A company runs a two-day “AI for managers” programme for sixty people. Feedback is excellent. Six months later, the honest audit shows those sixty managers using a chatbot to tidy emails and nothing else. The tool was learnt. The queue was untouched.
That is the usual outcome, and it is not a motivation problem. It is a unit-of-training problem. A programme that teaches capabilities cannot check whether anything changed, so it measures confidence on a feedback form. A programme that teaches three specific tasks can check whether an artefact exists on a Tuesday that did not exist before.
Why generic AI literacy produces nothing
Two findings explain the gap.
Brynjolfsson, Li and Raymond studied 5,179 customer support agents and found that access to a generative assistant raised issues resolved per hour by 14 per cent on average. The average hides the useful part: the gain was 34 per cent for novices and minimal for the experienced. The tool compressed a learning curve. It did not make experts faster.
Noy and Zhang ran a pre-registered experiment with 453 college-educated professionals on mid-level writing tasks. Average time fell by about 40 per cent and output quality rose by 18 per cent, graded by independent evaluators. Again, the weaker writers gained most.
Read together, the pattern is not “AI raises productivity”. It is “AI raises productivity where the work is a repeatable artefact and the person is not already good at it”. A manager who writes one board note a quarter stays a novice at that task forever. So the selection question is not “which tool”. It is “which recurring manager output has a reader and a checkable form”. Get that wrong and the licence spend produces a training completion rate.
The artefact test for choosing a task
Put every candidate task through three filters. It must pass all three.
- A named consumer. Someone downstream opens the output and acts on it. A meeting note nobody reads is a diary, not work.
- Digital input. The material the manager feeds in already exists as text: notes, an email thread, a spreadsheet, last month’s report. Tasks that begin with someone else’s phone call need a different design.
- A one-week failure signal. If the artefact is missing or wrong, someone notices within a week. Without that, you cannot tell automation from decoration.
Three manager tasks pass in most Indian organisations, in the order the work actually happens across a week.
Task one: the meeting note into an action log
The input is the manager’s own rough notes from a review, a standup or a project call. The output is four columns: action, owner, date, and the decision it came from. The consumer is the person who owns the action and the manager at the next review.
This is the highest-yield start because most managers already take notes and almost none convert them. The failure mode is not a bad summary. It is an action with no owner or a date never agreed. So the prompt must forbid invention. The instruction is not “summarise this”. It is: list only actions where a person and a date were actually said; where either is missing, write UNASSIGNED and leave the date blank.
That rule changes the review. A log with three UNASSIGNED lines is better than a fluent paragraph, because it names what the meeting failed to decide.
Task two: last week’s actions into this week’s one-to-one
The input is the action log from task one. The output is a four-line agenda: what was promised, what is late, one question to ask, one thing to decide. The consumer is the direct report, who receives it before the meeting rather than during it.
Managers skip this because the agenda takes fifteen minutes they do not have at 9pm. The model needs no judgement here, only a diff of two lists and the overdue items. Set a hard rule on length: four lines and no more. An agenda that runs to a page becomes a status report, and the report replaces the conversation.
Task three: the month into an exception list
The input is the manager’s monthly numbers or last month’s update. The output is two lists: what moved outside tolerance, and what needs a senior decision. The consumer is the skip-level or the business review.
This is the task most programmes get wrong. They teach managers to “use AI to write the report faster”, and the report gets longer, not clearer. Ask for exceptions only, against a threshold the manager sets before the prompt runs: any variance beyond five per cent against plan, any commitment that slipped by more than a week. The model sorts. The manager decides.
| Task | Input | Artefact | Reader | What failure looks like |
|---|---|---|---|---|
| Action log | Rough meeting notes | Four columns, one row per action | Action owner, next review | Action with no owner or invented date |
| One-to-one agenda | Last week’s action log | Four lines | Direct report, before the meeting | Agenda becomes a status report |
| Exception list | Monthly numbers or update | Two short lists | Skip-level, business review | Report grows longer, decisions do not appear |
Build it in ninety minutes, on their own documents
Do not run a session on a sample dataset. Use the manager’s real material, because the value is in the prompt matching their vocabulary. Split the ninety minutes into three parts.
