AI on the job: what do your metrics reward?
Introducing AI at work often comes with output targets: more documents, responses or proposals. The precise management question is this: how can managers set a goal for AI-assisted work without rewarding volume at the expense of the service delivered?
What the literature establishes
Edwin A. Locke and Gary P. Latham (2002, American Psychologist) synthesise research showing that specific, challenging goals can support performance under certain conditions, including commitment, ability and feedback on results. Goals direct attention and effort towards the dimensions identified as important. Applied to AI, this framework invites us to examine what a target prioritises: producing a response or actually resolving a request.
The gap between intention and reward
Steven Kerr (1975, Academy of Management Journal) analyses situations in which organisations reward behaviours different from those they claim to want. His article offers an analysis illustrated with examples, not an experiment on AI. The mechanism nevertheless provides a relevant hypothesis: if assessment prioritises the number of deliverables, employees may have an incentive to generate more outputs rather than check their usefulness.
The hasty conclusion
One might conclude that numerical targets should be removed whenever AI is involved. These studies do not justify that conclusion: an explicit goal can be useful, while a vague assessment of “quality” leaves other ambiguities unresolved. The issue is to align the stated target, the criteria for accepting the work and what actually counts in performance assessment. (our executive and employee training programmes)
What these studies cannot promise
Neither article addresses generative AI, and neither demonstrates that changing a metric will automatically improve a team’s results. In Lausanne, a medtech document, a teaching resource at a higher education institution, a communication from an international sports federation and a commercial proposal from a Lake Geneva region scale-up do not share the same acceptance criteria. A common volume metric may conceal these differences, but a poorly defined quality metric can conceal them just as easily.
A practical check in Lausanne
As part of SHR’s “IA on the job” programme, we propose that a Lausanne team run a two-week test on a single category of deliverables, with acceptance criteria written down before it begins. For each deliverable, record whether it is accepted at first review and the rework time required of its recipient, alongside the volume produced. Then calculate the proportion accepted without rework and the total rework time, and examine them alongside the metric used to assess the team. This check does not, on its own, measure AI’s causal effect; it helps establish whether the organisation rewards usable output or shifts work onto those receiving it. To go further: explore the AI on the Job training in Lausanne, or browse our executive and employee training programmes in Switzerland.
In pictures: AI on the Job in Lausanne
- ai on the job
- Lausanne
- research
- training lausanne