Five tests for whether AI agents are doing useful work, not just keeping busy

AI agent adoption is rising faster than measurable business impact, argues Appier chief executive Chih-Han Yu. His proposed answer is to manage agents around outcomes, reliability, model choice, learning and compute costs, rather than count deployments and call it progress.

Watch Desk analysis

What happened

In a Forbes Technology Council essay published on 9 November, Yu says 40% of respondents at organisations with more than $1 billion in annual revenue were scaling AI agents in at least one business function, up from 27% a year earlier. Yet the share of respondents reporting a positive effect on earnings before interest and tax remained at 37%, according to the McKinsey survey he cites.

Yu sets out five principles for operating agents: define the business outcome before deployment; check decisions at important points and let systems flag uncertainty; route tasks to models suited to their complexity; build feedback into workflows so results can inform future decisions; and track compute use as systems scale. His essay argues that success depends on how agents are operated, not simply how many are switched on.

Why it matters

The distinction is practical. A deployment count says a system is being used; it does not show whether customers are better served, errors are falling or costs are justified. Defining the intended result first gives teams a clearer basis to expand, refine or retire an agent.

Yu also makes a useful case for matching oversight to the work. A routine classification task and a multi-step decision do not need identical models or checks. And if feedback never makes its way back into the workflow, an agent may repeat the same mistakes with admirable consistency.

Our read

This is an industry executive’s framework, not evidence that these five practices alone deliver business returns. Still, it gives organisations a more useful starting point than another dashboard celebrating agent counts. Pick an outcome, measure the whole workflow and keep an eye on the bill. Busy is not the same as better, even when the busy thing has a rather convincing interface.

What to watch

• Whether organisations define outcome measures before expanding agent deployments.
• How teams test complete workflows, rather than relying only on model-level benchmarks.
• Whether feedback and compute costs are tracked as agent use grows.

Discussion spark: When judging workplace AI, should companies prioritise measurable business outcomes even if that slows deployment, or accept faster adoption while the benefits are still unclear?

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.