Twenty minutes: one golden example. The manager takes a real meeting from last month, writes the action log by hand, and that hand-written version becomes the target. Every later output is compared against it. This is the part that most programmes skip, and it is the only part that makes the check objective.
Forty minutes: the prompt card. One card, five fields: the input, the artefact shape, the fields that must never be invented, the length limit and one rejection rule. The rejection rule is the manager’s own: “if the log has no UNASSIGNED lines and I remember leaving the room unsure who owned something, the output is wrong.” That sentence is more useful than any prompt-writing tip, because it tells the manager what to distrust.
Thirty minutes: the second run. The manager runs the same prompt on this week’s real meeting and edits the output. The gap between the first and second run is the training. If the second run needs no edits at all, the task was too easy to be worth the licence.
This is where an AI training programme earns its fee or does not. Not in the number of tools covered, but in whether each manager leaves with a card that produces an artefact their colleague actually reads. The same discipline applies to the wider training portfolio: specify the output before you specify the content.
Worked example: fourteen managers, four weeks
Illustrative example. A mid-sized manufacturer puts fourteen plant and functional managers through the three tasks over four weeks, one task per week with a spare week for catch-up. The numbers are invented to show how the measurement works. They are not a client result.
Baseline week: the administrator collects the last four weeks of meeting notes and counts how many contain at least one action with an owner and a date. Suppose 9 of 46 do, about one in five. That baseline matters, because without it any later improvement is a feeling.
Week one: all fourteen produce an action log for their next real meeting. The administrator checks two things only. Does the log contain at least one UNASSIGNED line where the note was genuinely unclear, and did the manager send the log to the owners within a day? Suppose 11 of 14 send it. Three do not, and the reason is usually the same: their meetings have no owners at all, so the log is entirely UNASSIGNED and reads as an accusation. That is a finding about the meetings, not about the training, and it is worth reporting to the leadership team as such. It also belongs in the same honest frame as the smallest evaluation that survives a finance review, which keeps one leading indicator rather than a satisfaction score.
Week four: notes with a named owner and date rise from 9 of 46 to something like 21 of 46. That is a process measure, not a productivity claim. It says the artefact now exists, not that the plant runs better.
The cost side is plain. Fourteen licences at business tier for four weeks, about ten hours of an administrator’s time, and the manager hours. If nobody reads the artefact, that is the entire loss, and it is small next to a two-day offsite for the same fourteen people.
Where this method stops
Three tasks will not help a manager whose real problem is that nobody has defined what the meeting is for. If a review has no decisions, an action log will be empty every week, and the honest conclusion is that the meeting should be cancelled rather than automated.
The method also fails on judgement. Anything involving a performance rating, a grievance or a pricing concession should not be drafted by a model and forwarded. The risk is not tone. It is a fabricated detail: a date, a figure or a commitment that reads plausibly and was never made. Keep those conversations human and use the model only for the paperwork that follows a decision already taken.
A second limit bites in practice. A manager who does not read the artefact carefully will start forwarding the model’s version of events as the record, and the log becomes an authority it has not earned. The defence is procedural: the manager signs it, and the log says who wrote the first draft.
Finally, do not treat these three as a ladder. Some managers will only ever need the action log, and that is a fine outcome. Three tasks adopted by half a cohort beats six demonstrated to all of them. A wider set belongs in the e-learning build once the tasks are fixed, not before.
What to do on Monday
Take the next management cohort and cancel the tool tour. Ask each manager for last month’s meeting notes and have them write the action log by hand in the room. That hand-written log is the golden example. Then run the same meeting through the model and count the edits. Book the second session in the same week, because a prompt card not run twice in seven days is not learnt. Record the baseline count of notes with a named owner and a date before you start.
Sources
- https://www.nber.org/system/files/working_papers/w31161/w31161.pdf
- https://www.science.org/doi/10.1126/science.adh2586
- https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain
- https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born