How to Measure AI Workforce ROI in Hours, Not Percentages

"Productivity improved 20%." Every AI vendor says it. No grant assessor believes it, no board can audit it, and nobody can tell you where the 20% went. Percentage claims fail for three reasons, and until an organisation abandons them, its AI investment cannot be defended — or funded.

Why percentage claims are meaningless

First, no baseline. A percentage without a pre-intervention measurement is a comparison against nothing. Second, no attribution. Even with a baseline, a percentage cannot tell you whether the improvement came from the AI, from a headcount change, from a seasonal spike, or from a process change made in the same quarter. Third, no direction. A productivity percentage cannot distinguish between staff doing rote work faster and staff doing more high-value work — and those two outcomes are opposites in workforce terms. The first is automation of the wrong work; the second is redesign.

WDG(JR+) grant evidence requirements reject percentage claims for exactly these reasons. So do finance committees, eventually.

The three metrics that work

The workforce analytics dashboard we build into every Qlabs engagement reports three metrics, and only three:

• Hours reallocated per role — hours moved from rote tasks to Anchor tasks, per role, per month. This is the metric that proves the redesign worked: it has direction. Hours moving into Anchor work is the redesign happening; hours merely shrinking is not.
• Skills-assessment closure rate — the percentage of the training cohort that completed and passed the redesign training. A redesign whose training nobody finished did not happen, whatever the dashboards say elsewhere.
• Revenue-per-FTE delta — quarterly, against the pre-engagement baseline. The commercial consequence of the first two metrics, reported slowly and honestly rather than monthly and optimistically.

How the baseline is established

The baseline is captured in week one, before any build: 90-minute task-audit interviews with each role-family lead, producing a task-frequency matrix per role, followed by two weeks of time-logging across the role family to validate the interview data against observed behaviour. Interviews alone over-report Anchor work; time-logging corrects for it. The two together are the baseline — and because they predate the first line of code, they are auditable.

How the dashboard works

The dashboard goes live in week eight, once the first surfaces are in production and the time-logging pipeline is feeding it. It reports hours reallocated per role monthly, skills-assessment closure as the cohort progresses, and revenue-per-FTE quarterly. A mid-term re-audit at month six compares the live task mix against the redesigned target in the Job Redesign Workbook; an annual re-certification repeats the task audit to keep the workbook honest.

A baseline captured before the build, a dashboard live by week eight, a re-audit at month six. That sequence is the difference between a claim and a measurement.

Why these metrics satisfy WDG evidence requirements

Map them: role redesign documentation is evidenced by the hours-reallocated metric per redesigned role; the training plan by the skills-assessment closure rate; the productivity baseline by the week-one task-audit matrix; measurable outcomes by the revenue-per-FTE delta at quarterly re-audit. Each grant dimension has a metric with a timestamp and an owner. That is what "evidence" means in an assessment — not enthusiasm, not percentages, but artefacts with dates.

Start measuring before you start building

The ordering is the whole method: baseline first, build second, dashboard by week eight, re-audit at month six. Organisations that start measuring after the build can only ever produce reconstructions — and reconstructions fail assessment.

If you want AI investment that generates auditable ROI evidence, email hello@qlabshq.com.

A worked example: reading the dashboard

Take one redesigned role — the Advertiser Solutions Consultant — and read its dashboard line. Baseline, week one: 15% of role hours on Anchor work. Month one post-launch: 31%. Month three: 48%. Month six re-audit: 65%, matching the redesigned target in the Job Redesign Workbook. The line between those points is the redesign happening; the plateau at target is the redesign holding.

Now read the same role's companion lines. Skills-assessment closure reached 100% of the cohort by week ten — the training landed. Revenue-per-FTE for the role family turned positive at the second quarterly read. Three lines, one story, every point dated. A percentage claim cannot be read this way because it has no anatomy.

Why revenue-per-FTE is quarterly, not monthly

The temptation in every dashboard build is to report the commercial metric monthly, because monthly movement feels like progress. Revenue-per-FTE does not move monthly in a way that means anything: sales cycles, seasonality, and cohort timing swamp the signal. Reported monthly, it becomes noise that executives argue about; reported quarterly against baseline, it becomes evidence.

The discipline matters for a second reason: grant evidence. A metric that is re-cut monthly invites the suspicion of re-baselining — of moving the comparison point until the number flatters. A quarterly cadence fixed in the workbook before the engagement starts removes the suspicion by removing the choice.

Attribution: keeping the metrics honest

The hardest question in workforce measurement is attribution: did the hours move because of the redesign, or because of the reorg that happened in the same quarter? Three practices keep the dashboard honest. First, the baseline is per role and per task, so reallocation is visible at task level — a reorg moves headcount, not task mix inside a role. Second, the re-audit compares against the redesigned target in the workbook, not against last month, so drift and step-changes are distinguishable. Third, amendments to the workbook — new tasks, re-categorised tasks — carry dates and reasons, so the audit trail explains every discontinuity in the line.

Metrics with anatomy survive audit. Metrics with only a headline do not — and neither do the programmes they describe.

Instrumenting without surveillance

Time-logging fails when staff experience it as surveillance, and it becomes surveillance the moment it is individualised punitively. The baseline exercise logs at role-family level for reporting — individual entries exist only during the two-week calibration window and are aggregated before any manager sees them. The dashboard reports role lines, not person lines; hours reallocated is a role metric by construction.

That design choice is also what makes the metric defensible externally. A grant assessor asking how the baseline was captured is asking, among other things, whether the data was collected lawfully and proportionately — under PDPA, proportionality is not optional. Staff who understand that the instrument measures the role rather than the person log honestly; honest logs are the only kind worth having, because a calibrated baseline built on defensive logging understates rote work and overstates the redesign's later delta.

If you want AI investment that generates auditable ROI evidence, email hello@qlabshq.com.