The first hour of a Monday morning in a tasting room is mostly cleanup from the weekend and prep for the week. Glassware racks to reset and the open-bottle list to walk through with the cellar before the staff who come on at eleven arrive. There is a familiar shape to it. The shape does not announce itself, which is why most of the work inside it goes unexamined for years.
Two months into a real AI rollout, the same first hour looks different in small ways that compound. The shape is the same. The contents have shifted. The cleanest way to show that is to walk through one. The Monday I am about to describe is at our fictional demo winery, Not Really Wines — a working showcase at notreallywines.vercel.app of the kind of vault and workflow we build with real clients. Riley Park is the tasting bar lead. She started her rollout in early March, which puts her at the back end of the eight-week habit window Microsoft's research describes — three uses a week for seven to eight weeks is the point where a tool stops feeling like an assignment and starts feeling like a hand. Riley is on the other side of that line. This is what that side looks like.
7:15 — visit recaps for the weekend
Riley opens the side door and walks the bar before she starts anything else. The Saturday flight cards from the host station are still stacked next to the POS — guest names, the wines they tried, the staff initials of who poured, a few hand-written notes in the margin. There are twenty-three of them. Before the rollout, she would have transcribed five or six of the most promising into the CRM on Tuesday afternoon and the rest would have stayed in the pile until somebody bussed it into the recycling.
This morning, the cards go through the AI workflow Tessa Brennan, the hospitality director, sat down with her in week two to build. Riley snaps a photo of each card with the tasting bar's iPad, drops the photos into the visit-recaps folder she keeps inside the NRW vault, and opens the recap prompt she has been refining for six weeks. The output is a short recap per visitor — the wines they responded to and a one-line follow-up flag — formatted the way Tessa wants it. The first draft takes the model about three minutes for all twenty-three. The pre-rollout version took Riley twenty-five minutes for the six she got to. The other seventeen visits never made it into the CRM at all.
She reads the recaps before she accepts them. Two of them she rewrites. One she discards because the model misread the handwriting on the wine list. The model is not correct on the first try, and the workflow does not depend on it being correct on the first try. What the workflow does is move the version Riley reads, edits, and accepts into the CRM by seven-thirty on Monday rather than partially into the CRM by Wednesday afternoon. The same job, on a different clock.
7:35 — the follow-up note Maren cares about
Riley pours herself a coffee and pulls up the four visits from the weekend that Tessa flagged for a follow-up note. The flagged visits are the ones where a guest asked about a wine NRW does not currently make, or where a guest mentioned a private event they were planning. Before the rollout, the follow-up note for those visits would have waited until Tessa got a chance to draft it, which in practice meant two weeks. Two weeks is past the window where a follow-up still reads as warm. Most of those notes got drafted and most of them landed late.
This morning Riley drafts the four notes herself, with the voice file she and Tessa built in week two. The voice file lives inside Riley's starter folder, the same shape I described for the wine club concierge in the AI starter folder, line by line — different role, same skeleton. It is a folder of eight real follow-up notes Tessa has written over the years, plus a paragraph Tessa wrote describing what the notes are trying to do — warm but unhurried, with one specific reference to the wine the guest responded to and a soft invitation rather than a hard ask. Riley feeds the four flagged visits to the prompt and reads the drafts. Three of the four are close enough to send after a line of edits. The fourth she rewrites from scratch because the guest's situation was unusual enough that the model defaulted to a generic warmth that does not fit. She sends all four before eight, while she is still on her first coffee.
Same-day follow-up was the metric Maren Holloway, the CEO, was quietly tracking. Two months in, the answer is yes for almost every flagged visit. That is the change in the data Maren keeps coming back to. The CRM is showing the kind of touch volume she has been asking for since the day she opened the doors.
8:00 — the pattern the manager would never sit down to do
This is the part of the workflow that does not show up on any task list. Once a week, usually on Monday morning before the bar opens, Riley runs a short prompt across the recaps from the past two weeks. The prompt asks one question: what is showing up in conversations with visitors more than once that is not already on our radar? The output is rough. She reads it like she would read a hunch from a colleague — sometimes it lands, sometimes it does not.
This week the prompt pulls something out. Three different visitors over the weekend asked about a low-and-no wine pairing dinner — not a tasting flight, a sit-down dinner with food courses paired to the non-alcoholic wines. Riley would not have noticed the cluster on her own. She was too close to each individual visit. The pattern is not load-bearing yet. It is one weekend. But she opens a note in the vault titled "pairing dinner — three asks 5/10–5/11" and drops the three recap snippets in. If the same ask shows up next weekend, she has the start of a case to bring to Tessa and Maren.
