Behavior change does not happen in the workshop, and it does not happen in the coaching either. It happens in the first week afterward, in whether anyone’s manager asks to see the changed work. Both products exist to make that week more likely, and they are priced about nine times apart, so the question is which one your team is missing.
The vendor answer is predictable. Every page arguing that ongoing coaching beats a one-off session is published by somebody selling ongoing coaching, and its figures are either proprietary client data or a nineteenth-century memory experiment with one participant. The research that tested the question directly reaches an awkward conclusion: across two meta-analyses, neither the number of coaching sessions nor the length of the engagement predicted how much good the coaching did. That does not make coaching useless. It relocates what an enablement budget is buying.
The short answer, by situation
| Where you are | Buy this | Not this |
|---|---|---|
| Licenses bought, no training yet | One well-built session on the team’s own work, plus a short second sitting | A monthly retainer before anyone has a baseline |
| One session ran, usage faded | A named internal owner and a standing question slot | A second vendor session on the same material |
| Session taught features in one tool | Opportunity to use it, and a manager who asks | A coaching engagement to rescue a closed-skill session |
| Session taught judgment: what to delegate, how to check it | A follow-up layer, internal where possible | A single session and hope |
| No internal owner, 50+ people, real pressure to show return | A retainer, scoped to a defined handover date | An open-ended retainer with no exit |
| Trained team, two or three tasks, stable tools | Nothing further this quarter | Enablement as a subscription |
Read rows three and four together. That is the decision most budgets get wrong, because the right follow-up depends on what the first session taught, and the distinction has a name in the research.
Duration is the one variable the research rules out
Rebecca Jones, Stephen Woods and Yves Guillaume pooled 17 workplace coaching studies covering 2,267 people and found coaching produced a positive effect on outcomes overall, d = 0.36, with individual-level results at d = 1.24. Coaching works, which is settled well enough for a buying decision.
Then they tested whether more of it worked better, running linear and curvilinear models on the length of the engagement in weeks and on the number of sessions. Nothing reached significance. Their conclusion is one flat sentence: neither longevity in weeks of the coaching intervention nor number of coaching sessions moderated overall coaching effectiveness.
This is not an isolated result. Theeboom, Beersma and van Vianen had already reported an aggregated effect of g = 0.66 across coaching outcomes, and their moderator tests also found no relationship between the number of sessions and the size of the effect.
Two meta-analyses, built from different study pools, tested the thing the ongoing-coaching pitch rests on, and neither found it. Whatever the mechanism is, it is not the meter running. The components the studies have in common are goal setting, feedback and somebody expecting an answer, and a short engagement that does those things will beat a long one that does not. When a provider quotes twelve sessions instead of four and calls the difference depth, ask what each session is for.
The study everyone quotes is a study of both
If you have seen the claim that coaching produces four times the return of training, you have seen Olivero, Bane and Kopelman from 1997. Training alone raised productivity 22.4%; adding one-to-one coaching raised the gain to 88% in a single municipal agency. The number has circulated for nearly thirty years.
First, the caveats: 31 volunteers, one agency, no control group, action research, 1997. Nobody should set a budget on that, and the pages citing it never mention the sample size.
Second, and more useful: the design was not coaching instead of training. Every participant sat through a three-day workshop first. The coaching ran eight weeks afterward and consisted of goal setting, collaborative problem solving, practice, feedback, supervisory involvement, evaluation of end results, and a public presentation of the work.
That list is the only part worth keeping, and it is a specification rather than a product. Six of the seven cost nothing to run internally once somebody owns them, and the seventh, supervisory involvement, cannot be bought from a vendor at any price. A retainer supplies practice and feedback. It cannot supply managers who care about the output.
Open skill or closed skill decides whether you need the follow-up
The transfer-of-training literature has a distinction the AI enablement market has not caught up with. Blume, Ford, Baldwin and Huang pooled 89 studies and split training by its objective. Closed skills are those where the trainee is taught to respond in one particular way according to a set of rules, and their stated example is computer software training. Open skills are those where the trainee has considerable latitude in deciding a course of action.
