Training builds the skill. Whether it turns into better work depends on what happens in the weeks after the workshop, and that part usually has no owner.
Training builds capability, but capability only becomes results when the skill has a recurring task to live in and a person reinforcing it. Run the five-question test on one skill and one role, fix the first link that fails, and judge success by use, persistence and impact rather than attendance.
Say you lead learning and development at an insurer with 1,200 employees. In March, 400 people from claims, underwriting and HR sat through a two-hour workshop on using AI at work. The feedback scores came back at 4.6 out of 5. In June your CHRO asks a plain question: where is this showing up in the work? You open the dashboard and find a completion count, a licence count and nothing else. Nobody can tell you what changed on a Tuesday afternoon in the claims team.
That gap usually has an explanation, and it isn’t always a training problem. Training is the first of four links, and a skill has to pass through all four before it shows up as a result. We’ve already made the case for why organisations need to upskill people and for which AI skills are worth teaching. This piece picks up the day after the workshop.
| Link | What it does | What it looks like when it’s missing | Who usually owns it |
|---|---|---|---|
| 1. Training | Builds capability. Someone can now do the thing, in principle. | People can’t do it on their own file without the facilitator in the room. | L&D, often with an expert facilitator |
| 2. Workflow design | Gives the capability somewhere to be used: a real, recurring task it slots into. | Everyone says the session was useful, and nobody can name when they’d use it. | The process owner or line manager, with an HR business partner |
| 3. Reinforcement | Turns it into behaviour through repetition, review and a manager who checks. | A good first fortnight, then the old way of working comes back. | The line manager, supported by L&D |
| 4. Adoption | What you see when the new way of working sticks. It is the result of links 1 to 3. | Usage numbers move while the work itself looks the same. | Nobody directly, because it isn’t a lever |
Keep these four words and meanings consistent across your programme. Adoption is an outcome of the first three links, so there’s no separate way to pull on it.
If you work in L&D, you’ll recognise the pattern from Kirkpatrick’s model. Level 2 measures whether people acquired the skills. Level 3 asks whether they perform the critical behaviours on the job and are supported in, and accountable for, doing so.[1] Many AI programmes end up measured at levels 1 and 2, because those are the numbers a workshop produces by itself. The Tool x Workflows x Behavior idea in our piece on what makes someone AI-powered is the individual version of the same point. The rest of this page is the organisational version, one skill at a time.
The mistake in our insurer scenario is treating “nothing to show” as one problem. It’s at least three, and each has a different fix. So work at the smallest useful size: one skill, in one role. “Checking AI output against the source” for claims handlers is a unit you can test. “AI skills” for “employees” isn’t.
Here are five yes-or-no questions. Answer them for a skill and a role, in order. The first “no” is the link that’s broken.
| # | Question | Link it tests |
|---|---|---|
| 1 | Can someone in this role do the skill on a real file from last week, with nobody coaching them? | Training |
| 2 | Can they name the recurring task, and the trigger or day, where the skill gets used? | Workflow design |
| 3 | Is the AI step written into the checklist or template the team already works from? | Workflow design |
| 4 | Has someone looked at actual output from this skill in the last 30 days? | Reinforcement |
| 5 | Does the person’s manager ask about it in a 1:1 or team meeting without HR prompting? | Reinforcement |
Five yeses means you’re ready to measure (section six). Any “no” means fix that link first, and stop at the first one.
Stop at the first no, because a later fix can’t rescue an earlier break.
Here’s the test run on three skills in the same company. They all came out of the same workshop. They fail in three different places.
| Skill and role | Q1 | Q2 | Q3 | Q4 | Q5 | First no | The fix |
|---|---|---|---|---|---|---|---|
| Check an AI summary against the source, claims handler | Yes | Yes | No | No | No | Q3, workflow design | Add one line to the claim-file checklist: “AI summary checked against the original note, initials.” |
| Summarise interview notes, recruiter | Yes | No | No | No | No | Q2, workflow design | Pick one trigger, “within an hour of each first-round interview,” and write it into the scorecard template. |
| Draft a first-pass policy answer, HR advisor | Yes | Yes | Yes | No | No | Q4, reinforcement | Book a monthly 20-minute review of five real outputs with the HR lead. |
Illustrative answers for a fictional insurer. None of these three fixes is a second workshop, because link 1 was already working.
Where Q1 is a “no,” more training is the right answer. A facilitator who works on your own files shortens that curve, and a team learning together builds shared capability, which one enthusiast on their own can’t. The test just stops you from reaching for it when the broken link is further down the chain.
Don’t train a skill until you can name the recurring task it will live in.
There’s some support for this in the academic literature. A meta-analysis of 89 studies of training transfer reported that factors such as motivation and work environment predicted transfer more strongly for open skills, like leadership development, than for closed ones, like computer software.[2] Checking whether an AI summary is right is a judgment skill, so in our reading it sits nearer the open end. That’s our interpretation, not the authors’ claim, and it’s the reason the work environment gets two of the five questions.
