When a winery leader asks me how their AI rollout is going, the honest version of the answer lives inside four numbers from research published in the last twelve months. None of the four are vibes. Each has a specific source, a specific study size, and a specific implication for a winery with somewhere between twenty and fifty people on payroll.
A lot of the conversation about employee AI happens at the level of slogans. Everyone is using it. Nobody is using it. ChatGPT changed everything. ChatGPT changed nothing. I find the numbered version of that conversation much more useful, because it lets you compare your own winery to a known reference point instead of to whatever LinkedIn posted this morning.
Here are the four I come back to in nearly every conversation. I will name the study, the sample, and the so-what for a working winery in each one.
The 95 percent that goes nowhere
The first number is from MIT NANDA's "GenAI Divide" report, published August 2025. The headline finding: 95 percent of enterprise generative AI pilots produce no measurable profit-and-loss impact. That is the entire enterprise sample, across industries, after eighteen months of generative AI being widely available. Ninety-five out of every hundred rollouts went nowhere measurable.
The shape of the 5 percent that did work is more useful than the failure rate itself. Bought-or-partnered systems succeeded at 67 percent. Internally-built ones succeeded at 22 percent. Pilots that retained employee feedback and adapted lived past the novelty window; ones that did not, died around month three.
For a 20-to-50-person winery, the so-what is unglamorous. The winning move is almost always to pick a tool that already exists, attach it to one specific employee's recurring task, and feed the result back into the setup every two weeks. The losing move is to commission a custom build for "our winery's specific needs." Custom is where wineries go to spend money and end up at the wrong end of the 95 percent. I covered the failure-rate finding in more depth in the 95% failure rate, which is the starting point for this whole series.
Three times a week for eight weeks
The second number is from Microsoft's Work Trend Index research on AI habit formation. Employees who use an AI tool roughly three times a week for seven to eight weeks cross into a habit. Below that frequency, they bounce off — open the app, run one prompt, forget about it for ten days, treat it as a novelty when they come back. Above that frequency for that duration, the tool stops feeling like a separate destination and starts feeling like part of the job.
The implication is the entire shape of a winery's 60-day rollout. If three times a week for eight weeks is the threshold, then a Monday-Wednesday-Friday rhythm is the right cadence for the first two months, and the first metric worth tracking is how many named people are hitting that frequency in any given week. License-activation counts measure a different thing. Everything else is decoration on the behavior question.
The reason most winery rollouts fail this part is that nobody designs around it. Leaders buy ChatGPT Plus for the team in February and check in at the all-hands in May. By May, two of the eight weeks needed to form a habit have already been lost to the manager's silence. I worked through the math on this number in the 3x-a-week-for-8-weeks rule, which is the post most directly responsible for how I structure the play.
So-what for the winery: pick a Monday-Wednesday-Friday rhythm. Pick one person. Pick one task that recurs that often anyway. Run it from week one to week eight without skipping weeks. Habit forms or it does not. There is no middle outcome.
Eighty-one percent more satisfied
The third number is the one I find leaders dismiss fastest and the one I now spend the most time on. The Slack/Salesforce Workforce Index, published April-May 2025 with a sample of 5,156 desk workers, asked daily AI users how their work felt compared to a baseline year. Daily users reported a 64 percent productivity lift, a 58 percent focus lift, and an 81 percent lift in job satisfaction.
The productivity and focus numbers get nodded at. The satisfaction number gets dismissed as soft. I think that read is exactly backwards. Self-reported satisfaction is the one metric that predicts whether an employee keeps using a tool after the novelty wears off. A burnt-out tasting room manager who hates her week does not pick up a new tool in month two. A satisfied one does. The 81 percent finding is the leading indicator for whether month three exists at all.
