Back to Guides
Enterprise Software 15 min read

When a great workshop stops working: the missing follow-up architecture

When a great workshop stops working: the missing follow-up architecture

A workshop that worked and then stopped working is not a bad workshop. It is a workshop with nothing built behind it. The session had an owner, a budget line, a room and a date. The ninety days after had none of those, so the capability that existed at 4 p.m. on Thursday had nowhere to live on Monday.

That failure is structural, not motivational. Search for AI workshop follow up and you get email templates: send a recap within 24 hours, nudge again in three days, attach the slides. What you do not get is the short list of standing things that have to exist on a calendar before the trainer leaves the building.

What the decay looks like from the inside

The shape is consistent enough to plan around. Week one is loud: people try what they practiced, a few tell a colleague, and the admin console makes the workshop look like a success.

Week two is where the first real obstacle lands. Someone gets an output that is subtly wrong and has no one to ask. Someone else hits a question about what data they may paste in. Someone’s manager wants the deliverable by eleven, so they write it the old way, because the old way takes a known forty minutes.

By week four, the people still using an assistant daily are the ones who already were before the session. The rest have not decided against it. They have never had a second occasion where using it was the obvious choice, and the prompt they saved is three folders deep in a drive nobody opened.

None of that is about enthusiasm. Each step is an unanswered question meeting a deadline, and a follow-up structure is the set of places those questions can go.

Why AI fluency fades faster than a software rollout

Buyers plan AI enablement the way they planned a CRM rollout, then wonder why the playbook underperforms. The learning-transfer research says these are not the same kind of problem.

In their meta-analysis of 89 studies for the Journal of Management, Blume, Ford, Baldwin and Huang found that transfer predictors behave differently for open and closed skills. Closed skills have one correct procedure: enter the opportunity, attach the quote, move the stage. Open skills have no single right answer and must be adapted each time. Motivation and a supportive work environment related more strongly to transfer for open skills than for closed ones.

Using a language model well is an open skill. There is no correct keystroke sequence for turning a messy client thread into a status update, which is why it cannot be installed through a help-desk article. So the work environment decides more of whether the skill survives, and that is the part a one-day session cannot touch.

The claim is not that people forget. It is that an adaptive skill needs repeated contact with real, varied work before it settles.

The six standing objects

A follow-up plan is not a list of tactics. It is a short list of objects that exist whether or not anyone feels motivated, each with one named owner. Six is enough, and a 50 to 500-person company can staff all six without hiring.

Object What it is Owner Cost What breaks without it
The second session A 60–90 min working session, booked before the first ends Whoever booked the workshop 1.5 hrs/person, once The skill never meets week-two work
Office hours A 30-minute open slot, same time weekly Two internal champions 2 hrs/mo Every obstacle becomes a private dead end
The prompt library One folder, one naming rule, one person who prunes it A named curator 1–2 hrs/mo Saved work is unfindable, so it gets rewritten
Champions Two or three respected peers with a written mandate Their own manager 3–4 hrs/mo each Help depends on the trainer, who has left
The manager’s sentence Naming out loud which task should now be done differently Each functional manager 10 min The old way stays the safe way
The 90-day checkpoint A review against what was agreed before the workshop The budget holder 2 hrs, once Nobody learns whether it worked

Add the hours: 15 to 20 person-hours a month, concentrated in two or three people. Cost is almost never why follow-up does not happen. The owner column is. There is no Chief AI Officer at this size and usually no training function. The workshop was arranged by a COO, a chief of staff or a VP with a day job, and follow-up that is “everyone’s job” is nobody’s.

Why the second touch matters more than the first

Almost every published plan puts the follow-up touch close to the session: a nudge the next day, a check-in at the end of week one. If you need the capability in ninety days, that is the wrong shape.

In their synthesis of 839 assessments across 317 experiments, Cepeda and colleagues found that the best gap between practice episodes grows with how long the material has to last. Short gaps win when the test is tomorrow; longer gaps win when it is months away.

On a run sheet, that puts the second practice occasion two to three weeks out. At two days the room has not yet hit anything hard enough to learn from. At three weeks everyone arrives with a real failure and a question that matters.

