AI agent

Part of our topic guide on AI Skills for Business.

An AI agent is software built on an AI model that plans and carries out a multi-step task towards a goal, using tools, data or other systems, with limited step-by-step human input. That's different from a chatbot, which answers one prompt at a time and waits for you to direct the next step.

Chatbot to agent: where the line sits

Give a chatbot a question and it gives you an answer. Give an agent a goal, "find the three suppliers with the best on-time delivery this quarter and draft a comparison", and it breaks that goal into steps, pulls data from the relevant systems, checks its own progress, and comes back with the result. You set the destination: the agent works out the route.

The line is worth holding onto, because "agent" is being stamped on a lot of software that sits on the chatbot side of it. A tool that answers a single question, however well, isn't acting as an agent. It's the multi-step, semi-independent working-towards-a-goal that earns the name.

The loop that earns the name

What separates an agent mechanically is a loop, made of a few components:

• A goal or instruction: what you want done, set by a person.
• A planning step: the agent breaks the goal into smaller tasks and decides an order.
• Tools and access: it can call other software, search data, run calculations, or trigger actions in systems it's connected to.
• A feedback loop: it checks whether each step worked and adjusts before moving to the next one.
• A handback point: well-designed agents flag when a decision needs a human, rather than acting on everything unsupervised.

The "agentic" part is the planning and feedback loop; the handback point is what makes one safe to use on work that matters.

What working alongside one actually takes

You'll meet this term a lot over the next few years, and the hype around it moves faster than most people's understanding of what it actually does. That gap matters commercially: most AI and data initiatives don't fail because the technology is missing, they fail because nobody in the business builds the judgement to point an agent at the right problem, check its work, and know when to step in.

You don't need to build agents to work well alongside them. The skills that count in non-technical roles are writing a clear goal with the right constraints, checking the agent's output before it's used or shared, and knowing when a task is too judgement-heavy to hand over at all. In our experience, the people who get the most value from agents are the ones who already understand the task well enough to spot when the agent has got it wrong. The skill sits in the person, not just the tool: a data-and-AI literacy question as much as a technical one, which is why we treat it as foundational rather than something only "technical" roles need.