AI agents are being sold as workers, but proof is missing

Vendors are aiming agents at work, not chat

Dünya Gazetesi and the South China Morning Post describe AI agents being positioned for production and operational work rather than only chatbot-style use. The wording matters. Positioning is a market activity: it tells buyers what a product is meant to become, not what it is doing today.

The same reports point to a shift in the problem. Once software is described as a worker, the buyer’s question changes from “can it answer?” to “can it be trusted with a task?” That moves attention to permissions, supervision, audit trails and failure handling. Those are unglamorous disciplines, but they determine whether an agent can survive an ordinary Tuesday.

The gap is not between chatbots and agents, but between positioning and production.

The optimistic case is that agents are becoming operational tools. The cautious case is that the proof has not arrived.

The distance between these two states is easy to miss in vendor language. A product can be introduced as a digital worker before any employer has given it real duties. In that sense, the current market resembles an early hiring round: lots of résumés, few pay slips.

The gap is not between chatbots and agents, but between positioning and production.

Security and management are the new bottleneck

Anadolu Agency’s security framing and the Post’s management framing converge on the same operational point: agents that act on behalf of a company create a governance surface. A chatbot can be wrong in a visible way. An operational agent can be wrong in a quiet way, by taking an action, opening a ticket, changing a record or triggering a workflow before anyone notices.

That is why the language of “digital workers” matters, even if it is vendor terminology. It implies a lifecycle: onboarding, access rights, monitoring, review and retirement. In that framing, the hard part is not the model alone. It is the fleet around the model — the identities, logs, controls and escalation paths that let a business use many agents without losing track of them.

The available descriptions do not show a named organisation running such a fleet in live production. They show that the industry is selling the category and beginning to build the support layer around it. Those are different facts.

This distinction matters because governance costs arrive before returns. If an organisation must map processes, assign access, monitor behaviour and train staff before an agent earns its keep, the economics depend on repeated, reliable work. A demo can hide that burden. A production environment cannot.

Training is appearing before proof

One part of the support layer is training. Material reviewed for this story describes training on the use of AI agents in cybersecurity. That is a small but telling detail: if agents are going to touch sensitive systems, someone has to teach staff how to use them safely, and someone else has to decide what safe means.

Cybersecurity training can mean many things, from a short webinar to a formal curriculum. The descriptions do not specify scale, certification or outcomes. Still, its presence suggests that the market is not only selling agents; it is also selling the habits needed to keep them from becoming liabilities.

The missing piece is a live example with numbers. Without it, the story remains directional: agents are being pushed toward operational roles, and the surrounding controls are being discussed. Whether that becomes routine management is still unresolved.

The open question is whether this support layer matures faster than the evidence base. If named deployments, audited results and measurable outcomes appear, the “digital worker” label will start to look like an operational category. Until then, it is best read as a direction of travel, not a destination.