It is also why a day-after recap email is not a touch. Reading is not practice: Dunlosky and colleagues rated practice testing and distributed practice highest-utility, rereading near the bottom. The same test applies inside the session, which is why our guide to what a workshop should have people do specifies every exercise by the artifact it leaves behind.

The 90-day calendar

The cadence those findings produce, sized for one cohort:

When What happens Length Who runs it
Day 0 Each person names one task and a date 10 min Facilitator
Day 2 Library folder created from the prompts saved in the room 30 min Curator
Week 1 First office hours; agenda is whatever broke 30 min Champions
Week 3 Second working session: one failure and one win each 90 min Facilitator or lead champion
Weeks 4–8 Office hours weekly; library pruned monthly 30 min/wk Champions, curator
Week 6 Managers revisit their sentence one-on-one 10 min Managers
Week 10 Third session, narrower: one role, one workflow 60 min Facilitator
Week 12 Checkpoint against the intake agreement 2 hrs Budget holder

Two notes. The weekly slot stays on the calendar even in a week when nobody comes, because a slot cancelled once stops being a place people bring things. And the week-three session is what you protect when the quarter gets busy; cancelling it to keep the day-two email is backwards. The agreement the checkpoint measures against gets written during intake, which is why our account of what a team owes the facilitator beforehand insists on writing it down.

The manager is the binding constraint

Of the six objects, the one that costs ten minutes does the most work, and it is skipped most often.

BCG’s AI at Work 2025 survey, published June 26, 2025, found regular AI use among frontline employees stalled at 51%, with only about a third saying they had been properly trained. Two findings speak straight to follow-up: regular usage is sharply higher among employees who get at least five hours of training, and higher again where that training includes in-person sessions and coaching. The sharpest number is about leadership rather than curriculum. Without strong leadership support, 15% of employees felt positive about generative AI; with it, 55% did.

Strong support is not a town hall. It is a direct manager naming one recurring deliverable, saying they expect to see it produced differently from now on, then asking about it in the next one-on-one. Without that sentence a careful employee reads the situation correctly: the old method has never been questioned, so it is the safe one under deadline. No office-hours program outranks a manager’s silence, and from where the employee sits, the manager is the work environment.

What office hours should contain

Office hours fail predictably: they turn into a general discussion about AI. Three rules keep them useful.

Bring work, not questions. Entry is a real document, prompt or output that went wrong. Hypotheticals are endless and teach nobody.

Thirty minutes, same slot, no agenda. The value is that it is cheap to attend. A booked hour with a topic becomes a meeting to decline.

Answers go into the library. If a problem took ten minutes to solve, the solution belongs in the folder under a name the next person would search for. Otherwise the same question gets solved four times.

Staffing matters too. Office hours run by the outside facilitator work for a month and then stop, because the month-two questions are about this company’s documents, approval chain and data rules.

What a champion is allowed to do

A champion is not a volunteer with a badge. Three things go in writing, or the role decays into unpaid help-desk work the person quietly drops.

A mandate: the hours are part of their job, agreed with their manager. A boundary: they answer questions about using the tools well and do not rule on what data may be entered, which belongs to whoever owns the written rule. A term: three months, renewed on purpose, so it can end without anyone being told they failed.

Choose for respect over skill. The person whose example changes behavior is the one whose work the room already takes seriously, not the one with the highest tool fluency. Two champions in a 40-person department is enough; one is fragile, since a single vacation ends the program.

What follow-up costs and what to put in the contract

Hands-on AI workshops sit in a market range of roughly $2,000 to $15,000 per session, depending on cohort size and how far it is tailored to the team’s own documents. Follow-up is usually quoted separately or not at all, which is where the value leaks out. Get four things in writing before signing:

  1. A second session, dated. Not “a follow-up to be scheduled”. A date two to three weeks out, in the same document as the first.
  2. Who staffs office hours after month one, and the cost if that is the provider. If the answer is internal champions, name them and confirm their managers agreed.
  3. What the provider hands over. Exercise prompts, the library structure, and the facilitator’s notes on what the room struggled with. That last one is the most useful artifact and is rarely asked for.
  4. The 90-day review, with its measure named up front. A named workflow, produced a named way, by a named group.

