AI agents: what they are, how they work and examples for businesses

An AI agent is a program that is given a goal and works out for itself which steps to take to reach it, using external tools: a browser, a spreadsheet, the API of a CRM or accounts package. A chatbot answers a question. An agent keeps doing things until the task is finished.

Google Cloud defines them as "software systems that use AI to pursue goals and complete tasks on behalf of users". IBM's definition stresses one point: the agent designs its own workflow with the tools it has available.

Agents, chatbots and automation: the differences

The three terms get used as if they meant the same thing. They don't.

	Chatbot	Automation (workflow)	AI agent
Who decides the steps	Nobody, it just answers	Whoever programmed it	The model, case by case
Uses external tools	Usually not	Yes, always the same ones	Yes, and it chooses which
Copes with the unexpected	No	No, it stops or gets it wrong	Partly, it can try again
Example	A chat assistant on a website	Zapier copying leads from a form into the CRM	A system that finds 20 target companies, reads their websites and drafts the emails

The clearest distinction was written by Anthropic in December 2024. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths". Agents are "systems where LLMs dynamically direct their own processes and tool usage". The same piece has the most useful advice for a business: look for the simplest solution that works and add complexity only when you need it. Plenty of problems can be solved with a workflow, no agent required.

How an AI agent works

Google Cloud lists four components. In practice:

1. The model, the "brain": GPT, Claude, Gemini. It reasons and picks the next step.
2. The tools: web search, reading files, access to a CRM, sending email. Without tools an agent is just a chatbot.
3. Memory: short term (what it did in the previous steps) and long term (preferences, rules, history).
4. Instructions: its role, its limits, the rules it has to follow.

The loop never changes: the agent looks at the situation, chooses an action, carries it out with a tool, reads the result and decides whether it's done or needs another step. A simple task takes three or four steps. A complex one can take hundreds.

To connect the tools, the Model Context Protocol (MCP) has become the standard since 2025. It's an open standard created by Anthropic and now run by the Linux Foundation. We explain it in our article on WebMCP and MCP.

The AI agents available in 2026

The market changes every quarter. This is where things stood at the end of September 2026:

• ChatGPT Work from OpenAI, launched on 9 July 2026. It works across apps and files for hours on the same project and has a browser built into the desktop app. It replaced ChatGPT agent, which is no longer available.
• Claude in Chrome from Anthropic, available on all paid plans since 26 August 2026. It reads pages, clicks, fills in forms and works across several tabs. A safety classifier checks every action.
• Chrome auto browse with Gemini, in preview since 28 January 2026 for AI Pro and Ultra subscribers in the US only. It asks for confirmation before purchases and posts. Gemini in Chrome reached UK desktop users on 14 July 2026, but as of late July Google was still rolling out the agentic part in the US first.
• Perplexity Comet, a browser with a built-in agent, free for everyone since October 2025 and on iOS since March 2026.
• Microsoft Copilot Studio, with "computer-using agents" generally available since May 2026. Agentic browsing in Edge for Business is available worldwide with a Microsoft 365 Copilot licence, excluding the European Economic Area. The UK isn't in the EEA, so it isn't on the exclusion list.

To build your own agents there are platforms such as n8n and Zapier, or the frameworks published by the model providers themselves. Here the choice depends more on who will maintain the agent than on the technology.

Examples of AI agents in marketing and SEO

Three that we use at Visilay, with the limits we've run into.

Editorial agent. It runs three times a day on our blog. It picks a topic from the editorial plan, checks the topic isn't already covered, writes the article and publishes it. In August 2026 our Italian blog published 56 articles, in September 168, plus the English translations. The full story, with the numbers and the mistakes, is in AI agents for SEO.

Prospect research agent. It's given a sector and an area, finds the companies, reads their websites, checks whether they advertise on Google and prepares a profile for each one. A day's work becomes an hour. We still write the emails ourselves, though: the ones the agent wrote were correct and all identical.

Monitoring agent. It puts a list of questions to ChatGPT, Gemini, Claude and Perplexity and records which brands appear in the answers. It's the basis of measuring AI visibility.

Other common examples: agents that handle first-line support tickets, that reconcile invoices against orders, that prepare weekly Google Ads reports by reading the data straight from the account.

Where an agent pays off for a business

The rule we use: an agent is worth it when the task is repetitive, the rules can be written down, the result is easy to check and a mistake is cheap. If any one of those four is missing, a conventional workflow or a person is the better choice.

In the UK, AI adoption is wide but shallow. According to the ONS (Business Insights and Conditions Survey, June 2026, published 20 July 2026), around 35% of businesses with 10 or more employees used at least one AI technology, up from around 12% in late 2023. The figure is 28% for businesses with 0 to 9 employees and 49% for those with 250 or more. Only 10% of adopters say they use AI extensively. Insufficient expertise has delayed adoption for around 18% of businesses with 100 to 249 employees, and cost affects between 7% and 14% depending on size. The ONS doesn't yet measure agents: it lists "agentic workflows" among the things it plans to track. On the supply side, the government's AI sector study (published 3 September 2025) counted 5,862 AI companies in the UK with around £23.9 billion of AI-related revenue in 2024, up 68% on the year before.

The gap between large and small businesses shows up with agents too. In the McKinsey State of AI 2026 report, 40% of large companies are scaling agentic AI; among smaller organisations the figure is stuck at 22%.

What they cost

Three items:

• subscription or licence: from $20 to $200 a month for consumer agents (US list prices: Anthropic, for one, publishes its plans only in dollars), and a great deal more for enterprise platforms;
• model usage: if the agent is built on an API you pay per token, billed in dollars. A simple task costs a few cents, a long one that reads dozens of pages can cost a few dollars;
• the time of whoever sets it up and checks it, which at the start is the biggest cost.

For an order of magnitude: one Claude Sonnet query with web search, run through the API for one of our tests on 27 September 2026, cost about 7 US cents. The same question cost about 4 cents on Gemini 3.5 Flash and less than a cent on Perplexity Sonar.

What can go wrong

Gartner has predicted that over 40% of agentic AI projects will be cancelled by the end of 2027. Anushree Verma, Senior Director Analyst, describes them as "early-stage experiments driven by hype and often misapplied". The same analysis says only about 130 of the thousands of vendors offer real agents.

The limits we run into most often:

• confident mistakes: the agent gets something wrong and carries on as if nothing had happened;
• unpredictable costs: a task stuck in a loop keeps using tokens until someone stops it;
• security: an agent with browser access can be manipulated by instructions hidden in a web page;
• personal data: if the agent reads emails and documents, you need to know where that data ends up.

The safeguard is always a human check at the points where a mistake is expensive, plus a spending cap.

Further reading: Agentic SEO: what it is and what changes for businesses.

FAQs