AI agents and chatbots: what changes when AI can act

What is an AI agent, in plain terms?

An AI agent is a system that is given a goal, works out the steps, and carries them out using tools it has been given access to: a case file, a document library, an email draft, a workflow queue. A chatbot does one thing: it takes a question and returns text. A person then decides what to do with that text.

The difference is not intelligence. The same model can sit behind both. The difference is reach. A chatbot's output stays on the screen. An agent's output can become a changed record, a sent message or a routed case.

That is why we describe the shift as one from accuracy to authority. With a chatbot, the risk is a wrong answer that a person may or may not catch. With an agent, the risk is a wrong action taken on the organisation's behalf.

How does the same request look when each one handles it?

The chatbot was not wrong; it was general. The agent did the work an officer would otherwise do, and stopped where a person needed to decide. On the Delegation Ladder, Cader Group's framework for AI authority, this agent sits at Draft with binding consequences: a reply reaches a member of the public, so a named officer signs before it leaves.

Where do agents fit today, and where don't they?

Agents fit work that is frequent, follows a known procedure, draws on records the organisation already holds, and has a natural point where a person reviews the result. Enquiry handling, document checks, case triage and report preparation are typical.

They fit poorly where the procedure is unclear, the data is scattered or unowned, or the value of the task has never been measured. The research forecasts both growth and attrition. Gartner forecasts that at least 15% of day-to-day work decisions will be made autonomously by agentic AI by 2028, up from 0% in 2024. It also forecasts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls (Gartner, June 2025; forecasts).

The second forecast names the failure modes worth planning against: value that was never defined, and controls that were never built.

What controls does an agent need?

Three, built in from the start rather than added after a demonstration.

• Approval points. A defined step where a named person reviews and approves before anything binding happens: before a reply is sent, a record is changed or a case is closed. Where a mistake is recoverable, review can move to exceptions over time, on evidence.
• Logs. A record of what the agent read, what it produced, which tools it used and who approved it. Logs make errors traceable and give the organisation the evidence it needs to move an agent up, or down, a level.
• Limits. Which systems the agent can read, which it can write to, what it may never do, and when it must hand over to a person. An agent that drafts replies does not need permission to send them.

Approval points matter more than they first appear, because people tend to trust a confident system. In a randomised study of 758 consultants, those using AI were 19 percentage points less likely to reach the right answer on a task the AI handled badly (Dell'Acqua et al., Harvard Business School working paper, 2023). An approver who signs without reading is not a control.

What must a team learn to build one?

Not programming. With today's no-code platforms, the skills that decide whether an agent works are about the task, not the code:

• Mapping the work. Breaking a task into steps, and marking which ones need judgment.
• Preparing the sources. Knowing which documents and records the agent will read, and who keeps them current.
• Writing instructions and limits. Telling the agent what to do, what not to do and when to stop.
• Testing on real cases. Running the agent on the organisation's own requests, including awkward ones, and checking the output.
• Placing the approval point. Deciding where a person signs, and making sure the approver has time to read.

Skills in this area are still uncommon. In one vendor benchmark, 13% of employees were assessed as accomplished in agentic AI before training (Workera, 2026 AI Skills Enterprise Benchmark; Vendor data). Our Agentic Build program is three or eight days and ends with a working, demoed agent and a pilot handover plan naming the owner, approval route and next steps. For more on who can do this work, see Can non-technical staff build AI agents?

Measure where your own team stands: take the free Readiness Snapshot →