Why do single-metric targets go wrong?
Single-metric targets go wrong because an AI employee optimises exactly what it is measured on, without the social instinct that stops a person from technically satisfying an instruction. Set publish four articles a week and you will get four articles a week regardless of whether any of them were worth writing.
This is not a flaw so much as a property to design around. The system is doing what it was asked; the instruction carried an unstated assumption that the work would also be good.
What is a counterweight, and why does every target need one?
A counterweight is a second measure that makes gaming the first one unprofitable. Pair volume with a quality or cost measure so that satisfying one at the expense of the other fails the pair.
Publishing volume pairs with a measure of whether published pages actually earn visibility. Ticket resolution pairs with reopen rate or satisfaction. Meetings booked pairs with meetings held. Spend efficiency pairs with an absolute volume floor, so the cheapest strategy is not simply spending nothing.
- Output volume pairs with a quality or acceptance measure
- Speed pairs with an error or rework measure
- Efficiency pairs with an absolute volume floor
- Coverage pairs with a depth or accuracy measure
Should targets measure things it fully controls?
Targets should measure outcomes the system influences rather than outputs it controls outright, because fully controlled outputs are trivially satisfiable and say nothing about value.
The balance matters in both directions. A target entirely outside the system's influence, such as total company revenue, produces noise rather than direction. The useful band is close enough that its work moves the number and far enough that it cannot simply manufacture it.
How many targets should a role have?
Three to five targets is the working range for a single Digital FTE. Fewer than three usually means an unbalanced pair; more than five dilutes prioritisation and makes trade-offs invisible.
Rank them explicitly rather than presenting them as equals. When two targets conflict, and they eventually will, the ranking is what determines the decision, and an unstated ranking becomes whatever the system inferred.
How do you set the starting level?
Set the starting level from the recorded baseline rather than from ambition, and record that baseline before the first run. Without it you cannot distinguish progress from normal variation.
For a first target, a modest improvement on the baseline is more useful than a stretch. The purpose of the first cycle is to establish that the measurement works and the mechanism responds; aggressive targets set before anyone trusts the measurement tend to produce arguments about the measurement.
How often should targets change?
Review targets monthly and change them deliberately rather than continuously. A target that moves every week is not a target; it is a running commentary, and the system cannot plan against it.
At each review ask three questions: does this target still matter, is it still measurable the same way, and has the counterweight held. Retire targets that have stopped mattering rather than accumulating them, because a role carrying nine legacy targets is a role with no priorities.
What do good targets look like for two different roles?
For a Search Visibility Digital FTE, a good target set as a labelled example might be: pages published that pass the fact gate, with a counterweight of send-back rate below a level the manager sets; citation presence on a fixed prompt set sampled weekly, with a counterweight that no page carries an unsourced claim; and search console impressions for the target queries where the console is connected. Each item lives in the Targets tab with a status and the evidence behind it, and the FTE's monthly report states which were met and which were not.
For a Recruiter Digital FTE the set is different in content and identical in shape: time from application close to shortlist, with a counterweight of hiring-manager send-backs on the shortlist; share of interviewed candidates who met the hard requirements; and candidate communication drafted within a target window. Nothing in either set measures a decision about a person, because those stay human. Targets that follow this shape, an outcome with a counterweight and evidence, are what make the monthly review a reading of the record rather than a debate.