Record the demonstration and the rules. Do not record the practice. That split gives you the reusable material you are paying for without charging the session itself for it, and it is the shape of AI workshop recording that holds up six months later.
The question usually arrives for good reasons. Nine people cannot make the date. Two work a different shift. Someone in finance wants the file for next year’s hires. A recording looks like free reach on a budget line you only get to spend once.
The quiet part is that the recording is not free. It changes who shows up, how honestly they work in the room, and what the session can ask people to do. Those costs land on the day. The benefit lands only if someone opens the file, and most of the time nobody does.
The answer, segment by segment
A private session on your team’s own work is not one thing. It has four or five parts with different value on video and different costs when the red light is on.
| Session segment | Record? | Why |
|---|---|---|
| The opening task demonstration | Yes, as a short clip | One task shown once on a real document is the most reusable thing in the room |
| The data rules and verification steps | Yes, as a short clip plus a one-page written version | New hires need this in month one, and it does not date |
| Hands-on practice on the team’s own documents | No | This is where real client names, salary lines, and wrong answers appear on screen |
| Questions and the parts people got stuck on | No video, but capture the questions in writing | The questions are gold; the footage of a person struggling is a liability |
| Closing commitments and saved prompts | No video; the written artifacts are the output | A list beats a recording of a list |
Two of five segments are worth a camera, and both are short. Nothing in that table requires a three-hour file, and the whole-session recording buys you the two useful segments at the price of the three that carry a cost. The design logic behind those segments, including why the practice block is the expensive part of the day, sits in our session design guide.
What recording costs before anyone watches it
It moves attendance. The cleanest measurement comes from a university setting, where someone finally ran the numbers properly. Edwards and Clinton compared matched cohorts of 161 and 160 students before and after lecture capture arrived and published the result in Higher Education in 2019. Attendance fell sharply once the recordings were available, attendance predicted grades, and viewing the capture showed no significant relationship with grades once attendance was accounted for. The recording did not repay the attendance it cost.
Offices are not lecture halls, and your team is not a cohort of undergraduates. The mechanism transfers anyway, because it is about calendars rather than about students. A 10am session with a recording promised is an optional 10am session, and on a Tuesday with four other claims on the hour, optional loses. We see it in the room. When the invitation promises a recording, the session starts lighter than when it does not, and the people who stay away are rarely the ones who needed it least.
It changes what people are willing to try. This matters more in an AI session than in almost any other training, because the exercise is to put your own unfinished work on a screen and get it wrong in front of colleagues. Houtti and colleagues interviewed 22 working professionals for a 2023 paper at the ACM conference on computer-supported cooperative work and found recording to be one of four features that reshape how people behave in a meeting. Participants turned their cameras off, changed their display names, and moved from speaking to typing in chat. One described not wanting to be immortalized doing something and becoming a meme. Several reported self-consciousness about permanence even while assuming nobody would ever watch.
Carry that into a room where the ask is to paste a real client email into a chat window and find out what the model gets wrong. The person most in need of the three hours is the one most likely to go quiet. You can buy a recording or you can buy candor, and in a hands-on session candor is the product.
It narrows the material. A facilitator who knows the file will be stored somewhere permanent stops working from live documents and starts working from sanitized samples. The session then teaches a generic workflow rather than yours, which is the exact failure mode that makes catalog courses weak, as we covered in the case for building the curriculum from your own work.
The three-hour file nobody opens
Assume you record anyway and the day goes fine. What is the file worth?
Guo, Kim and Rubin studied 6.9 million video sessions across four edX courses and presented the results at ACM Learning@Scale in 2014. It was the largest study of instructional-video engagement at the time. Engagement falls off with length, viewers rarely finish anything over nine minutes, and the authors recommend six minutes or less. Among the formats they compared, high-quality recordings of classroom lectures were the weakest, which is precisely what a session recording is.
So the artifact you are negotiating for is a long file in the weakest format, aimed at colleagues who did not have 90 minutes for the live session and now need 180 for the video.
Then there is storage. Microsoft Teams expires meeting recordings 120 days after creation by default unless an admin changes the policy. The permanent onboarding asset that justified the whole request is often deleted before the next hire starts, and nobody notices because nobody opened it.
