Technology
AI Agents Move From Chatbots to Coworkers
For two years, office workers have treated artificial intelligence as a clever assistant: ask a question, get an answer, copy the result into a document. That era is ending. A new generation of AI agents can now plan a task, open the right applications, gather information from several sources and deliver a finished piece of work with only light supervision. They are showing up in law firms, marketing departments, logistics companies and local government offices, and the change is less dramatic than the headlines suggest but more consequential than most managers expected.
The difference is persistence. A chatbot forgets the conversation when the window closes. An agent keeps a goal in mind across hours or days: research twenty suppliers, compare their prices, check delivery times, and draft a recommendation. It can send emails, update spreadsheets and flag the cases it cannot resolve. Early users describe the experience less like chatting and more like delegating to a diligent but literal-minded junior colleague.
Productivity gains are real but uneven. A survey of 400 companies that deployed agents this year found that teams saved an average of six hours per employee per week on routine research and drafting. Yet a third of the same companies said the savings were partly offset by new work: reviewing agent output, fixing confident mistakes and writing clearer instructions. "The agent does exactly what you asked," said one operations director at a mid-sized insurer. "The problem is that what you asked is rarely what you meant."
"The agent does exactly what you asked. The problem is that what you asked is rarely what you meant."
Accountability is the open question. When an agent books the wrong shipment or sends a client an outdated price list, who is responsible? Employment lawyers say the answer, for now, is the human who assigned the task. Several large employers have written policies requiring a person to approve anything an agent sends outside the company. Others have gone further, logging every agent action so mistakes can be traced and reversed.
Workers themselves are split. Some welcome the relief from repetitive work and say they finally have time for the parts of their jobs that require judgment. Others worry, reasonably, that the tasks being automated are the same ones that once taught newcomers the business. A junior analyst who never compiles a report by hand may never learn how the numbers fit together. Training programs are being rewritten to teach oversight: how to check an agent's sources, how to spot a plausible-looking error, and when to take a job back.
What comes next is less about capability than about trust. The technology is already good enough for many routine jobs. Adoption will be decided by whether organisations can build the habits around it: clear instructions, careful review, honest accounting of errors. The agent is not a colleague. But in a growing number of offices, it is already a coworker, and the job of managing it belongs to everyone.