A separate point: if you’re still deciding whether the organisation is ready for any of this, that’s a different job, and our AI readiness check for teams and organisations covers it. The test above assumes people have already been trained.
A recruiter at a 300-person logistics firm learns to turn her interview notes into a structured summary with AI. She uses it twice, likes it, and then stops. Her manager assumes the skill didn’t stick. When asked, she says it’s faster to just type the scorecard herself, because the applicant tracking system wants four specific headings and the summary didn’t match them. The extra step cost more than it saved.
Her capability was fine. The workflow had no slot for it. That’s the pattern at link 2: a skill with no named task, no trigger and no place in the template becomes an extra thing to remember, and extra things tend to lose to the way people already work.
A workable slot has five parts, and you can fill them in during a 30-minute conversation with the team:
Once the slot exists, it’s worth being just as explicit about who does what inside it, because this is a task where AI does real work.
| What AI does | What the human still owns | How it gets checked |
|---|---|---|
| Reorganises the recruiter’s notes under the scorecard’s four headings and flags anything the notes don’t cover | Every rating, every judgment about the candidate, and the final wording | The hiring manager reads the scorecard against the recruiter’s raw notes for one candidate in five |
| Drafts a follow-up question list from the gaps | Deciding which gaps matter and which questions are fair to ask | The recruiting lead reviews five scorecards a month against the notes |
The AI step is preparation. Ratings about people stay with a person, and the check is scheduled so it doesn’t depend on anyone feeling suspicious.
Every role fills the same five parts with different content. If you can’t fill all five for a role, that’s useful information. It usually means that role isn’t ready for that skill yet, and a different skill might be the better one to start with.
Training has a line in the budget and an invitation in the calendar. Reinforcement is mostly a manager asking one question in a 1:1, and that has no line anywhere. So it tends not to happen, and the first fortnight’s enthusiasm fades back into the old routine.
Nobody gets a calendar invite for it.
Gallup’s 2026 survey of U.S. workers points the same direction. Among people who use AI weekly or more, those whose managers actively support AI use were much more likely to say AI improves how work gets done: 33% versus 9%.[3] That’s self-reported and correlational, so it doesn’t prove a manager’s support causes the difference. It does fit what the Kirkpatrick material says about behaviour needing support and accountability, and it tells L&D where the lever is: with the person who sees the work.
Reinforcement needs a named person whose job includes asking, and a date on which they ask.
Here’s a schedule you can hand to a line manager. It takes about an hour across the first quarter.
The manager picks one real task and the person uses the skill on it. Not a practice exercise.
The person and a peer review five real outputs together for 15 minutes. They note what each one needed fixing.
The manager asks in a 1:1 for the three questions below and writes down the answers.
L&D checks persistence: is the skill still being used without a reminder?
The manager and L&D review impact against the baseline taken before training.
A starting schedule. Shorten it for fast-cycle work and stretch it for tasks that only come around quarterly.
The manager’s three questions at week four:
Question two is the useful one. If the answer is “nothing,” the person may not be checking the output at all, which is a link 1 problem hiding as a link 3 one. L&D can then use AI on the managers’ notes, as long as you strip names and check your data policy first.
| I’m in learning and development. Below are 12 short check-in notes from line managers about one AI skill, “checking an AI summary against the source,” used by claims handlers. Names have been removed. For each note, do three things: 1. Say which link looks broken: Training (can’t do it alone), Workflow design (no clear task, trigger or template step), Reinforcement (nobody checks or asks), or None. 2. Quote the phrase in the note that made you decide. 3. If the note doesn’t give enough to judge, say “Not enough information” and don’t guess. Then give me a count per link and the three notes that are hardest to classify, so I can read them myself. Return a table with columns: note number, link, quoted phrase. |
The AI sorts and quotes. You read the hard cases yourself, and the counts are a prompt for a conversation with managers, not a score for any one of them.
“Employees” is too broad a group to plan for. A recruiter, an HR business partner and a people manager meet the AI step in very different places, and they stall at different links. The table below is a set of working hypotheses from workshop conversations, not findings. Use it to guess where to look first, then confirm with the five questions.
| Role | A skill worth starting with | Where it tends to stall | First fix to try |
|---|---|---|---|
| Recruiter | Structuring interview notes | Link 2: the ATS or scorecard has no slot for it | Write the AI step into the scorecard template |
| HR business partner | Preparing for a hard conversation | Link 3: used once, then forgotten when the calendar fills | A fixed 15-minute prep slot before each case meeting |
| Learning designer | Drafting scenarios and quiz items | Link 1: output looks fluent and gets accepted without a check | Practice on a real course with a peer reviewing the draft |
| People manager | Drafting 1:1 agendas and feedback | Link 3: nobody asks them, so nobody notices if they stop | Their own manager asks the week-four questions |
| Finance analyst | Variance commentary first drafts | Link 2: the close process is fixed and the step isn’t in it | Add the step to the close checklist, with an owner |
| Customer support agent | Reviewing AI-suggested replies | Link 3: speed targets reward sending, not checking | Make the check a counted part of the quality review |
Working hypotheses, not findings. Your own answers to the five questions beat this table every time you have them.
