What are AI agents, and how can businesses use them safely?

You've probably seen "AI agents" mentioned basically everywhere: on software websites, in sales pitches and in the news. The term is used for very different things, though, so it isn't always clear what you're being offered.
In this guide we explain what an AI agent is, what it can do and the main types you'll come across, and we share what to check before you use one in your business.
What is an AI agent?
An AI agent is an AI tool that can carry out a task for you from start to finish. It works out the steps, uses your other systems (such as email, files or your CRM) to do them, and checks how it's getting on, often with little help from a person.
The big difference from a chatbot is that an agent can actually do things. If you ask a chatbot about overdue invoices, it will tell you how to chase them. An agent can find the overdue invoices, write the reminder emails and send them for you.
AI agents vs chatbots vs AI assistants
Chatbot | AI assistant | AI agent | |
What it does | Answers questions | Helps with a task you're doing | Works towards a goal |
Who decides the steps | You | You, one step at a time | The agent |
Can it take action? | Rarely | Sometimes, when you ask | Yes, often across several systems |
Example | A website FAQ bot | Drafting a reply in your inbox | Handling a refund request from start to finish |
Real products don't always fit neatly into these boxes, and some products sold as agents are really chatbots with a new label. Gartner calls this "agent washing". In 2025 it estimated that only about 130 of the thousands of vendors claiming to sell agentic AI were offering the real thing. So it's worth asking exactly what a product does before you pay for it.
How AI agents work

Most agents work in roughly the same way, whatever job they're doing.
You give it a task, for example "set up a meeting with Laura from veritas_fox next week".
It works out the steps. It needs to check your calendar, check Laura's availability, pick a time and send an invite.
It does the work in your systems. It uses whatever it's been given access to, such as your calendar, email, CRM or a web browser. OWASP describes this as an AI being able to call functions and work with other systems.
It checks the result. If Laura is busy all week, it tries another option or asks you what to do.
It finishes, or hands back to you. It either completes the task or passes it to a person when it gets stuck or reaches a decision it isn't allowed to make.
Step 3 is where you need to be careful. The systems you connect an agent to decide what it can do for you, and also what it could get wrong.
Everyday business tasks
Common uses include:
Admin and finance: matching invoices to purchase orders, flagging ones that don't match, and chasing late payments.
Sales: researching a new lead, updating the CRM after a call, and drafting a follow-up email for someone to check.
Scheduling: finding a time that works for several people and sending the invites.
Research: gathering information from a set of sources and pulling it into a short summary.
Different types of AI agents
There are two useful ways to sort agents: by the job they do, and by how much freedom they have.
By the job they do
Customer-facing agents talk to your customers, answer questions and handle requests such as refunds or bookings.
Workflow agents work in the background on internal processes, such as invoices, onboarding or reporting.
Coding agents write, test and fix software.
Research agents search for information and pull it together.
Browser or computer agents use websites and apps the way a person would, clicking, typing and filling in forms.
Multi-agent systems are several agents working together, each handling part of a task and passing work between them.
By how much freedom they have
This matters more for your risk than the job they do.
Level | What it means | Example |
Suggests | The agent proposes, a person does | Drafts an email for you to send |
Supervised | The agent acts, a person approves the important steps | Prepares a refund, a person approves it |
Autonomous | The agent acts without checks | Issues refunds on its own |
The same agent can often be set up at any of these levels. Starting at "suggests" or "supervised" and only moving up once you trust it is usually the safer route.
What to pay attention to before using an AI agent
Check it's really an agent. Ask the supplier what the product actually does on its own, and what still needs a person. If it mostly answers questions, it may be a chatbot with a new name.
Know what it can reach. List every system, account and type of data the agent is connected to, and cut anything it doesn't need. OWASP warns about agents with too much functionality, too many permissions and too much autonomy. We explain this in more detail in our guide to AI blast radius.
Decide what needs a person's approval. Payments, deleting records, emailing customers and changing live systems are good places to start.
Expect mistakes. Agents can get things wrong and sound sure of themselves. The Replit agent told its user the deleted data couldn't be recovered, which turned out to be wrong.
Plan for it being tricked. Agents read emails, web pages and documents, and hidden instructions in that content can steer them. The UK's National Cyber Security Centre has warned that this kind of attack, called prompt injection, may never be fully fixed.
Keep a route to a person. For customer-facing agents, make it easy for people to reach a human, as Klarna found out.
Be clear on the value. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear business value or weak risk controls. Agree what success looks like before you start.
Check the data protection side. If the agent handles personal data, your UK GDPR duties apply as normal. Talk to whoever looks after data protection before you switch it on. If the supplier processes that data for you, you also need a data processing contract with them.
Name an owner. Someone should be responsible for each agent: what it's allowed to do, when its access is reviewed, and how to switch it off. The owner is also in charge of the risk registery.
Keep a record of what it does. Make sure the agent’s actions are logged, so you can see what it did and why if something goes wrong or a customer complains.
You can use our tool, Vulpes, to track all of this.

