With the same study time, recall was higher when learning was split across sessions than when it came in one sitting, in all but 12 of 271 comparisons the research has pooled.
A One-Day AI Workshop or a Programme: Which One Lasts?
By James Cotton · Last updated · Sources checked · 11 min read
Part of our topic guide on AI Skills for Business.
By James Cotton, Founder, iO-Sphere
What happens when the same hours are split up?
Hold the hours fixed and they can be spent in one sitting, which researchers call massed practice, or split across sessions with gaps between them, which they call spaced. The two have been compared on learning and training of many kinds, from word lists to clinical decisions.
The largest synthesis, Cepeda and colleagues' 2006 review, gathered 839 assessments of spaced practice from 317 experiments in 184 articles. In that review, with study time held equal, spacing came out ahead in all but 12 of 271 comparisons, and average final recall was 47.3% for spaced study against 36.7% for massed.
The tasks were verbal recall, such as word lists and passages of text, mostly in laboratories, and none was a workplace skill. Within those limits, the authors concluded that "separating learning episodes by a period of at least 1 day, rather than concentrating all learning into one session, is extremely useful for maximizing long-term retention".
In Cepeda and colleagues' 2008 online study, 1,354 adults with a mean age of 34 learned trivia facts, reviewed them once after a gap and took a final test up to a year later; for the same study time, the best gap gave 64% more recall than no gap at all.
The evidence is thinnest where workplace training lives. In the 2006 synthesis, 80% of comparisons tested recall less than a day later, only 4% after more than a month, and 85% of the data came from young adults; the authors asked for "much less emphasis on convenient single-session studies".
Kim and colleagues' 2019 study of 10,514 employees at five companies, answering short question sets on one vendor's training platform, found the laboratory pattern held: the best gap between practice questions grew as the time to the test grew. It measured answers to knowledge questions and did not look at behaviour on the job.
The counter-evidence comes from Donovan and Radosevich's 1999 meta-analysis, a study that pools the results of earlier studies, here 63 of them. The spacing effect shrank as tasks grew more complex: measured as an effect size, a standard measure of how far apart two groups' results are, it was 0.97 for the simplest cluster of tasks and 0.07 for the most complex.
Less rigorous studies reported bigger effects, and the authors found the strong effect "limited to relatively simple tasks", though for complex tasks "representative of organizational training settings" longer rest periods appeared to help. They could not measure how much of what was learned was carried back to the job, and warned that under certain conditions the effect "may even disappear or reverse itself".
Nothing above rests on Ebbinghaus's forgetting curve. Its 2015 replication by Murre and Dros used one person and 70 hours of nonsense syllables. The curve measures savings, the relearning time saved, which the replicators call "very different" from ordinary memory tests. By Ebbinghaus's figures, about a third was saved after a day, about a fifth after a month. It never measured how much of a course people forget.
Why does a training day always seem to have worked?
The 2008 online study warns that compressing learning into too short a period is likely to produce "misleadingly high levels of immediate mastery that will not survive the passage of substantial periods of time". An end-of-day quiz or feedback form catches that mastery at its peak, so it will rate a single day well whether or not anything survives the month.
The confidence people feel at the end of a session misleads in the same direction. In Roediger and Karpicke's 2006 study, students learning prose passages either re-read them or tested themselves. Re-reading did better when the final test came five minutes later; at two days and at a week, self-testing did better, even though re-reading had made the students more confident.
For Learning at Work 2023, the professional body CIPD surveyed 1,108 people responsible for learning at work, online with YouGov in January and February 2023; 91% were UK-based. In that survey, 41% said their organisations had used external conferences, workshops and events in the past two years, up from 27% in 2021.
The survey also asked about transfer, which training research still defines as Baldwin and Ford did in 1988: applying on the job what was learned, and keeping it up over time. Asked by CIPD for their top three priorities, 8% put "speeding up the transfer of learning back into the workplace" among them, and CIPD concluded that learning teams were "prioritising inputs" over "outputs".
