The buyer rated it 4.1. The fourteen thousand people who had to use it were never asked.

Two populations, measured separately

The first decision of the engagement was to refuse the average. The platform has two user populations with almost nothing in common: about 600 HR operators who open it several hundred times a year, and about 14,000 employees who open it roughly five times a year and forget it in between. Scoring them together would have produced one number describing nobody, so the Task surface was scored twice and reported twice.

The two readings landed in opposite quadrants of the say-do matrix, on the same screens, in the same week.

What the measurement said

For employees, the Task surface scored 2.2 behaviour against 4.1 perception: silent friction, concentrated in time-off requests and first-week onboarding. Median time to file a single day of leave was 4 minutes 10 seconds across a window that opened when the employee decided to request it and closed when the record was filed. Thirty-one per cent of all requests were ultimately entered by an HR operator on the employee's behalf, a number that existed in the database from the beginning and had never been reported, because no dashboard had a row for it.

For operators, behaviour scored 2.9 against perception 2.1: loud friction, and every item on it was already written down. Bulk approvals, filtering, the export. Real problems, correctly identified, and a receipt rather than a discovery.

Baseline confidence was 0.72. The behavioural side was close to fully instrumented, which is normal for a system running on infrastructure the client owns. The perception side was weak and declared weak: operator perception came from vendor-run interviews, and employee perception from a client-administered engagement survey, which the collection-conditions discount takes to 0.8 of its already reported weight because the employer is the one asking.

Ninety-six per cent of the users had never been asked anything. The four per cent who had been asked were describing a different product, accurately.

The four surfaces at baseline

Composite computed on the employee population, who are 96% of users, under internal-tool weighting: Task 0.40, Access 0.30, Commitment 0.15, Belief 0.15. The operator reading is reported beside it and never folded into it.

The treatment

Per the silent-friction playbook: subtract, do not explain. The time-off form asked for a cost centre code. Employees did not know their cost centre code, so they asked someone, or guessed, or gave up and mailed HR, which is what the thirty-one per cent was made of. The code was already inferable from the employment record. The field was removed, populated automatically, and left editable for the three per cent of requests where it legitimately differs.

The second change was the policy question. Deciding whether a day counts as sick leave, carers' leave or unpaid was the largest single component of the four-minute window, and the answer lived in a linked forty-page PDF. The three qualifying rules were rewritten into one sentence each and placed at the point of the choice. Nothing was redesigned; text was moved from a document into a form.

What did not work

4m10s → 50s

Median time to file a time-off request

31% → 9%

Requests entered by an operator on the employee's behalf

2.2 → 3.6

Task surface, behaviour, employee population

What it cost, in the only currency available

There is no conversion event in an HR platform, so the finding was priced in hours. Roughly 61,000 leave requests a year across the client organisations, at 3 minutes 20 seconds saved each, is about 3,400 employee hours returned. The assisted-entry drop from 31% to 9% removed roughly 13,400 operator interventions a year at a little under four minutes each, which is another 880 hours and is the number that carried the renewal conversation, because it lands in a team whose headcount is on a budget line.