Where UK and European law come in
There's no separate law for AI agents. What applies depends on what the agent does, who it talks to and where it's used.
EU AI Act: telling people they're talking to AI. Since 2 August 2026, AI systems that interact directly with people, which the European Commission says includes chatbots, AI agents and avatars, must be designed so that people know they're dealing with AI, unless it's obvious. The Commission's guidelines go further for agents: an agent should also make clear who it is acting for. The guidelines aren’t legally binding, but national regulators are expected to use them when they check compliance.. This duty sits with the provider. If you build or commission an agent and put it into service under your own name, that could be you. Fines can reach €15 million or 3% of worldwide turnover, whichever is higher. For SMEs and start-ups, the lower of the two applies.
EU AI Act: high-risk uses. An agent used for something like screening job applicants or assessing creditworthiness can fall into the high-risk category, with much stricter rules. Following the Digital Omnibus on AI, those obligations apply from 2 December 2027. AI built into products covered by EU product safety law, such as machinery or medical devices, follows from 2 August 2028.
UK businesses aren't automatically outside the Act. It can apply if you sell AI into the EU or if your AI's output is used in the EU.
EU AI Act: AI literacy. If the Act applies to you, you must take measures to support the AI literacy of the staff who use or oversee your agents. This has applied since 2 February 2025, whatever the risk level of the agent. Since the Digital Omnibus, you don’t have to guarantee a set level of knowledge, but you do need to take real steps, such as training people on the agents they actually use.
UK GDPR. If an agent handles personal data, the usual rules apply, including keeping that data secure and reporting breaches to the ICO within 72 hours where they're likely to put people's rights and freedoms at risk. If an agent makes significant decisions about people with no meaningful human involvement, such as turning down a loan or a job application, the automated decision-making rules in UK GDPR Articles 22A to 22D apply. Since 5 February 2026 these allow such decisions in more cases, but you must tell people about the decision, let them make representations and challenge it, and let them ask for a person to review it. Stricter limits apply where sensitive data, such as health information, is involved.
ISO/IEC 42001. The international standard for AI management systems gives you a structure for deciding which agents you use, who owns them and how you manage their risks.
Frequently asked questions
What is an AI agent in simple terms? An AI system that can work towards a goal by itself: it plans the steps, uses tools like email or other software to carry them out, and checks its progress.
What's the difference between an AI agent and a chatbot? A chatbot answers questions. An agent can take action, such as updating records, sending emails or booking meetings, often across several systems.
Are AI agents safe for small businesses to use? They can be, if you limit what they can access, keep a person approving important actions, and name someone responsible for each one. The risk grows with how much an agent can reach and do without checks.
Do I have to tell customers they're talking to an AI agent? Under the EU AI Act, from 2 August 2026, AI systems that interact directly with people must be designed so people know they're dealing with AI, unless it's obvious. The Commission's guidelines also say an agent should make clear who it's acting for. If you serve customers in the EU, assume you need to make both clear.
What is agent washing? Gartner's term for vendors rebranding existing products, such as chatbots or automation tools, as AI agents without adding real agent capabilities.
Where should I start? Pick one routine, low-risk task, give the agent the least access that works, keep a person approving its actions, and measure whether it's actually saving time before expanding.
Thinking about using AI agents?
If you're already using an agent, or a supplier is trying to sell you one, it helps to know what it will be connected to, who's responsible for it, and which rules apply before it goes live.
In our free consultation, we'll talk through the agents you're using or considering, help you work out where the risks are, and give you the questions to ask your supplier. You'll come away knowing what to put in place first.
This article is general information, not legal advice.
Sources
Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner, 25 June 2025
LLM06:2025 Excessive Agency, OWASP GenAI Security Project
Prompt injection is not SQL injection (it may be worse), National Cyber Security Centre, December 2025
Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act, European Commission, C(2026) 5054 final, 20 July 2026
EU AI Act, Article 2: Scope, European Commission AI Act Service Desk
EU AI Act, Article 99: Penalties, European Commission AI Act Service Desk
Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal of the European Union, 24 July 2026
Personal data breaches: a guide, Information Commissioner's Office
ISO/IEC 42001 explained, ISO



Comments