AI Agents Move From Demo to Real Work

The chatbot answered; the agent acts. The difference sounds slight, yet it is vertiginous. An agent receives a goal — assemble a dossier, reconcile invoices, organise a trip — then decomposes, executes, checks and retries until the result is reached, using digital tools the way a colleague would: email, spreadsheets, business applications.

After a year of carefully staged demos, the first serious deployments are painting a contrasted picture. On well-bounded processes — data entry, sorting, matching, document synthesis — agents plough through a remarkable and steady volume of work. Set loose on vague objectives, they wander just as steadily, piling up plausible but useless actions.

Managing the Inhuman

The stakes are shifting toward an unexpected skill: knowing how to delegate to a machine. Defining the perimeter, setting checkpoints, deciding what requires human sign-off — these are management gestures, applied to a worker that never sleeps but understands nothing outside its frame.

The organisations that succeed are not those that automate the most, but those that carve up work the best. The agent is neither an oracle nor an intern: it is a new kind of workforce, one that forces us to relearn the old art of handing over a task.