Back to Guides
Enterprise Software 15 min read

What actually sticks six months after an AI workshop

What actually sticks six months after an AI workshop

Six months after a one-day AI session, expect three or four narrow behaviors to have survived and most of the rest to be gone. That is the normal result, not a verdict on the trainer. What persists has a shape: a substitution embedded in a deliverable produced every week, a saved artifact reachable without remembering anything, a changed default in a tool already open. What evaporates has a shape too: technique with no recurring trigger, breadth, and anything requiring recall of what was said in the room.

AI workshop retention is a question nobody answers honestly, partly because the phrase hides three questions. It can mean whether people kept the skills, whether the company kept the people, or whether the vendor keeps your data. This piece answers the first, the one a training budget is judged on.

Three questions hiding in one phrase

Search the term and the results split three ways without acknowledging each other. Software companies with “Workshop” in the product name publish data-retention policies. Human-resources writers ask whether AI upskilling stops people from quitting. Learning-and-development blogs restate the forgetting curve and sell reinforcement packages.

The third is the right question and the worst-served. The pages answering it quote percentages with no study attached: one vendor PDF contrasts 23% sustained adoption for generic training against 67% for role-specific training and names no source. Another publishes a curve reaching 97% retention at two years, a projection from a sales model rather than a measurement of anyone.

The honest starting point is that AI-specific evidence at six months is thin. What exists is a long literature on training transfer, one large 2026 survey that splits AI users at the six-month line, and one published experiment that took the tools away to see what people had absorbed.

What the transfer research says about month six

The most-documented six-month figure in corporate training comes from Saks and Belcourt’s survey of training professionals across 150 organizations, published in Human Resource Management in 2006. Reviews citing it agree on the twelve-month end state: roughly a third of trainees are still applying what they learned a year later.

The fuller chain quoted everywhere, 62% applying right after training, 44% at six months, 34% at a year, travels through secondary summaries more than through the paper itself. Treat it as an order of magnitude, not a measurement of your team. The useful reading is the slope: roughly two in five still using the material at six months, with the decline steepest early. The rival claim that only 10% of training ever transfers deserves retiring, since it traces to a 1982 trade-magazine article where a training manager offered it off the cuff with no study behind it.

Two mechanisms explain most of the spread between teams, and both are specified before a session rather than rescued after one. Cepeda and colleagues’ synthesis of distributed practice, in Psychological Bulletin, found that spacing practice beats cramming it, and that the longer the horizon you care about, the longer the gap between sessions should be. If you care about month six, sessions a week apart beat sessions an hour apart.

Roediger and Karpicke’s work on test-enhanced learning, in Psychological Science, is sharper. At a five-minute delay, restudying material beat being tested on it. At two days and a week, the order reversed and retrieval won clearly. Rereading also made people more confident while leaving them with less. A workshop ending in a recap slide produces the confident version. One ending with people producing work under observation produces the other.

Why month six is the checkpoint that matters

Month six is where novelty stops paying. BCG’s fourth annual AI at Work survey, published in June 2026 with 11,749 respondents across twelve markets, cut its sample by how long people had been regular AI users and found the drivers invert.

Among those regular for under six months, the largest driver of enjoying the work was cognitive stretch, at 43 percentage points. Past a year, that fell to 21 points while a clear AI strategy rose to 30 and clear guidance on saved time to 25. BCG’s own phrasing is that the honeymoon does not last unless leaders supply strategic clarity.

Read as a retention finding, that is a warning about what your month-three dashboard was measuring. Early use is partly powered by the interest of something new, and that fuel runs out in the first half-year. What is left afterward runs on structure: a real reason, a named deliverable, a manager who asks.

The same survey shows where programs lose the thread. 74% of frontline employees now count as regular AI users, up 23 points in a year, and 42% of those save a full day or more each week. But 66% get limited or no guidance on what to do with the time, and 58% do not redirect it toward anything strategic. Meanwhile 88% expect to need major upskilling within five years while 36% feel properly trained, both unchanged from 2025.

That gap is the argument for treating AI enablement as a program with checkpoints rather than an event, which is the case laid out in our guide to proving capability moved.