There is a wine-industry overlay on this finding. Tasting room, club concierge, and DTC coordinator roles have higher-than-average turnover, partly because the work involves a lot of repetitive output drafting and a lot of emotional labor with customers. The Slack/Salesforce finding suggests that even a partial reduction in the drafting load — visit recaps, follow-up notes, club cancellation responses — moves the satisfaction needle in a way that compounds with retention. The retention gain on its own probably pays for the rollout, even before you count the time saved on the underlying task. This is part of why the Klarna story is worth a separate look, and I will get into what Klarna learned in the next post in this series.
So-what for the winery: track satisfaction at the role level, not just hours saved. A short, in-person, 15-minute check-in once a quarter with each role-holder catches this. A quarterly engagement survey across the whole team does not.
One champion per fifteen to twenty employees
The fourth number is the cohort-sizing rule, and it is the one that gets the most pushback from leaders who pride themselves on running lean. Hartz AI's documented work on champion programs lands on a ratio of one champion for every 15 to 20 employees. Their finer-grained breakdown is 2 to 3 active champions per 10 to 12 people while a team is actively in adoption phase, dropping to one per 15-to-20 once the habits are settled.
For a 20-person winery this means one champion in adoption phase, and probably one champion long-term. For a 40-person winery this means two to three during adoption, and probably two long-term. For a 60-person winery this means three to four during adoption, and likely three long-term. There is a real upper bound here. Anything tighter than 1-to-10 and the champion's coworkers start treating the champion as a watcher rather than a peer. Anything looser than 1-to-20 and stuck employees give up silently before the champion notices.
The deeper finding under the ratio is that the champion role works by peer proximity, not by expertise broadcast. A champion who has fifty people to support cannot run the small in-person interventions that produce adoption — the 90-minute starter-folder build, the show-me-on-your-phone moment, the five-minute "it hallucinated, now what" demo. Stretched too thin, the champion becomes a help desk; the help desk is the thing this entire approach is trying to avoid.
So-what for the winery: count your headcount, divide by 15 to 20, round up, and that is your champion cohort during adoption. If the math returns a champion who is the wrong person, find a better person before you find a different number. I worked through the selection criteria in more depth in the right first AI champion, and the cohort-sizing rule sits on top of those criteria — get the person right first, then check that you have enough of them.
What the four numbers do together
Read end to end, the four numbers describe a system. The 95 percent failure rate sets the prior — most rollouts go nowhere. The 3-times-a-week-for-8-weeks threshold sets the behavior the system is built around. The 81 percent satisfaction lift sets the outcome metric that predicts whether the system survives. The 1-per-15-to-20 ratio sets the staffing model that makes the behavior possible at the scale of a working winery.
Notice what these numbers do not include. There is no average dollar spend per employee. No average prompt count. No average license activation rate. The reason is that those numbers are vanity at the SMB scale — they tell you about input, not about whether the play worked. The four numbers above are the ones I have found predict whether a winery will be running a working employee AI program at the end of the year or unwinding a stalled one.
Each of these has held up across separate research teams in 2025 and early 2026, with sample sizes large enough to take seriously. The MIT NANDA report is the standard reference for the failure rate. The Microsoft Work Trend Index is the standard reference for the habit threshold. The Slack/Salesforce Workforce Index gives the satisfaction finding the largest publicly-available sample I have seen. Hartz AI is the most-cited source on champion cohort sizing in the practitioner literature.
Bringing the four to the webinar
The four numbers form the spine of the 60-day play I walk through in the upcoming session. If you want to see how they translate into a calendar — week one, week three, week six, week eight, what each named person is doing on each of those weeks — that is what the Employee AI Engagement webinar covers. I will spend most of the hour on the play, the failure modes I see most often in wineries, and the Q&A. Bring your headcount and your hardest-to-staff role; that combination tends to surface the highest-leverage starting point.
If you would rather work through this for your specific winery first, that is what I do at NunnCurtis Labs. The shape of the engagement maps onto the four numbers above: pick the right task, run the 8-week build, measure the satisfaction-and-time outcome, size the champion cohort to your headcount.
Part of a 20-post series on employee AI adoption for wineries — see the full series under AI Adoption.