They then asked how much the work environment mattered to whether the training transferred. The correlation was .26 for open skills and .04 for closed skills. That gap is the whole argument, and it cuts both ways:
- A session that teaches features is closed-skill training. Here is how to summarize a thread in your approved assistant, here are the six things the sidebar does. Support and climate contribute almost nothing, because there is nothing to decide: either the task comes up and people use the feature, or it does not.
- A session that teaches judgment is open-skill training. Which parts of this proposal should go to an assistant and which should not, how to tell a confident wrong answer from a right one, what belongs in a prompt that the model cannot infer. People have latitude, which means they can quietly decide not to bother, and that is where the environment carries weight.
Most corporate AI workshops are sold as the first and pitched as the second. If yours taught judgment, the follow-up layer is not optional. If it taught six features, buying a retainer to rescue it treats a design problem as a frequency problem, and the design is worth fixing first in the session design guide.
The same study puts a number on who in the environment matters. Supervisor support correlated .31 with transfer. Peer support correlated .14. A champion network is roughly half a manager.
Internal coaching beat external coaching by seven to one
The Jones meta-analysis tested four practice factors. The format did not matter: face-to-face came in at d = 0.29, blended at 0.28, statistically indistinguishable. Coaching that used multisource feedback inside the process showed smaller effects than coaching without it, 0.21 against 0.88.
The one that should change your purchase order is the coach. Effects were weaker when the coach was external to the organization, d = 0.20, against d = 1.40 for coaches employed internally, with non-overlapping confidence intervals. The authors had predicted the opposite and said so.
They did not isolate the cause. The plausible reading is that an internal coach knows the work, is there when the task is live, and is still there in six weeks. An external one arrives, is excellent, and leaves.
For an AI enablement budget this flips the usual shape. The expensive part, the outside expert, belongs where outside expertise is scarce: designing the session, teaching the judgment, building the material on your own documents. The repeated contact that keeps the habit alive belongs to someone on your payroll. That handover is a specifiable line item, costed in what a train-the-trainer program costs, and it settles in the same conversation as who keeps the curriculum once the trainer leaves.
The arithmetic nobody puts on the page
Comparison pages in this market describe both models and price neither. The public numbers:
| Purchase | Public market range | What you get |
|---|---|---|
| One hands-on session | Roughly $2,000 to $15,000 by length, headcount and customization | A designed session on your own work, plus whatever material you keep |
| Quarterly plan | Mostly Serious lists $10,000 a year for four sessions | Scheduled contact, vendor-owned agenda |
| Monthly plan | $18,000 a year for twelve sessions at the same provider, $30,000 for 24 or more | Frequency, and a standing relationship |
| Advisory retainer | Roughly $2,000 to $5,000 a month for monthly office hours; $5,000 to $15,000 for embedded weekly work | Consulting, priced as consulting |
| Per-employee enablement subscription | One network publishes $2,000 to $3,500 per employee per year | At 40 people, $80,000 to $140,000 annually |
Verify these against the vendor’s current page before quoting them internally; corporate training catalogs change quarterly.
Set the first row against the last. Those are not competing products at one price point, they are different orders of magnitude, and the evidence above does not support the gap. Before committing to the bottom row, price the version where you buy row one, hand the material to an internal owner, and spend the difference on a custom automation project where the return is measurable. What drives the first number is broken down in the one-day workshop pricing.
What to buy, in order
A sequence that matches the evidence rather than the sales deck.
- One session built on the team’s real work. BCG’s 2025 survey of more than 10,600 leaders, managers and employees across 11 countries found only a third said they had been properly trained, and regular use rose sharply at five or more hours of training with access to in-person training and coaching. Dose and hands-on contact have public evidence behind them. Session frequency does not.
- A named internal owner, agreed before the session. One person who keeps the prompts and the data rules current and runs the question slot.
- A short second sitting while the skill is warm. Agenda set by what people tried and where it broke. The timing case is made in full in the piece on whether you need an annual refresher, and the answer is weeks, not months.
- Managers asking for the changed work. Supervisor support at .31 is the cheapest high-value item here and the one most often left out of the plan. One line in a weekly check-in.
- One timed task per quarter. Pick a real deliverable, time it before and after, write it down. Without it you will be comparing opinions at renewal.
- A vendor retainer, if steps two and four failed. As a correction, with a date on which it ends.