What survives and what evaporates

The split is predictable enough to plan around.

Survives to month six Fades by month six
One substitution inside a weekly or monthly deliverable A technique with no recurring trigger
A saved artifact: a prompt in a shared doc, a template, a project file A clever approach demonstrated once on stage
A changed default in a tool already open all day A tool requiring a new tab and a decision
Knowing which task to hand over and which to keep Knowing a tool’s full feature list
The habit of checking output before sending it Any specific interface walkthrough
A named person people ask when something looks wrong A help channel nobody posts in

The pattern underneath: what survives is attached to something that happens on its own. A weekly report arrives whether anyone remembers the workshop or not. A saved prompt needs no recall, a default needs no decision. Everything in the right column needs someone to remember, intend and choose, six months after anyone last mentioned it.

This is why breadth is the expensive mistake. A session covering fifteen capabilities across four products leaves fifteen things with no trigger. One covering a single substitution, practiced on the team’s own documents until it is boring, leaves one thing attached to a recurring event. The second is less satisfying on the day and is most of the difference later.

Quality needs its own look. Faster output that nobody checks is not what you bought, and the instrument for scoring capability rather than enthusiasm is in our framework for scoring a team before and after.

The one-day withdrawal test

The cheapest way to find out what a team retained is to take the tools away for a day and watch what breaks.

Atlassian’s Teamwork Lab ran exactly this in June 2026 and published the results. Around 250 talent-acquisition staff were invited and 123 analyzed from daily surveys, each serving as their own comparison across one AI day and one AI-free day.

The findings are a retention report in disguise. 82% felt the urge to reach for AI on the no-AI day, nearly one in four out of habit. 47% said tasks felt uncomfortably hard without it, and 46% said their non-AI skills had gone rusty. On the AI day, priority work felt 44% easier and stress ran 22% lower, while 44% of recorded reflections flagged the work of checking and fixing output. After the team compared notes, 89% reported a clearer shared sense of when to use AI and 92% were more confident about when not to.

Repurposed as a six-month audit, the design costs one day and answers what a dashboard cannot: which tasks have genuinely moved, because their absence is felt; which use is habit rather than judgment; and which skills have atrophied, the finding nobody plans for and the one most worth having early.

Three rules make it work. Announce it as an experiment, not a test, or people will perform. Have each person write down what they could not finish and what they did instead, because that list is the data. Hold the comparison conversation the same week, since Atlassian’s largest effects came from the discussion rather than the days themselves.

The six-month review, in numbers

Beside the withdrawal day, four numbers are enough. Each needs a threshold written down before you look and one named owner, because a review three people own is a review nobody runs.

Measure How to get it Healthy at month six What a weak result means
Repeat use on the named behavior Manager confirms it happened in 3 of the last 4 weeks Above 40% of the trained group The behavior had no recurring trigger
Spread across teams The same number, per team, never averaged No team below half the group rate A manager problem, not a capability one
Quality of one real sample 10 minutes per person, scored by their own manager No decline against the pre-training sample Speed went up and checking stopped
Output measure on the deliverable The figure recorded before training Still moved, and the gap has not closed The gain was novelty, now spent

The 40% threshold sits near the transfer literature’s six-month figure rather than above it. A program holding two in five people on a named behavior half a year later is performing as well as the general evidence predicts, and anyone promising 80% is promising something nobody has measured.

Two of those four numbers depend on a figure recorded before the first session, the part teams skip and cannot reconstruct, since recalled timings drift toward whatever answer is wanted later. What to capture is set out in our look at what to measure the week before training starts.

Month six is also the fourth checkpoint rather than the first, and it reads differently if you have the earlier three, which is the sequence in our breakdown of the 30, 60 and 90-day gates.

What has to be specified at purchase

The features that make something findable at month six are bought in advance, not retrofitted at month five. Four are worth asking about before signing, and with market rates for team workshops running roughly $2,000 to $15,000, none is an expensive add-on.

One named behavior, in writing. Which deliverable, produced how often, by whom. A proposal listing capabilities instead is selling breadth, and breadth leaves nothing to find.

Practice on your own documents. A session run on a vendor’s sample data teaches a vendor’s workflow. People retain what they produced, so they should produce their own work in the room.