The vendors who sell you replay access know all of this and sell it anyway, because for a seat-based product it is the right call. Section lists its AI crash course at $195 for 90 minutes of live instruction with recordings and materials afterward, and self-paced catalog libraries are recordings by definition. Replay is a sensible deliverable when the buyer is one person who may miss a webinar. It is a weak deliverable when the buyer is a company paying $2,000 to $15,000 of market rate for a private session built on its own documents.
Replay is not practice
The argument for recording rests on an assumption nobody states out loud: that watching the session is a reasonable substitute for attending it. A long line of memory research says it is not.
Roediger and Karpicke ran the canonical test in Psychological Science in 2006. A week after studying a passage, people who had practiced recalling it held about 61% of the material against 40% for people who had reread it for the same amount of time. Doing the retrieval beats reviewing the content, and the gap widens as time passes.
An AI session is retrieval practice with a facilitator correcting you. The colleague watching the video gets the review condition and none of the correction. That person is not a late attendee. That person read the notes.
This is the same confusion that makes license dashboards look healthy while nothing changes on Tuesday, which is the argument we make at length in our case against buying licenses and calling it enablement. Attendance counts and view counts are both easy to collect and neither one tells you whether anyone can do the work.
The practical consequence: if nine people cannot make the date, the answer is a second session, not a video. A repeat of a three-hour block costs real money, though it is normally quoted below the first session because the material already exists. It is also the only option that builds capability in those nine people, rather than a record that they were sent a link.
Record these five things instead
Ask for durable artifacts rather than footage. Ranked by what still gets used a quarter later.
- Short task clips, two to six minutes each. The facilitator demonstrating one task end to end on a real document, screen only, no faces. Four or five of these cover a function’s core workflows and land inside the engagement limits that the video research supports. These are the clips a new hire opens.
- The prompt set the team wrote and tested. Not a handout of 50 generic prompts. The eight or ten that this team wrote on their own documents, with a note on what each one is for and where it broke. The ones people wrote themselves are the ones still in use next quarter.
- The verification checklist. What to check before anything goes to a client, what the tool does when it does not know, which numbers get confirmed against the source. One page, written by the facilitator from what went wrong in the room.
- The data rules. Which fields never get pasted into which tool, which tools are approved, who to ask when it is unclear. One page, and the item most likely to be reread.
- The before-and-after of one real task. The Thursday report as it was produced in September and as it is produced now, with timings. This is the artifact that unlocks the second budget, and it is the one nobody thinks to ask for.
Name all five in the scope of work and the recording question mostly dissolves, because what the finance team wanted was never the footage. It was evidence and a reusable asset, and a page of data rules plus a two-minute clip of the demonstrated task serves both better than a video of a dozen people practicing.
When recording the whole session is the right call
Four cases where recording the full block is worth its costs.
A genuinely distributed team. When people are in six time zones and no single hour works, the session is already remote and the camera is already on. Record it, keep cohorts small, and accept that the replay group will need a live clinic of its own. The format differences are set out in our comparison of on-site and remote delivery.
Training your internal trainers. If the point of the engagement is that your own people run the session next time, the full recording is a training manual for them rather than a learning asset for the room, and the economics are different. We work through those in the train-the-trainer cost breakdown.
A rollout past 100 people. When the same session runs eight times, recording the first delivery lets you fix the pacing before the other seven, and it gives the second facilitator a reference to calibrate against.
A documented compliance requirement. If your regulator or your insurer requires evidence that staff were trained on approved use of these tools, record the rules segment and keep the attendance record. That is a compliance artifact, not a learning one, and it can be 12 minutes long.
Outside those four, the full recording is a habit rather than a decision.
Put the recording terms in the agreement
Settle the AI workshop recording terms before the session, because afterward you are negotiating over a file that already exists.
Who owns it, and what may you do with it? Many trainers record for their own library, their marketing, or their next client. Get it in writing that recordings of your staff working on your documents are yours, are not used in the vendor’s marketing, and are not reused with other clients. If the vendor retains a copy, ask for how long and under what access controls. Ownership of the material a session produces is a wider question than the video, and the procurement checklist covers the rest of it.
