What is an AI agent? Explained without the jargon

By Babagana Zannah, team lead · Published 4 July 2026 · Updated 6 August 2026

Babagana Zannah
Leads engineering teams; puts AI to work inside one of the UK's largest companies

"Agent" is a common technical word, but suppliers do not all use it in the same way. This guide explains the useful distinction: how an agent differs from a prompt-led chatbot, which workflows may be suitable, what controls matter, and when you do not need one at all.

No assumed knowledge. If a sentence here makes you re-read it, that's our failure, not yours.

What does an AI agent actually do?

An AI agent can be configured to work through a task step by step. It may read permitted information, choose among allowed next actions, use an approved tool, check a result and either continue or escalate. What it can complete — and what it records — depends on the workflow, access, rules and controls around it. The word "agent" does not mean that a job will always finish successfully.

The aim of a setup like this is to reduce repeated handling while keeping a person responsible for approvals and exceptions. Whether it achieves that aim has to be measured against the practice's own baseline; it is not an automatic outcome of using an agent.

What's the difference between an AI agent, a chatbot and ordinary software?

A prompt-led chatbot helps a person produce an answer. Rule-based automation applies predefined logic. An AI agent can combine language-model decisions with permitted actions across a workflow. The boundaries overlap — some products combine all three — so the useful questions are which actions are allowed, where human approval sits and what happens when the system is uncertain.

A chatbot answers you, then stops

A prompt-led product can be useful for drafting a reminder, summarising a document or exploring an idea. That does not by itself create a connected operating workflow: permissions, routing, approvals, records and exception handling still need to be designed. Product features and prices change, so check the supplier's current terms rather than relying on a generic comparison.

Ordinary software follows one rule, forever

Think of a mailbox rule that routes messages from one sender into a folder. Fixed rules can be dependable and proportionate when the input and destination are predictable. When a message is ambiguous, a person or a separately designed classification step may still be needed.

An agent handles the messy middle

An agent can be configured to classify bounded variation. For example, it may distinguish a likely document from a question and route uncertainty to a person instead of acting. Performance depends on the data, rules and test cases, so unusual formats and low-confidence decisions should be treated as exceptions rather than promised away.

What can an AI agent safely do in a small business?

Repetitive work with a clear owner, inputs, finish line and escalation path is usually the best place to investigate. A written process is a useful starting point, not proof that an agent can safely take it on. Access, data sensitivity, technical feasibility, error consequences and the required human checks still need to be assessed.

• Chasing documents and information. A scoped workflow may prepare approved reminders, track responses and route questions or missing items to a person.
• Sorting a defined mailbox route. It may classify in-scope messages, draft routine replies for approval and route exceptions; this is not a promise to manage an entire inbox.
• Preparing reports and packs. It may place permitted data or exports into an agreed template, with a person checking the result.
• Checking invoices and admin. Agreed rules may flag likely exceptions for review; professional or payment decisions stay with authorised people.

And what shouldn't an agent do? Anything that needs professional judgement, or carries real risk if it goes wrong: giving clients advice, signing work off, making final decisions about money or people. Those stay with your team. The agent's job is to clear the routine work out of their way, so they have more time for the judgement calls.

For a longer list of candidate jobs and the questions to ask before scoping one, see our guide to how AI can help your small business.

Why call them "AI helpers" instead of agents?

Because the word "agent" confuses more people than it helps. In Britain, an agent sells houses or spies for a living. "AI helper" says what the software actually is: something that helps a person do their work, without replacing them. Everywhere else on this site, "helper" means exactly what this guide calls an agent.

The industry will keep inventing new words for all of this. You don't need to learn them. If a supplier can't explain what their software does in one sentence your whole team would understand, be careful.

How do you stay in control of an AI agent?

Control has to be designed, not assumed. Agree which actions need approval, which data and tools the system may access, which events are recorded, when it must stop, and who owns each exception. Ask for those boundaries in writing before access is granted.

Approval in Quiet Quarter

Compatibility is scoped, not assumed

Logging, in practice. Define the action and exception events the operating record must capture, who may read them, how long they are retained and how incidents are reviewed. Quiet Quarter's agreed outputs can include chasers prepared or sent, replies classified, documents filed where technically feasible, tracker updates, exceptions and weekly summaries. That is a scoped operating record, not a claim that every internal model decision can be reconstructed.

If you want the fuller picture — where your data goes, who can see it, and what agreements sit underneath — it's all on our data safety page, written so you can forward it to your IT adviser.

When is an AI agent not what you need?

Quite often. If you mainly want help with writing and questions, a prompt-led tool may be enough. If the task follows one fixed rule, a built-in rule or checklist may be more proportionate. Use an agent only when a measured, repeatable workflow justifies the added setup, controls and operating cost.

• You want better letters, emails and summaries. Compare current prompt-led products and their supplier terms before paying for a connected workflow.
• Every email from one sender should go to one folder. A built-in mailbox rule may solve that without an AI build.
• The job only happens occasionally. The setup and control effort may be disproportionate. A checklist, template or one-off service may be enough.
• The job is done differently every time, and no two people agree how. Fix the process first. An agent pointed at chaos just produces faster chaos.

The honest test starts with evidence. Measure the current workflow, include review and exception time, then compare that baseline with the full setup and operating scope. Our guide to what AI automation costs a UK small business shows what to include without promising a generic payback.

Sources checked

Reviewed 6 August 2026 against current UK regulator and cyber-security guidance.

• ICO: guidance on AI and data protection
• NCSC: guidelines for secure AI system development

Questions owners ask about AI agents