Why is task count the wrong measure?
Task count is the wrong measure for an AI employee because activity is cheap and easy to generate. A worker that drafts a hundred emails nobody sends has completed a hundred tasks and produced nothing. The number rises while value stays flat.
Productivity has to be measured at the outcome: what changed in the business because the work was done. For a Digital FTE that means the targets it was hired against, which are explicit in its workspace.
What are the three measures?
The first measure is outcomes against targets: transactions reconciled, exceptions resolved, pages published and their visibility, candidates shortlisted and time to shortlist, tickets resolved at first line. Each Digital FTE's Targets tab records the contract and the status of each item with evidence.
The second is approval rate: the share of proposals approved without change, with send-backs and their reasons. The third is cost per outcome: the tokens consumed by the role divided by the outcomes it delivered, visible because usage is reported per FTE.
- Outcomes: against explicit targets, with evidence per item
- Approval rate: approved unchanged, edited, sent back, with reasons
- Cost per outcome: tokens per role divided by outcomes delivered
Why does approval rate matter?
Approval rate matters because it distinguishes work that is done from work that is right. A Digital FTE with a rising approval rate is producing proposals the manager agrees with, which is the basis for promotion on the autonomy ladder. One with a falling rate is drifting from what the team wants, and the send-back reasons say why.
It is also the honest measure of the human time the FTE costs. A high approval rate means approvals take seconds; a low one means the manager is editing, which is time that belongs in the cost side of the comparison.
How do you calculate cost per outcome?
Calculate cost per outcome from the workspace's usage per FTE and the outcomes recorded in Targets over the same period. As a labelled example: if a Finance Controller Digital FTE consumed a certain number of tokens in a month and reconciled a certain number of transactions with a certain number of exceptions resolved, the cost per reconciled transaction is the tokens divided by the count, converted at your plan's rate.
Do this per role, because roles differ: a page that passes the fact gate costs more tokens than a reconciled transaction. Compare each role's cost per outcome with what the outcome costs another way, from a baseline recorded before the pilot.
What does honest reporting look like?
Honest reporting includes what did not improve and what is unmeasured. The Performance tab on ftes.ai reports targets met, targets missed, and metrics that cannot yet be measured, rather than showing only the good news. A monthly review written by the Digital FTE in the first person states what it did, what worked, and what it recommends changing.
A productivity report that shows only gains is a marketing document. One that shows the misses and the unmeasured is the one a finance leader can use.
What does the monthly productivity report look like?
The monthly productivity report is the Digital FTE's own first-person review in its Performance tab, read alongside the record. It states each target and its status with the evidence: transactions reconciled, tickets resolved at first line, pages published and their citation presence, candidates shortlisted and time to shortlist, whichever the role owns. It states the targets it missed and why, and the metrics it cannot measure, labelled as unmeasured rather than estimated. It ends with what it recommends changing.
Next to it the manager reads three numbers: the approval rate with send-back reasons grouped by category, usage in tokens for the role, and cost per outcome from the two. The report is honest by design because the performance view shows only what the platform measured, and it is useful because the three numbers point at decisions: a falling approval rate in one category points at the remit, a rising cost with flat outcomes points at the model tier or the scope, and a target met with evidence points at a class of action ready for the autonomy ladder. Thirty minutes a month, from the record, is the whole review.