Microsoft published a practical guide this week walking anyone through building a working AI agent inside Microsoft 365 Copilot, and the notable detail is that it requires zero coding. You open the Agents section, click New Agent, describe what you want in plain language, and Copilot drafts a working version for you to test and adjust.
That is a real shift from a year ago, when building any kind of automated assistant meant hiring a developer or learning a scripting tool. Now the job looks more like writing instructions for a new employee: define the problem, decide what information the agent can see, set its tone, and tell it what to produce, a report, a spreadsheet, or a reply to a customer email.
The distinction Microsoft draws between a chatbot and an agent is the one worth remembering. A chatbot answers a question when you ask it. An agent works on its own: it can watch a shared inbox overnight, sort messages by urgency, draft replies to routine questions, and flag anything sensitive for a person to check in the morning. That is the difference between a tool you use and a worker you supervise.
The demand for this is already large. Survey data from PwC found that a large majority of companies say agents are already running somewhere in their business, and most of those say the agents are producing a measurable lift in productivity. Executives are not waiting for a mature version of this technology, they are building with what exists today.
But cheap and fast agent building has a cost that shows up later, not now. Gartner has estimated that more than 40 percent of agentic AI projects will be canceled by the end of 2027, pointing to rising costs, unclear payoff, and weak controls over what these agents are allowed to do. Separately, researchers tracking what they call agent sprawl have found companies including Lyft, GitLab, and FICO already dealing with employees who built agents faster than IT departments could track them, each one holding its own set of permissions to company email, documents, and systems.
That second problem deserves attention. An agent that reads a team's emails and drafts replies is only as safe as the access it was given and the instructions it was written with. If several people on a team each build their own version with slightly different rules about what counts as sensitive information, nobody has a clear picture of what company data is moving through which agent.
The sensible way to use this feature is to treat agent building the way you would treat hiring someone new. Write down what the agent is allowed to see, check its work regularly, and keep a simple list of which agents exist and who owns each one. The building part just got easy. The part that still needs a person is deciding what these things should and should not be trusted to do on their own.