In the same survey, 11% strongly agreed they had a process for tracking learning transfer and a further 44% agreed, yet CIPD judged that "very few set up processes to do this with intent".
What decides whether training changes the work?
The claim that only 10% of training reaches the job comes from a 1982 practitioner article's estimate of how much 1980s US training spending resulted in transfer, later repeated in Baldwin and Ford's 1988 review. Ford, one of the review's authors, co-wrote with Yelon and Billington a 2011 journal article whose title calls it "the 10% delusion"; it found the belief accepted "despite the lack of empirical, behavioral evidence".
Training does change behaviour on average: across very different programmes, Arthur and colleagues' 2003 meta-analysis found a "medium to large" effect on job behaviour. Training professionals in 150 organisations, asked for a 2006 study by Saks and Belcourt, estimated that 62% of employees apply training immediately, 44% after six months and 34% after a year.
The same study found what mattered more: what the workplace did before and after training was more strongly related to transfer than what happened during it. A 2011 narrative review by Grossman and Salas, built on Baldwin and Ford's model, likewise named opportunity to perform and follow-up among the factors with the strongest, most consistent links to transfer.
Behaviour-modelling training, in which trainees are shown a skill modelled and then practise it, shows what the workplace adds. Across 117 studies, Taylor, Russ-Eft and Chan's 2005 meta-analysis found that knowledge faded over time while skills and job behaviour held or grew. Transfer was greatest when practice used trainees' own scenarios, when trainees set goals, and when their managers were trained too.
In Lacerenza and colleagues' 2017 meta-analysis of 335 samples of employee leadership training, the analysis of design features supported spaced training sessions, practice and feedback among those linked to more effective training.
The strongest single result is a randomised trial by Kerfoot and colleagues, published in 2010. From 2007 to 2009, 95 primary care clinicians in eight US Veterans Affairs medical centres were split into two groups. One group received e-mails over 36 weeks, each presenting a case and asking whether a PSA test, a prostate cancer screening test, was appropriate, with feedback; the other received nothing.
In the trial, those sent the e-mails ordered fewer inappropriate screening tests than the control clinicians, during the 36 weeks of e-mails (10.5% against 14.2%) and over the 72 weeks after they stopped (7.8% against 13.1%). Spaced practice on a decision taken from the job changed their work, and the difference lasted after the practice ended. These were US clinicians, making one decision.
The Center for Creative Leadership calls 70-20-10 a "guideline" that "emerged from" its research, gives no method for the percentages, and says coursework has an "amplifier effect" on the other 90%. Clardy's 2018 review found "the evidential basis for the 70% rule" weak.
Johnson, Blackman and Buick's 2018 qualitative study of the Australian public sector found it was "not achieving the desired outcomes", and named as one misconception the belief that training alone would change managers.
Is AI training any different?
The studies above were of word lists, trivia facts, leadership and clinical decisions, and carrying them over to AI training is our reasoning. It starts from field studies of AI at work, where the gains come task by task and differ from one person to the next.
In Dell'Acqua and colleagues' 2026 field experiment with 758 Boston Consulting Group consultants, those given GPT-4 completed more tasks, faster and at "significantly improved quality" on 18 tasks inside the frontier, the line the researchers drew between tasks within the AI's capabilities and one chosen to fall outside them. On that task outside it, they were 19% less likely to produce a correct solution.
In Brynjolfsson, Li and Raymond's 2025 study of 5,172 customer-support agents, access to a generative-AI assistant raised issues resolved per hour by 15% on average. Less experienced agents gained most, while the most experienced saw small gains in speed and small declines in quality. Both studies measure people using a tool at work, each in one firm and one line of work.
Which of a person's tasks fall inside the line, and which outside, is found by trying the tool on their own work and checking what comes back. That takes the weeks between sessions, in a practice loop on the team's recurring tasks, and a single day can only start it.
Even the effect of a single exposure fades. In Noy and Zhang's 2023 online experiment, 453 professionals did writing tasks and half were given ChatGPT. Against the half not exposed, they were twice as likely to report using it in their real job two weeks later, and 1.6 times as likely two months later. It was one session, with no training around it, and the later use was self-reported.
