One Team, One Quarter: The Cycle in Practice

The essays in this series describe one learning cycle. Find the skill gap in the work itself, practise against it little and often, measure the work again the same way, and read the results by team and skill, never by name. This essay follows one team through that cycle, sets out what it costs, and says what is still untested.

Five claims sit behind the cycle, one from each essay. Training can only show that it worked by reading the work itself, before and after. Adults learn the way everyone does, so training should start from what each person already knows. Skills that are not practised fade within months. Learning data helps when it describes the task and the team, and harms when it ranks or watches people. And workplace English stalls for the same reasons, and moves with the same cycle.

One Team

The team below is invented, to show the cycle from start to finish. It is a customer service team of 24 people who answer customers by email.

The cycle starts with a diagnostic. An AI model reads a month of the team’s written replies against an error map: the mistakes people in this job usually make, such as a reply with no clear next step, a refund rule stated wrongly or a question left unanswered. A trainer checks a sample of what it flags before anything is reported. The trainer is not the team’s manager and has no say over anyone’s pay or position.

The diagnostic finds that four replies in ten give the customer no clear next step, and most of them are replies to refund requests. That is the skill gap, and it came from the work, not from a course catalogue or an annual appraisal.

Practice replaces the half-day course. Every week each person gets a ten-minute task built from the team’s own anonymised emails: rewrite a reply so the next step is clear, or spot what is missing from one. New starters get worked examples first, and the scaffolding fades as they improve. Experienced staff go straight to the hard refund cases.

The feedback is about the task. The AI drafts it, and the trainer approves it before anyone sees it. The human stays in the loop, and the feedback points at the reply, not at the person who wrote it.

After a quarter, the same reading runs again. Replies with no clear next step have fallen from four in ten to one in ten, and the ones left are mostly about one refund rule, which tells the company where the next round of practice should go.

The team’s manager sees the report, and the report shows the team and the skill. No group in it is smaller than five people, each learner sees their own progress against their own past, and nothing goes into a performance review.

What It Costs

A training manager’s first question is who does the reading, and what it costs.

Most of the cost of training is the learners’ time. In Spain, an hour of company training cost about €75 in 2020, and €60 of that was the paid time of the people being trained (Eurostat, 2020, a year affected by COVID). The trainer, the materials and the fees were the smaller part.

That changes the sums. A half-day course for 24 people takes about 96 hours of paid time in one go. Ten minutes a week for twelve weeks takes about 48, and the trainer’s review, at 30 minutes a week, adds 6 more. These figures are illustrative.

Kept up all year, the practice takes more hours than one course, not fewer. What changes is what the hours buy. A course is paid for once and has faded within months. Practice spread across the year keeps the skill in use, and each quarter’s reading shows whether the hours are working.

Two costs come before any of that. The error map has to be built and checked once for each job, and the software has to be paid for.

The AI does the first reading, and a person checks it. A trained educator still gives better feedback than AI. In a US study of 200 essays, trained humans gave better feedback than ChatGPT on almost every measure (Steiss and colleagues, 2024). Automated feedback still improves writing, with a medium average effect across studies (Fleckenstein and colleagues, 2023). AI drafts, the trainer checks.

Honest Limits

Three parts of this cycle have not been tested at work.

The error maps exist for English writing, where FeedForward Learning was built. A map of how people go wrong in sales calls or safety procedures would have to be built and checked for each job before the diagnostic could be trusted.

Comparing a learner with their own past rather than with colleagues comes from education research, not workplace trials. It is a design principle for work, still to be tested.

Spaced practice keeps skills alive in healthcare trials, but nobody has shown that it reduces injuries at work.

The way to find out is to run the cycle with one team for one quarter, with the measure agreed before it starts.