Discussion

AI agents need a performance review, not just a cloud bill

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#14481

Companies can count what their AI agents cost, but may struggle to say what work they actually get in return. A Forbes Technology Council essay proposes measuring agents against verified results, quality and the human effort still needed to supervise them, rather than treating activity or a “task complete” message as proof of success.

Watch Desk analysis

What happened

Rotem Alaluf, founder and chief executive of enterprise AI provider Wand AI, argues that businesses should manage AI agents as part of the workforce, not just as lines in a software budget. His proposed “agent-hour” records an agent’s time under a defined objective and governance controls, including its authority, actions and escalations.

That measure is an input, not a verdict. Alaluf says companies should also track verified output, quality, rework, escalations, human review effort and total cost. In a reconciliation team, for example, hours alone say little; the number of exceptions resolved, escalated or reopened gives a clearer picture of whether the work was useful.

His practical starting point is narrow: choose one business function, give each agent a named owner and defined objective, then compare its governed hours and costs with evidence of results. A completion message is not that evidence. Check the ledger, completed transaction or independently reviewed sample instead. Alaluf sets out the argument in a Forbes Technology Council essay published on 18 November.

Why it matters

A company can spend less on staff and still create more work if people have to repair, check or repeat what an agent produces. Conversely, agents might make previously neglected work affordable. Looking only for jobs replaced misses both outcomes; looking only at model fees misses the work itself.

The distinction is useful as businesses move from trying out chatbots to assigning systems ongoing responsibilities. If managers cannot connect an agent’s activity to independently checked results, they have a poor basis for deciding whether to expand its remit. A busier dashboard is not a performance review, however confidently it says “done”.

Our read

Alaluf’s strongest advice is also the least futuristic: start with one function, give the agent a real owner, and verify the result outside the agent’s own report. The proposed agent-hour may help track capacity, but it should sit beside outcomes and oversight, not become a new badge of productivity. Conveniently, counting hours is much easier than proving value. That is precisely why the two should not be confused.

What to watch

  • Whether companies publish measures linking agent costs to independently verified outcomes.
  • Whether they include rework and human supervision when reporting AI productivity.
  • Whether agent-hour accounting clarifies performance or simply adds another dashboard metric.

Discussion spark: If a company cannot verify an AI agent’s claimed result without a person checking it, should that work count as AI productivity or as supervised human work?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.