The core promise of enterprise AI was automation: buy the tool, cut the workflow, remove the headcount. A South China Morning Post report on corporate adoption points to a different outcome — companies are now managing AI systems the way they manage people, as digital workers that require oversight, assignment, review and correction.

That shift reframes a budget line. Software licenses scale without supervision. Digital workers, in the form of AI agents that take actions across systems, demand something closer to people management: role definitions, task delegation, output checks and intervention when behavior drifts.

What does 'digital worker' mean in practice?

A digital worker is an AI agent — or a fleet of them — assigned ongoing responsibilities rather than one-shot tasks: drafting responses, moving data between systems, executing multi-step processes. Unlike traditional automation scripts, these agents operate with probabilistic outputs. They can be wrong in ways a script cannot.

The distinction matters for product teams building on top of them. Deterministic automation fails loudly and visibly. Agentic systems fail quietly, producing plausible work that is subtly incorrect unless someone reviews it.

Where does this leave managers?

The report's framing implies a new middle layer of work between the AI and the business outcome: humans who allocate tasks to agents, monitor performance and step in on exceptions. This is supervision labor, and it does not disappear as adoption widens — it grows with the number of agents deployed.

The failure mode is familiar to anyone who has run an offshore team or a complex vendor relationship. Delegation without a feedback loop produces drift. Agents optimized for narrow objectives will pursue those objectives in ways the business did not intend, and the cost of catching that late exceeds the cost of the labor saved.

The conditions under which this model works are specific: tasks must be checkable, failure must be cheap to detect, and a human must hold authority to stop or correct the agent. Where verification is expensive — subjective quality judgments, ambiguous edge cases, high-stakes irreversible actions — the digital-worker model degrades into old-fashioned manual work with extra steps.

Why the automation pitch understated the problem

Vendors sold AI as a replacement for labor. What enterprises received is closer to leverage on labor: fewer people doing the execution, more people doing the oversight. The net headcount math depends on how much supervision each agent requires, and that ratio is still poorly measured in most organizations.

For product managers, the practical read is this: roadmaps that assume AI agents remove operational burden need a corresponding line item for the operating burden they create — monitoring tooling, escalation paths, evaluation criteria and the staff time to run them.

The direction of travel is toward agent fleets large enough that managing them becomes a discipline of its own, with tooling and roles to match.