In research co-developed with Skills England, published in June 2026 and applying to England, Dr Nisreen Ameen of Royal Holloway, University of London, writes that "AI training needs to last longer than a single session". Her evidence is mainly qualitative: 23 online workshops with around 150 organisations, ten case studies and a survey of 536 responses collected through Amazon Mechanical Turk.
Because the tools change quickly, she argues, training has to be easy to update, with plans to review and refresh learning after delivery; follow-up check-ins and short refresh sessions help reinforce it.
Can we trial AI training before committing?
You can, and we recommend it, as long as the trial has the programme's shape: run it with one team, not as a one-day taster, which tests a different format. Judge it weeks after the last session, on the work itself, against a team that has not yet been trained.
The Magenta Book, HM Treasury and the Evaluation Task Force's evaluation guidance for central government (May 2026 edition), says pilots let a design, its implementation and its outcomes be tested "at a smaller scale", and that early testing can explore why an intervention might not work. Using it for a company's training pilot is an analogy.
It also says such testing "does not replace" a robust evaluation of the whole intervention, and that a randomised trial or other comparison "may still be required". For a company, the comparison is a team that has not had the training yet, measured over the same weeks on the same kind of work.
People's own estimates can drift from the clock: in a 2025 randomised study by METR, a preprint, 16 experienced open-source developers forecast that AI tools would cut their task time by 24% and afterwards believed they had cut it by 20%; measured, the tools added 19%.
A 2010 meta-analysis of 89 studies by Blume and colleagues found that measuring transfer through the same source in the same setting "consistently inflated" the results, so let the work or a manager supply the measure. The checks that show a change in the work weeks later can be agreed before the pilot starts.
For a worked example, Google's AI Works report (2025, research by Public First) is vendor research: Google trained people on its own tools. It trained three groups in short formats. Small-business staff were offered five one-hour webinars and four 30-minute lunchtime sessions, spread weekly "for regular exposure and repetition", or a single full day in person. The report gives no separate result for the full day.
It measured use by self-report straight after training and three months on, with no comparison group, and calls itself "not a formal academic study". It says a few hours of training doubled or tripled AI use; among small-business staff, daily use rose from 29% to 60% and weekly use to 86%. Participants estimated saving roughly 122 hours a year, the kind of self-estimate the developers' study shows can drift.
Which should you book?
When someone asks us whether to book a workshop, a pilot or a programme, we start from what their team has to change in its work, and only then from the cost and who meets it: the training budget, an apprenticeship for one person, or both.
If everyone needs to hear the same thing once, such as a tool going live or the rules in a staff AI use policy, one session does that job. Plan for it to be remembered as a briefing.
If a team has to work differently, plan several sessions over weeks, practice on its own tasks between them and a check weeks after the last. For that team, we offer team training and custom programmes.
If you are not yet sure, run the pilot above with one team before buying for everyone.
If the day is already booked, keep it and make it the first session: Baldwin and Ford's 1988 review recorded 1960s evidence that difficult tasks went better with a longer first session followed by shorter, more frequent ones. Before the day runs, put the follow-up the Ameen research calls for in the diary: practice on real tasks, a second session and a check weeks later.
Frequently asked questions
How far apart should the sessions be?
How long the team must remember the material decides the gap, which grows with that time. In the 2008 online trivia study, the best gap was about 20 to 40% of the time to a test a week away, and about 5 to 10% for one a year away. Too long costs much less than too short, though the 2006 synthesis could not say for certain how long is best.
Does it matter whether the sessions are online or in person?
Google's AI Works found "no significant difference in AI attitudes or usage based on delivery format, whether in-person or through webinars", concluding that well-structured, relevant content mattered more than the mode. That compared in-person and online delivery inside a vendor pilot with no control group, not one day and several weeks. Lacerenza's 2017 meta-analysis of leadership training linked on-site, face-to-face delivery that was not self-administered to more effective training.
Upskilling a whole team?
Tailored data and AI training for organisations, from data literacy to technical upskilling, built around your team's real work.