A provider who treats all four as reasonable is selling capability. One who calls them scope creep is selling a day. The wider case for reading these engagements as capability rather than content sits in our argument against treating license rollout as enablement.

Three follow-up patterns that do nothing

The recap email. It documents the session for the people who were in it. Useful as a record, worthless as practice, and it takes the slot a dated second session should hold.

The assigned course. Buying every attendee a self-paced AI course turns an adaptive skill back into passive content, and completion rates on unrequested assignments tell the rest. Fine for people who ask; not a structure.

The open channel. A chat channel with no owner, no ritual and no seeded content goes quiet in about two weeks, after which its silence reads as proof that interest faded. A channel works only as the surface of something else: office hours posted there, library additions announced, week-three wins collected.

All three are cheap, which is why they get chosen, and all three fail the same test: nobody has to do anything on a given day.

How you know it held

At week twelve, three questions settle it, none of them a usage dashboard.

Can a named group produce a named deliverable the new way, on a deadline, without help? Did the library grow past the prompts saved in the room, added to by people who are not champions? Can each manager name one task their team now does differently?

Three yes answers mean the capability moved; two mean the structure held in parts, and you can see which part. What makes this measurable is decided before the session, the argument of our workshop design playbook; the shape of a session worth following up on is walked through in our anatomy of a 90-minute session.

If you ran a workshop that worked and are watching it fade, the fix is not another workshop. It is six named owners and four dates. Book a free AI-readiness call and we will map the structure onto the AI enablement your team has already paid for.

Frequently asked questions

How long after an AI workshop should the first follow-up happen?

The first working session belongs two to three weeks out, not the next day. The spacing research finds the best gap between practice episodes grows with how long the skill must last, and by week two people arrive with real failures to work on. A light check-in in week one is fine, but it is not the follow-up.

What should an AI workshop follow-up session cover?

Participants’ own failures and wins since the session, not new material. Each person brings one output that went wrong and one that worked, and the room works through them. New content in the second session signals the first did not land.

Who should own AI workshop follow-up at a company with no training team?

Split it three ways. The person who authorized the workshop owns the calendar and the 90-day review, two internal champions own office hours, one curator owns the prompt library. The split is what makes it survivable when all three have other jobs.

How many office-hours sessions does a team need?

Weekly for the first eight weeks, 30 minutes, same slot. Attendance will be uneven. The slot matters more than the turnout, because a cancelled slot stops being somewhere people bring a problem.

What does AI workshop follow-up cost?

Internally, about 15 to 20 person-hours a month, held by two or three people. Externally, providers either include a second session in the $2,000 to $15,000 market range for hands-on workshops or quote it separately, so ask which before signing.

Why does usage look fine in the admin console but nothing changed?

Seat activity measures logins, not capability. People open an assistant, ask something low-stakes, then go back to the old method for work that matters. The test is whether a named deliverable is now produced a different way under a real deadline.

Should managers attend the same workshop as their teams?

They should attend something. A manager who has never used the tools cannot name the task they expect to see done differently, and that sentence carries more of the plan than anything else. A shorter manager-only session works if a joint one is impractical.

Is follow-up still needed if the team already used AI before the workshop?

Yes, with different weights. Prior users need the library and the champion role more than office hours, because their obstacles are about sharing and scale rather than first contact. Same six objects, different emphasis.

Key takeaways

  • A workshop that fades did not fail. Nothing was built to catch it, and the catch is six objects with six named owners.
  • Treat LLM work as an open skill. Work environment matters more for open skills than for procedural software, so one session cannot be the program.
  • Put the second practice session two to three weeks out. The best interval grows with the retention horizon, and a day-two email is not practice.
  • The manager’s one sentence does more than any other ten minutes in the plan. Without it, the old method stays the safe method under deadline.
  • Contract for the second session, the handover and a named 90-day measure before signing.

Last Updated: Nov 13, 2026

DJ

Dirk Jan van Veen, PhD

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles