A dashboard nobody trusted, so everyone went back to spreadsheets
A non-profit's staff could operate the new dashboards perfectly well and still went back to the spreadsheet, because nobody had ever taught them to believe what a chart was telling them, so I rebuilt the training around that.
The problem
A UK non-profit rolled out Power BI to its 150 staff eighteen months ago. Weekly active use of the dashboards sat under 15%, and it stayed there. Most staff had quietly gone back to the spreadsheets they were meant to have moved on from, and the dashboards built to replace those spreadsheets sat unopened.
The obvious read was that people hadn't been shown properly. It wasn't true. Staff who'd been through the original training could apply a filter, switch a view, do everything the course had asked of them. What they couldn't do, and what nobody had ever taught them, was decide whether to believe the number in front of them once they'd built it. Someone looking at an unfamiliar chart still reached for the spreadsheet she already trusted, not because she couldn't read Power BI, but because she didn't trust what it was telling her. This wasn't a skills gap. It was a trust gap, and the first rollout had only ever tested for the first one.
The learner and the constraints
Two very different people sit at the end of this. Operational and programme staff, some with real anxiety about anything that looks like maths, who need to read a chart calmly rather than freeze at a number. And decision-makers, fluent enough to have a dashboard opened for them in a meeting, who defaulted to gut feel the moment a chart said something they didn't expect. Neither group needed to be shown how Power BI works again. Both needed a reason to believe it.
The question I set myself: how do I teach people to trust a chart, not just operate one?
The approach, and why
I did not repeat the rollout. I built training around the moment someone doubts a number, because that is the moment the first course never reached.
- Find the real gap before designing anything. I mapped the original training against Gilbert's Behaviour Engineering Model, environment against person, and the pattern was stark: strong coverage of what staff knew and could do, almost none of what the organisation had ever told them about judging whether a number was reliable. That gap, not a generic skills gap, was the finding.
- Split the two audiences instead of building one course. A tiered pathway, working at different Bloom's levels, routes anxious operational staff and confident decision-makers separately rather than through a shared curriculum. Recognising a dashboard element is not the same skill as judging when a metric needs challenging, and treating them as one course was the first rollout's real mistake.
- Teach the reversal, not the walkthrough. The signature piece is a chart that looks like it says one thing and, read properly, says the opposite. A bar chart of monthly workshop attendance shows one month as the shortest bar, and the fast read is "cut the programme back." Look again and only three sessions ran that month, so each one was actually the best attended all year. The honest read is "run more sessions, not fewer." That reversal is the whole point: a fast, confident glance gets it wrong, and the training rehearses the slower, correct read until it becomes the habit.
- Build every number twice, for everyone. Low-stakes, self-paced practice, and every statistic shown as plain language and as a chart, so a learner who freezes at a number in isolation has a second way in. The chart also carries an accessible data table alongside it, holding the same figures, so the reversal works exactly as well for a screen-reader user as for anyone looking at the bars.
- Measure trust, not attendance at the training. The evaluation tracks weekly-active dashboard use and self-serve report generation at Kirkpatrick Levels 3 and 4, never completion, because completion was the first rollout's success measure and it produced sub-15% adoption regardless. Built to WCAG 2.2 AA throughout.
Artefacts
A lightweight diagnostic mapped to Gilbert's BEM, a tiered pathway map for the two populations, the sample microlearning module with its rubric, and an adoption dashboard mock-up. All available to walk through on request.
The outcome
The honest number is the one the organisation started with. Under 15% weekly-active use, eighteen months in, is the only real figure here. Everything the redesign is built to move is a design target until a real cohort has run through it.
What I took from it
The first rollout did not fail. It succeeded completely at teaching people to use the tool, and that success is exactly why nobody noticed the real gap for eighteen months. You can teach someone to operate a dashboard in an afternoon. Teaching them to believe what it is telling them, especially when the chart looks wrong and isn't, takes something closer to practice than instruction. That is the difference between a course people pass and training that gets a dashboard opened instead of a spreadsheet.
Further reading: why AI should support, not replace, the designer, measuring behaviour, not completion.