Artifacts that leave with the team. Prompts, templates and checklists in a place the team already opens, not a deck and not a portal with its own login.

Spacing, not a single block. Two or three shorter sessions a week or more apart beat one long day for a six-month horizon. The single day is priced more attractively because it costs less to deliver.

One thing not to buy on a vendor’s word: an unlimited reinforcement retainer sold against a retention number nobody can source. Run the withdrawal day first, find which behavior is load-bearing, then buy reinforcement for that one. The wider case sits in our argument that buying licenses is not the same as building capability.

Frequently asked questions

How much of an AI workshop is still in use after six months?

Plan for roughly two in five people still applying the material, concentrated on whichever behavior was attached to a recurring deliverable. The general transfer literature puts six-month application around 44% and twelve-month application near a third. No AI-specific study of comparable quality exists, so a precise AI number is a vendor model.

Why does month six matter more than day 30 or day 90?

Novelty stops paying in the first half-year. BCG’s 2026 survey found cognitive stretch was the top driver for people regular with AI under six months, while strategic clarity took over past a year. Earlier checkpoints measure enthusiasm plus logistics; month six measures structure.

What is the cheapest way to find out what survived?

Take the tools away for one working day, record what people could not finish and what they did instead, then hold the comparison conversation that week. Atlassian’s published version cost two days and produced findings no usage dashboard can give, including which use is habit rather than judgment.

Which parts of AI training reliably survive?

A single substitution inside a deliverable that recurs on its own, a saved artifact reachable without recall, a changed default in a tool already open, and the habit of checking output before sending it. Anything needing someone to remember a technique and choose to apply it is gone.

Is the fade a training problem or a tool problem?

Usually a trigger problem. The most common cause is that the taught behavior was never attached to a recurring event, so nothing in a normal week prompts it. Check that before concluding the session was weak or the tool was wrong.

Do refresher sessions work, and how often?

Spaced practice has strong evidence behind it, and the longer the horizon you care about, the wider the gaps should be. For six months, two or three short sessions across the first two months beat a single refresher at month five, which arrives after the behavior has lapsed.

We recorded no baseline. Can we still measure retention?

Partly. Repeat use, spread across teams and a quality look at real work samples can all start late. The output measure cannot be recovered, because remembered timings bend toward the answer people now want. Pick a team the training has not reached, record its state this week, and compare there.

Should we re-train the same people or train a new group?

Re-train the same group only if the withdrawal day shows the behavior was understood and lapsed for lack of a trigger, in which case the fix is a trigger rather than a session. If the behavior never started, repeating the session repeats the result.

Does a prompt library help retention?

Only when it lives where people already work and someone owns it. A shared document a team edits survives. A separate portal with its own login does not, and a library nobody has changed in six months means the behavior went with it.

Who should own the six-month review if there is no AI lead?

Split it by who already holds the information. Usage belongs to whoever administers the tools. Repeat use and the quality look belong to line managers, since judging a work sample needs knowledge of the job. The output measure belongs to the function owning the deliverable.

Key takeaways

  • Expect three or four narrow behaviors to survive a one-day session and most of the rest to be gone by month six. That is the normal result.
  • What persists is attached to something that happens anyway: a recurring deliverable, a saved artifact, a changed default. What fades needs someone to remember and choose.
  • The transfer literature puts six-month application near 44%. The AI-specific numbers in circulation are vendor models.
  • Month six is where novelty stops paying, and BCG’s 2026 data shows the drivers of sustained use inverting across that line.
  • A one-day withdrawal test is the cheapest honest instrument. What people cannot finish without the tools is what they absorbed.
  • Four numbers are enough: repeat use on a named behavior, spread by team, one quality sample, and the output measure recorded before training.
  • What leaves something findable at month six is specified at purchase: one named behavior, practice on your own documents, artifacts the team keeps, spacing.

To find out what your own team kept, book a free AI readiness call. We will walk through the behavior worth naming, how to run the withdrawal day without it becoming a test, and what a six-month result should be allowed to decide.

Last Updated: Nov 13, 2026

AW

Arthur Wandzel

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