Role matters for impact as well as for where things stall. A study published in the Quarterly Journal of Economics looked at a generative AI assistant deployed to customer support agents reported a 14% average gain in issues resolved per hour, a 34% gain for novice and low-skilled agents, and minimal gain for experienced and highly skilled ones.[4] It measured an assistant rolled out inside a company, not a training programme, so don’t read it as a training result. What it does show is that one tool can produce very different gains across experience levels in a single team.
For L&D that has a practical consequence. If you measure impact as one number across a whole department, a big gain for newer people and a flat line for senior ones can average out into something that looks unremarkable. Cut the impact numbers by role, and by tenure where you can, before you decide a skill isn’t working.
Training completion and licence counts are the numbers everyone can pull, which is why they end up on the dashboard. They tell you the programme happened. They don’t tell you whether the claims team checks summaries differently now. The lens we use at Future Factors has three steps: use, persistence and impact. Use is whether people are doing it. Persistence is whether they’re still doing it a month on, without being reminded. Impact is whether the work itself is better, faster or different.
Attendance tells you the room was full.
| Step | The number or signal | Who pulls it, and when | What it tells you |
|---|---|---|---|
| Use | Share of claim files where the “AI summary checked, initials” line is filled in | Claims team lead, weekly for the first month | Whether the skill has a place to live. A low number points to link 2. |
| Persistence | The same share in month three, plus how often managers’ 1:1 notes mention it | L&D at week 8 and week 12 | Whether it became a habit or a burst. A drop after week four points to link 3. |
| Impact | Rework on summaries sent back for correction, minutes to prepare a medical summary against the pre-training baseline, errors caught at review | Claims manager with L&D, quarterly | Whether the work changed. Read it as direction, because other things move these numbers too. |
An illustrative sheet. The impact column only works if someone captured the baseline before the training started.
The baseline is the part that’s easy to miss, and it can’t be rebuilt afterwards. If you want to write up the result for a finance lead once you have it, our piece on what counts as proof for an AI initiative covers baselines and how strong different kinds of evidence are. And if the programme itself needs maintaining as the tools change, keeping an AI upskilling programme from going stale is the companion read.
Count a skill as adopted only if it’s still being used a month after the last session, without a reminder.
That’s a deliberately strict rule, and it’s a judgment call on our part rather than a published standard. Some tasks only come around quarterly, so stretch the window to fit one full cycle of the work. The principle holds even then: one cycle without a prompt, not one week after a workshop.
You can start small: one skill, one role and a month.
One pair first. Widen to a second role only after the first one passes the persistence check.
A team can run all of this internally with the tables on this page, and many will. Expert-led training adds something at the links where people most often need it: it shortens the learning curve at link 1, it builds shared capability across a whole team rather than in one enthusiastic person, and it helps people practise on their own real files instead of generic examples. A facilitator who has watched skills stall at links 2 and 3 can also help your managers write the slot and ask the week-four questions. If you’d like that support, our corporate workshops are built around this chain.
Which skill and role will you take through the five questions first?
Check use within the first month and persistence at about two months. Impact needs at least one full cycle of the work itself, so for a weekly task that might be 8 to 12 weeks, and for a quarterly one it could be longer. Take the baseline before the training starts, because it can’t be reconstructed later.
The line manager does the reinforcing, because they see the work. L&D supports by supplying the schedule, the three check-in questions and the persistence check at week eight. If reinforcement is left to HR alone, it tends to turn into reminder emails that nobody connects to a real task.
Then that role probably isn’t ready for that skill yet, and a different skill may fit better. If you can’t name the task, the trigger and the place it’s written down, the training has nowhere to land. Choose a skill that matches work the person already does on a regular rhythm.
A readiness assessment looks at whether people and conditions are ready before a rollout, and scores the team and the organisation. This test starts after training and follows one skill in one role to find the first link that’s broken. They’re companions: readiness first, then this once people have been trained.
Yes. A manager can run the five questions, write the slot and ask the week-four questions with no one else involved. Expert-led training still helps at the first link, because it shortens the learning curve and builds shared capability across the team instead of leaving the skill with one person.
The distinction between the learning and behaviour levels comes from Kirkpatrick Partners’ description of the Kirkpatrick Model. The meta-analysis of 89 training transfer studies is Blume, Ford, Baldwin and Huang (Journal of Management, 2010), and I read its published abstract; the link between judgment skills and open skills is our interpretation. The manager-support figures are from Gallup’s 2026 U.S. workplace survey and are self-reported. The customer support figures are from Brynjolfsson, Li and Raymond (NBER 2023, QJE 2025). The 30-day, week-four and week-eight timings are Future Factors working heuristics. Widely repeated figures about how much training people forget were left out because I couldn’t trace them to a primary source. The insurer, the recruiter and the claims examples are illustrative, not real cases.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.