AI agents explained: what they are and where to start

The latest AI phrase is "agents". Software vendors are adding them to their products, and you may have been asked whether your business has a plan for them yet.

It's natural to wonder whether this is something you should be doing. The good news is that the idea is simpler than the noise suggests, and the sensible way to begin is small, careful and well within reach of an ordinary business.

This article explains what an AI agent is, how it differs from the chat assistants you may already use, where agents genuinely help today, and how to pick a good first candidate.

Chatbot, assistant, agent: what's the difference?

The words get used loosely, so it helps to set out a simple distinction. The difference is mostly about what the AI is allowed to do.

An assistant gives you an answer. An agent does some of the work: it can open a document, search a system, fill in a form or prepare an email, decide what to do next based on what it found, and keep going until the task is done or it needs a person.

What an agent is made of

Behind the marketing, most agents have the same four ingredients:

• A goal. "Prepare a draft quote for each new enquiry" is a goal. "Help with sales" is not.
• Tools. The things the agent is allowed to use: reading an inbox, searching a price list, creating a draft in the accounts system. Each tool is a door, and you decide which doors are open.
• Instructions. How you want the work done: your pricing rules, your tone, what to do when something is unclear.
• A way to check. A list of what "finished" looks like, so the agent can tell whether it has done the job, and a point where a person reviews the result.

Compared with the small automations we wrote about recently, the difference is flexibility. A traditional automation follows a fixed rule: when X happens, do Y. An agent can handle more variety, because it reads the situation and chooses its next step. That's what makes agents useful, and why they need sensible limits.

Where agents help a small business today

The best early uses are tasks that are frequent, follow a recognisable pattern, and end with something a person can quickly check. A few realistic examples:

• Triaging enquiries. An agent reads incoming emails or web forms, works out what each one is about, looks up whether the sender is an existing customer, and sorts them into the right queue with a short summary and a suggested reply.
• Preparing quotes for approval. For a business with a price list and standard options, an agent can turn an enquiry into a draft quote, flag anything unusual, and leave it for someone to check and send.
• Chasing paperwork. An agent checks which certificates, timesheets or signed documents are overdue, drafts polite reminders, and lists who has been chased and when.

In each case the agent does the gathering, sorting and drafting, and a person makes the decision. Many businesses find most of the time saving sits there, because the preparation is the slow part.

You may not need to build anything to try this. Several mainstream tools now include agent features:

• ChatGPT agent can browse websites, work with files and connected apps, and carry out multi-step tasks. OpenAI says it asks for permission before actions with real-world consequences, such as making a purchase.
• In Microsoft 365, Agent Builder lets you create simple agents in plain language, with knowledge such as SharePoint content. What an agent can reach depends on your licences: grounding agents on your organisation's files needs a Microsoft 365 Copilot licence or pay-as-you-go billing. We supply Microsoft licences and know these tools well, but the right choice depends on your business.
• Google's Workspace Studio lets Workspace users describe an agent in plain language and have Gemini build it, working across Gmail, Docs and Drive and some third-party services.
• Anthropic's Claude Cowork takes on multi-step tasks on your behalf, and is available on Claude's paid plans.

Features and plans change often, so check the current details for any tool before you commit.

What agents aren't good at yet

Agents are promising, and it's worth being clear about their limits so you choose the right first task.

• Judgement calls. Anything involving a relationship, a negotiation or a sensitive conversation still belongs with a person.
• Long, open-ended tasks. The longer and vaguer the task, the more room for a small misunderstanding to carry through.
• Messy information. An agent is only as good as what it can find. If price lists are out of date or files are scattered, it will struggle, just as a new member of staff would. Our article on getting your business information ready for AI covers the housekeeping that helps.
• Being trusted blindly. Agents can make mistakes with confidence, so checking their work belongs in the design from the start.

Sensible guardrails

A few simple principles keep an agent safe and useful. They're the same ones you'd apply to a new starter.

• Give it only the access it needs. If the job is drafting quotes, it doesn't need to send emails or change prices. Read-only access is a good place to begin.
• Keep a person in the loop for anything that matters. Sending to a customer, spending money, changing records and deleting anything should wait for someone to approve it.
• Be careful with content from outside. Emails, attachments and web pages can contain text that tries to give an agent new instructions. The UK's National Cyber Security Centre explains this risk, known as prompt injection, and recommends limiting what a system can do rather than relying on the AI to spot it. Limited access and human approval do exactly that.
• Keep a record. Choose tools that log what the agent did and why, so you can review its work and learn from it.
• Follow your AI policy. The same rules about approved tools and personal data apply. If you don't have one yet, our article on an AI use policy your team will actually read is a good starting point.

How to spot a good first candidate

A good first agent task usually ticks most of these boxes:

• It happens often, at least weekly
• It follows a recognisable pattern, even if the details vary
• The information it needs is in one or two well-kept places
• The output is a draft, a summary or a sorted list that a person can check in a minute or two
• A mistake would be easy to spot and cheap to fix

Enquiry triage and paperwork reminders often fit well. Anything that touches payments, or that a customer sees without review, is better left until you've built some confidence.

A sensible first step

This week, list three tasks that someone in your team does repeatedly and that involve gathering information before a decision. For each one, ask: could a capable assistant prepare this for us to check? Pick the one with the lowest risk and the clearest finish line.

Then try it with a tool you already have, on real but non-sensitive examples, with the agent producing drafts only. Run it alongside the normal process for a couple of weeks and compare the results.

If you have developers in your team or among your suppliers, our technical companion article on building a first agent with a safe architecture follows later this month.

If you'd like help choosing a good first task or setting up a careful trial, our AI Adoption service is designed for exactly this, and we're always happy to talk it through.