AI agents: what changes when AI stops answering and starts acting

For two years, AI in business was mostly a chat window: you asked a question, got some text back, and someone copied it into another system. In 2026 the question has become what AI can do on its own, inside the company’s systems. That is what we call an agent.

What the data says about capability

The most widely cited measurement comes from METR, an independent institute that evaluates AI models. Rather than asking whether a model gets a question right, it measures the length of the tasks a model can complete, expressed as the time a human professional would take to do them. The conclusion is clear: for around six years, that length has doubled roughly every seven months (Kwa et al., 2025).

The length of the tasks AI can complete has been doubling about every seven months. The measurement, however, uses a 50% success threshold.

The detail matters. The threshold used is completing the task in half of the attempts. A system that succeeds half the time is impressive in a laboratory and inadequate in an operation.

What the data says about reliability

Researchers at Carnegie Mellon built a simulated software company, with virtual colleagues, documents, messages and real tools, and set agents to work inside it. The best agent tested completed 30% of the tasks autonomously (Xu et al., 2025). Simple tasks were often solved; long ones with several dependencies, rarely.

On the market side, Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027. It gives three reasons: rising costs, unclear business value and inadequate risk controls (Gartner, 2025). It is an analyst’s forecast, not a measurement, but it captures well what happens when an agent is launched without a defined scope or rules.

What this means in practice

Taken together, these three sources tell a consistent story. Agents are improving quickly and already do useful multi-step work. But reliability falls as tasks get longer and more complex. Three design rules follow:

1. Short, well-defined tasks first. For example, reconciling a delivery note with an order, preparing the invoice for a trip or spotting what is missing from a claim. On tasks like these the success rate is high and errors are easy to catch.
2. On sensitive actions, a person decides. The agent gathers the information and makes a proposal; a person approves payments, contract changes and communications with customers.
3. Know when to stop. When information is incomplete or contradictory, the agent hands the case to a person, with the context already gathered. An agent that recognises when it doesn’t know is easier to supervise and deserves more trust.

The difference between answering and acting

An assistant saves the person asking a few minutes. A well-designed agent takes a whole task off someone’s list. What separates the two is mostly design: the scope, the permissions, the approval points and the record of every step. These decisions also address one of the reasons for cancellation that Gartner points to, inadequate risk controls.