What is an AI agent?

An AI agent is software that uses AI to carry out a task across several steps. You give it a goal, and it can find information, use connected business systems and decide what to do next.

For example, it could read a customer enquiry, check product availability and prepare a quote for you to review.

You choose which information it can access, which actions it can take and when it needs your approval.

An AI agent at work

Here is how an agent configured to handle quote requests might work through a customer enquiry.

Example: an agent set up for quoting

  1. Step 1: Receive an enquiry

    A customer emails asking for a quote.

  2. Step 2: Understand the request

    The agent identifies products, quantities and missing details.

  3. Step 3: Check business systems

    It looks up approved prices and availability.

  4. Is anything missing from the request?

    1. Yes, details are missing

      Step 4: Clarify missing information

      It prepares a question for the customer, such as which size or delivery date they need.

      When the customer replies, the agent goes back to step 2.

    2. No, the request is complete

      Step 5: Prepare the quote

      A staff member reviews the draft and approves it before sending.

Chat, automation or agent?

These are three ways of using software to get work done. The difference is who decides the next step.

Basic AI chat

You ask questions and work with the answers. You decide what happens next.

Example: “Help me write a quote.”

Fixed automation

Software follows rules and steps that were set in advance. It does the same thing every time.

Example: Every submitted enquiry creates a customer record.

AI agent

AI chooses the next step based on the task and what it finds along the way.

Example: It checks an enquiry, identifies missing information and prepares a quote using connected systems.

These are ways of working, not separate kinds of product. A chat window can also contain an agent: you type a request and it works through the steps for you.

If a task is straightforward and happens the same way every time, a fixed automation may be all you need. It is usually cheaper and more predictable than an agent.

What an agent is made of

Most agents are built from the same five parts. Here is each one, using the quoting agent as the example.

Instructions

Its job, your business rules and its boundaries.

Quoting example: Quote only from the approved price list. Never offer more than 10% discount.

Information

The documents and records it is allowed to access.

Quoting example: Your price list, product catalogue and delivery terms.

Tools

A tool is a connection that lets the agent look something up or take an action in another system.

Quoting example: Check stock in your inventory system. Save a draft quote in your CRM.

Permissions and approvals

What it can do on its own, and when a person must review first.

Quoting example: It can look up prices by itself. A staff member approves every quote before it is sent.

AI model

The AI that reads the request and chooses the next step.

Quoting example: It works out that “20 of the blue ones” means 20 units of a specific product code.

Going deeper

Optional reading. These are the technical ideas behind agents. Each has a short explanation and an example, with more detail if you want it.

The software that runs the agent

Agent harness

The AI model does not act on its own. Software around it gives it instructions, runs the tools it asks for, checks permissions, keeps a record and decides when to stop.

In the quoting example

In the quoting agent, the harness gives the model your pricing rules, runs the stock lookup when asked and stops once the draft quote is ready for review.

Technical detail: Agent harness

The harness runs a loop: send the task and context to the model, receive a request to use a tool, check it against permissions, run the tool, return the result, and repeat until the task is done or a limit is reached.

It also manages how much context the model sees, retries, time limits and logs. Claude Code, OpenAI Codex and Microsoft Copilot Studio each include their own harness, which is one reason the same model can behave differently in each.

Connecting AI to your business systems

MCP

The Model Context Protocol (MCP) is one common way to connect an AI application to other software. A system offers an MCP server that describes what the AI can look up or do.

It is one option, not a requirement. Agents can also connect through a system’s usual integration or API. With or without MCP, each connection still has to be set up, signed in and given the right permissions.

In the quoting example

Your inventory system offers an MCP server. Once it is connected and given read access, the quoting agent can check stock for each product in an enquiry.

Technical detail: MCP

Anthropic introduced MCP in November 2024 as an open standard. Since December 2025 it has been governed by the Agentic AI Foundation, part of the Linux Foundation.

