage rate. The training-day instinct and the usage-rate instinct come from the same place: the wish that you could turn AI adoption into a top-down rollout. You cannot. It is closer to how the staff learns a new POS system from each other than how they learn a compliance module from corporate.
Why six hours feels safer than ninety minutes
Owners default to the training day for structural reasons, not analytical ones. A six-hour training day is legible. It can be calendared, budgeted, photographed, and reported to the board. A ninety-minute working session with the tasting room manager is invisible from any reporting tool the GM is using. It does not show up in a vendor contract. It produces no binder.
Most of the leaders I work with carry a quiet suspicion that the invisible version cannot be the right move, because it is too small. Six hours of mandatory training feels proportional to the budget of an AI rollout. Ninety minutes of working alongside one person does not.
Proportion is the wrong heuristic here. The 95% failure rate that MIT NANDA published in 2025 says the proportional response is also the one that does not work. The 5% of pilots that produced measurable impact were the ones where the work had been redesigned around the tool, in conversation with the person doing the work. Training duration did not separate them. BCG's research, which I have cited in this series before, puts seventy percent of the effort on people, process, and workflow. The training day puts about three percent there. The other ninety-seven percent is spent on slides about what AI is.
The disproportion is what works. A single ninety-minute conversation, repeated weekly for eight weeks with one staffer, sits at a higher leverage point than a six-hour event with fourteen staffers. The leverage point is the part the training day cannot reach: the staffer's Thursday morning.
To make that concrete, picture how this plays out at Not Really Wines, our fictional demo winery in Healdsburg. The training-day version brings the whole team in on a Wednesday: tasting bar staff, the production assistant, the DTC director, the hospitality lead, the operations lead, the wholesale rep who just joined in May. Six hours, a trainer flown in, a binder per person. The DTC director leaves with two ideas she might try. Nobody else opens anything. The trainer's number is forgotten by Friday. The ninety-minute version skips the training day entirely. The hospitality lead sits with the DTC director on a Tuesday afternoon. They build one folder for the cancellation-reply task the DTC director runs every Monday. Eight weeks later that task takes four minutes instead of forty. The tasting bar lead notices, asks how, and the play has its second user. The cost difference between those two versions is a five-figure swing on the training-day side, and a two-hour swing on the working-session side. The adoption difference, eight weeks in, is one user against fourteen non-users.
A note on the webinar
I am running a free webinar that walks through the full 60-day version of the play this post describes — including the ninety-minute starter session, the second-user move that arrives in week five, and the rituals that hold the play together past the brain-fry window in month two. If you have a training day on the calendar and you are starting to suspect it is not going to do what you hoped, the webinar is for that exact situation. Check the webinar schedule for the next date and to register.
The team is already training itself, on their phones, between shifts. The open question for a winery owner is whether the work the winery pays for ever gets to ride along.
Part of a 20-post series on employee AI adoption for wineries — see the full series under AI Adoption.