Steps one and three need hours your team does not obviously have, and for a group that bills its time that constraint shapes the plan more than any learning theory, which is the subject of scheduling around a team that bills by the hour.
When the retainer is the right purchase
Four cases. No internal owner and no candidate, where a vendor slot is the only contact that will exist. Open-skill work at scale, where several functions are each making judgment calls about what to hand over. A moving tool estate, where the approved assistant changes fast enough that someone has to track it. And a handover you can see on paper, where the contract says when the vendor stops.
A retainer with no handover date is not an enablement plan. It is an outsourcing arrangement for a capability you said you wanted to own, which is the same confusion that puts license spend in the enablement column, as our case against buying licenses and calling it enablement sets out.
Frequently asked questions
Is a one-off AI workshop or ongoing coaching better for behavior change?
Neither on its own. Two meta-analyses found no relationship between the number of coaching sessions and the size of the effect, so frequency is not the mechanism. What has evidence behind it is one well-designed session, an internal owner who keeps contact going, and managers who ask for the changed work.
Does a longer coaching engagement produce more change?
The published research says no. Jones and colleagues tested the length of the engagement in weeks and the number of sessions, linearly and curvilinearly, across 17 studies and 2,267 people, and found no moderation. Theeboom and colleagues reached the same conclusion.
Should the ongoing layer be internal or from the vendor?
Internal, where you have a plausible candidate. In the one meta-analysis that tested it, internal coaches produced d = 1.40 against d = 0.20 for external coaches, with non-overlapping confidence intervals. Buy design and teaching from outside, and keep repeated contact inside.
What does ongoing AI coaching cost compared with one workshop?
A hands-on session sits in a market range of roughly $2,000 to $15,000. Published annual plans run about $10,000 for four sessions, $18,000 for twelve and $30,000 for 24 or more. The ongoing option typically costs three to nine times the one-off, so it needs a reason.
Is the 22% versus 88% coaching statistic reliable?
Not as a budget input. It comes from 31 volunteers in one municipal agency in 1997 with no control group, and it measured a three-day workshop followed by coaching rather than coaching instead of one.
Do employees forget 90% of training within a month?
That figure traces to Ebbinghaus and its modern replication, which used one participant memorizing nonsense syllables and says nothing reliable about workplace skill. Capability does fade without practice, but the honest version of that argument comes from the training-transfer literature, not a memory experiment with a sample of one.
How do I tell whether my session needs a follow-up layer?
Ask what it taught. If people were shown specific features in one tool, the work environment correlates .04 with whether training transfers, so a retainer will not help much. If they were taught judgment about what to hand to an assistant and how to check it, the figure is .26 and the follow-up earns its place.
Will an internal champion network substitute for coaching?
Partly, and less than it is sold as. In the transfer meta-analysis peer support correlated .14 while supervisor support correlated .31. A champion group is useful, and it is roughly half a manager. Fix the manager conversation first.
Can we run the follow-up with no budget at all?
Often, yes. A standing half hour where whoever is stuck brings the live task, a shared place for prompts that worked, and one manager question in a weekly check-in cost nothing but calendar time. That is most of the component list from the study everybody misquotes.
How soon should we expect to see changed behavior?
Within weeks on the tasks the session covered, or not at all. If nothing in anybody’s output looks different a month later, more sessions will not fix it. Go back to whether the session taught the right kind of skill and whether anyone is asking for the result.
Key takeaways
- Frequency is not the mechanism. Two meta-analyses tested whether more sessions or longer engagements produce bigger effects, and neither found a relationship.
- Internal beat external by a wide margin, d = 1.40 against d = 0.20. Buy design from outside and keep the repeated contact inside.
- Whether you need an ongoing layer depends on what the session taught: work environment correlates .26 with transfer for judgment skills and .04 for feature training.
- Supervisor support correlates .31, peer support .14. The cheapest item in the plan is a manager asking to see the changed work.
- The one-off and the subscription are three to nine times apart in price, and a retainer without a handover date is an outsourcing arrangement.
Work out which of the two you are missing before you sign either. A free AI-readiness call will tell you whether your team needs a session, an internal owner or a manager conversation, and what the first ninety days should cost.
Arthur Wandzel
