From answering to acting

A chatbot responds to a message. An automation script follows fixed steps. An AI agent sits between the two: it is given a goal, works out the steps, uses tools to gather information or take action, checks the result and decides what to do next.

Ask a chatbot about an order and it tells you the returns policy. Ask an agent and it can look up the order, confirm it qualifies for a change, make the change within its limits and send the confirmation.

The parts of an agent

  • Model: the language model that reads, reasons and chooses the next step.
  • Instructions and skills: how the work should be done, written clearly enough to follow and to test.
  • Tools: permissioned connections to systems such as a CRM, calendar, database or email, each with defined inputs and outputs.
  • Knowledge: retrieved documents and records that ground the agent in your facts.
  • Memory: what it needs to remember within a task and, where appropriate, between tasks.
  • Guardrails and checkpoints: limits enforced in code, and the points where a person must approve.

Autonomy is a dial, not a switch

Agents do not need to be fully autonomous on day one. A practical path has four levels:

  • Suggest: the agent drafts, a person decides and acts.
  • Act with approval: the agent prepares the action and a person approves it with one click.
  • Act within limits: the agent acts alone below set thresholds and escalates above them.
  • Act and report: the agent runs routinely while people review samples and exceptions.

Move up a level only when evaluation data shows the agent is reliable at the current one. Different actions can sit at different levels: an agent may update a CRM record on its own but still need approval to issue a refund.

Where agents fit best

Agents shine when work involves unstructured inputs such as emails, chats and documents, touches several systems, and calls for judgement that follows recognisable patterns. Lead qualification, support resolution, invoice processing and scheduling are typical examples.

They are a poor fit where a simple rule would do, where every case is genuinely unique, or where there is no way to check whether the result was right. In those cases, plain automation or a human expert is the better tool.

Making agents safe to run

  • Give each agent the narrowest tool and data access it needs.
  • Enforce limits such as amounts, record types and recipients in code, not only in the instructions.
  • Treat content from outside, like emails and web pages, as untrusted, so hidden instructions cannot redirect the agent.
  • Log every run: inputs, sources, tool calls, outputs, confidence and cost.
  • Test against historical cases before going live, and re-test after every change.
  • Name a person who owns exceptions and reviews performance.

Measuring an agent

Track the same numbers you would for a team: volume handled, share resolved without escalation, accuracy on sampled reviews, time to resolution, cost per task and customer or employee satisfaction. Compare each with the baseline you measured before launch, and review the trend every week in the early months.