An MCP server can offer tools (actions), resources (data) and prompts to an MCP client inside the AI application. Sign-in is optional in the standard; servers reached over the internet usually use its OAuth 2.1-based authorisation. Either way, the permissions you grant decide what the agent can do.

When agents work together

A2A

Sometimes one agent hands part of a job to another agent, such as a specialist agent run by a supplier. Agent2Agent (A2A) is an open standard for that hand-over.

Most businesses start with one agent. A2A only matters when agents from different systems need to work together.

In the quoting example

The quoting agent asks a freight company’s agent for a delivery price, then adds it to the draft quote.

Technical detail: A2A

Google introduced A2A in April 2025, and it became a Linux Foundation project in June 2025.

Each agent publishes an Agent Card: a short document describing what it can do, where to reach it and how to authenticate. Agents then exchange tasks, status updates and results using JSON-RPC, gRPC or plain HTTP. A2A connects agents to other agents; MCP connects an application to tools and data.

Finding answers in your business documents

RAG

Before answering, the system searches your documents for relevant passages and gives them to the AI. This helps answers reflect your own policies and prices.

It can show where an answer came from when it is set up to. It does not guarantee the answer is correct, so important answers still need checking.

In the quoting example

A staff member asks whether the trade discount applies to a large order. The system finds your pricing policy and answers with a link to the relevant section.

Technical detail: RAG

Retrieval-augmented generation (RAG) works in two stages. First, documents are split into passages and indexed, often as embeddings in a vector database. Then, when a question arrives, the most relevant passages are retrieved and added to what the model is given.

Answer quality depends on the documents themselves, how they are split, the search settings and keeping the index up to date. Access should match who is allowed to see each document.

Deciding when a person must approve

Human in the loop

You decide which actions an agent can take by itself and which need a person to review first. This is a setting that has to be configured. An agent only asks for approval where it has been set up to.

In the quoting example

The quoting agent can look up prices on its own, but every quote waits for a staff member to approve it before it is sent.

Technical detail: Human in the loop

Approval steps are usually built into the harness or workflow. The agent pauses, sends the proposed action to a person, for example in Teams, Slack or email, and continues only once it is approved. Each decision should be logged.

Good candidates for approval are actions that are hard to undo, such as sending to customers, making payments or deleting records.

Questions businesses ask.

What is an AI agent?

Software that uses AI to carry out a task across several steps. You give it a goal, and it finds information, uses the business systems it is connected to and decides what to do next, within the limits you set.

How is it different from an AI chatbot?

A chatbot answers your questions and you decide what to do with the answers. An agent can take steps itself, such as looking up stock or drafting a quote. A chat window can contain an agent, so the difference is what it is allowed to do, not how it looks.

Can it work with our existing software?

Often, yes, if the software can be connected. Many systems offer integrations, APIs or MCP servers. Each connection has to be set up and given the right permissions, and some older systems are harder to connect.

Does it run automatically?

It can. Depending on the setup, an agent starts when a staff member asks, when something happens such as a new enquiry arriving, or on a schedule.

Can we approve actions before they happen?

Yes, if it is set up that way. You choose which actions need approval, and the agent waits for a person before doing them. Approval does not happen by default, so it has to be configured.

What happens if it makes a mistake?

Agents can make mistakes, so limit what they can do and review important work. Approval steps, logs of every action and narrow permissions make mistakes easier to catch and fix. Start with tasks where an error is easy to spot.

Do we need an agent, or would a simple automation work?

If the task follows the same steps every time, a simple automation is usually cheaper and more predictable. An agent is worth it when each case is a little different and needs judgement, such as enquiries that are often incomplete.

What determines the cost?

It depends on the software you use, how much AI usage the agent needs, hosting, the integrations with your systems, and ongoing support. A narrow agent connected to one or two systems costs far less than one that spans your whole business.

Find a useful first task for your business.

Tell us how your team handles enquiries, quotes or admin today. We will suggest one task an agent, or a simpler automation, could take on first.

Talk to us