Notice and consent. Faegre Drinker’s 2025 guidance on recording and AI meeting assistants is the practical summary: put notice in the invitation, announce it verbally at the start, obtain active consent where the platform supports it, restrict access to named people, and set a retention and deletion schedule up front. Several states require consent from all parties to a confidential conversation, so a quiet default-on recording is a poor idea even inside your own building. Check with counsel rather than with your AV vendor, whose incentive runs one way.
What is in frame. If you do record, record the facilitator’s screen and not the room. Faces, laptops, and the whiteboard where someone wrote a client’s name are what turn a training file into a problem.
Who maintains it. Name the owner, the storage location, and the review date, or the recording expires into a dead link. A vendor willing to name the owner and the review date in writing is promising something checkable, which is the subject of our piece on what a credible provider will commit to.
Frequently asked questions
Should you record an AI workshop?
Record the demonstration and the rules segments as short clips, and do not record the hands-on practice. The practice block is where real documents and wrong answers appear on screen, and where recording suppresses the willingness to try that makes the session work. Two short clips plus written artifacts give you everything a full recording would, without the cost on the day.
Does recording an AI workshop reduce attendance?
Yes, in the closest setting anyone has measured properly. Edwards and Clinton compared matched student cohorts before and after lecture capture and found attendance fell sharply once recordings were available, with viewing showing no significant relationship with results once attendance was accounted for. A promised recording turns a booked hour into an optional one.
Can people who miss the session just watch the recording?
Not with the same result. A week after studying, people who practiced recall held about 61% of the material against 40% for those who reread it. Someone watching the video gets the review condition with no facilitator correcting their work, so they finish informed rather than able. Run a second session for the group that missed the first.
How long should a recorded AI training clip be?
Two to six minutes per task. The largest study of instructional video, covering 6.9 million sessions, found engagement falls with length, viewers rarely finish videos over nine minutes, and recordings of classroom lectures were the weakest format compared. Four short clips beat one long file by a wide margin.
Who owns the recording of a corporate AI workshop?
Whatever the agreement says, which is why it belongs in the scope of work. Ask for your recordings to be your property, excluded from the vendor’s marketing, and not reused with other clients. If the trainer keeps a copy, pin down the retention period and who may access it.
What consent do we need to record a training session?
Notice in the invitation, a verbal announcement at the start, active consent where the platform offers it, restricted access, and a written retention and deletion schedule. Several states require all parties to consent to recording a confidential conversation, so treat silence as a refusal and confirm your position with counsel.
Is it cheaper to record once than to run the session twice?
On the invoice, yes. On outcomes, no. A recording produces viewers and a second session produces practitioners, and only one of those shows up in the work. A repeat delivery is normally quoted below the first session because the material already exists, and it is the right spend if those people are expected to change how they work.
What should we keep from the session instead of the video?
Five things: short clips of each demonstrated task, the prompts the team wrote and tested on their own documents, a one-page verification checklist, a one-page set of data rules, and the before-and-after of one real task with timings. Name all five in the scope of work before the session runs.
Where do recordings go wrong in practice?
They expire or they go unowned. Microsoft Teams deletes meeting recordings 120 days after creation by default unless an admin changes the policy, so the onboarding asset that justified the request is often gone before the next hire arrives. Name an owner, a storage location, and a review date, or do not bother.
Key takeaways
- Record the demonstration and the rules. Do not record the practice. Two of a session’s five segments are worth a camera and both are short.
- Recording moves attendance, and the trade is not favorable. Matched-cohort research found attendance fell sharply once capture was available, attendance predicted results, and viewing did not make up the difference.
- A hands-on AI session runs on candor. Interviews with 22 professionals found people turn cameras off, rename themselves, and retreat to chat when a meeting is recorded, which is the opposite of what the practice block needs.
- A three-hour file is the weakest video format aimed at the people with the least time. Engagement research says two to six minutes per task, and the default retention policy on Teams deletes the file after 120 days anyway.
- Ask for five durable artifacts instead: task clips, the team’s own tested prompts, a verification checklist, a page of data rules, and the before-and-after of one real task. Settle ownership, consent, and retention in the agreement, not on the day.
Decide this before you book, not in the room. A free AI-readiness call will tell you which segments of your session are worth keeping on video, what the written artifacts should contain, and whether a second live session beats a replay for the group that cannot make the date.
Arthur Wandzel