The pattern-recognition use case is the one I find hardest to predict in advance with a client, because the patterns that matter at a specific winery are local and seasonal and the prompt that surfaces them has to be built in conversation with the staff who already feel the floor. Six weeks ago Riley would have said the prompt was a waste of time. By week six she had started looking forward to running it on Monday mornings. That shift — from compliance to curiosity, on a recurring task — is the clearest behavioral signal that the rollout has stuck.
8:30 — the question Tessa is going to ask
Around eight-thirty Riley swings into the back office to drop off the morning's recaps. Tessa is on her laptop with a cup of tea and the schedule open in front of her. The Monday morning conversation between the two of them used to start with Tessa asking what came in over the weekend. Now it starts with Tessa asking what Riley noticed.
The change in the opening question is the part of the rollout that is the easiest to miss and the one I would point a winery owner at if they asked me what to look for at month two. The recaps are now an input Tessa can read at her own pace instead of a backlog she has to push Riley through, and the same-day follow-ups have moved the conversation from "what did we miss" to "what did we hear." The pattern note that sometimes turns into something is a column of insight that did not exist eight weeks ago. None of those changes are large, and all of them are durable, because the workflow that produces them is one Riley enjoys enough to keep running on her own.
This is also the morning where Tessa stops asking Riley to show her the prompt on the iPad. By month two she has watched it enough times. Today she pulls Riley's iPad over and runs the recap prompt herself on a card Riley has not gotten to yet. She edits the output, accepts it, and hands the iPad back. That is the second-employee moment the right first AI champion is supposed to produce, and it is happening organically because Tessa has been watching Riley do useful work with a tool for eight weeks. The play does not need a second-user kickoff meeting. The handoff is in the gesture.
What month two does that month one does not
A month-one rollout is mostly a learning loop. Riley spent the first four weeks figuring out where the tool fit and when it produced output she had to rewrite from scratch. She used it three times a week with Tessa nudging her once a week to keep going. The output was uneven and the time savings were inconsistent.
Month two is when the loop closes. The prompts she uses are no longer the ones she copied from the starter folder — they are versions she has edited four or five times to match the way her tasting room runs on a Saturday night. The follow-up notes are signed off without Tessa reading them because Tessa has seen enough of them that she trusts Riley's edits. The Monday pattern prompt has shifted from a chore to a piece of the routine. The tool has stopped being a foreground object and become part of how the work gets done. That move from foreground to background is the rollout outcome that matters. Everything before it is rehearsal.
The reason this lands for tasting room teams faster than for other functions is that the work already had a digital tail — the recaps and the follow-up notes inside the CRM — that was getting written shorter and slower than it needed to be. AI did not invent a new task for Riley. It cleared the bottleneck on a task she already knew how to do. The three roles where this pattern repeats most cleanly are covered in the three winery roles where AI lands fastest post. The tasting room is the one I would point a winery at first, because the volume of weekend touch points is where the time savings stack the fastest.
What this Monday morning is not
A few caveats so the sketch is honest. Riley is not running a closed-loop autonomous workflow. Every output the model produces passes through her hands before it goes anywhere a guest will see. The reason the workflow is fast is that her editing pass on a draft she half-likes is much faster than her drafting pass from a blank page would be. If the model were producing copy that went out unread, the math would change and the trust would erode. The role of the staffer at the keyboard is the part of the workflow that is doing the load-bearing work.
Riley is also not using AI on the parts of her job that are physically anchored. The cellar walk-through with the production assistant looks like it did eight weeks ago. The conversation with a regular at the bar looks like it did eight weeks ago. The judgment call about whether to pour a guest a comp glass on her third visit this month is the same call Riley would have made in February. The job is mostly the same. The paper trail behind the job is what has changed.
And Riley is not getting paid more for any of this, which is the thing Maren and I talked about at the end of week three. The compression Riley is producing has shown up as her being able to leave on time more often and as the parts of her job she least enjoys taking less of her week. The promotion question is on the table for next quarter, framed as a hospitality lead role with the kind of pattern-recognition responsibility she is already starting to exercise. The framing matters because the person who clears the eight-week threshold is the person Maren wants growing into a bigger hospitality role, not the person she wants buried inside an automation conversation.
A note on the webinar
The Monday morning I have described is what the 60-day play produces on the back end. The reason I keep writing about the tasting room is that it is the function where the time savings are easiest to point at and the function where the second-employee dynamic shows up earliest. If you are getting ready to run this play at your winery, or you are six weeks into one and trying to figure out what month two should look like, the webinar walks through the full play including the timing question — which is why running this two months before harvest is a bad idea and where in the calendar it fits instead.
Check the webinar schedule for the next date and to register.
Part of a 20-post series on employee AI adoption for wineries — see the full series under AI Adoption.