An AI agent is a system built on a language model that works towards a goal by deciding its own next steps and using tools, such as searching your files, updating a CRM record or sending an email. A chatbot answers. An agent acts.
The idea is real and useful, but the label is stretched. Plenty of products sold as agents are chatbots or fixed automations with a new name, and many genuine agent projects stall because the job didn’t need one. This guide explains what the term means, where agents earn their keep in a small firm, and how to try one without handing a piece of software the keys to the business.
- OpenAI defines agents as “systems that independently accomplish tasks on your behalf”. An agent combines a model, tools and instructions.
- Anthropic separates agents, where the model directs its own steps, from workflows, where steps are fixed in code. It advises using the simplest option that works.
- Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing cost, unclear value and weak risk controls.
- Keep a person in the loop for anything sensitive or irreversible, such as refunds, payments or decisions about people.
What makes something an AI agent?
The two companies that publish the most practical guidance on this, OpenAI and Anthropic, describe agents in similar terms.
OpenAI’s practical guide to building agents (updated January 2026) says an agent has three parts: a model that reasons and decides, tools that let it gather information and take actions in other systems, and instructions that define how it should behave. The model runs in a loop: look at the goal, choose a tool, read the result, decide what to do next, and stop when the job is done or it needs help.
Anthropic’s engineering guide, “Building effective agents” (December 2024), draws a useful line. Workflows are “systems where LLMs and tools are orchestrated through predefined code paths”. Agents are systems where models “dynamically direct their own processes and tool usage”. In a workflow, a developer decides the order of steps. In an agent, the model does.
Most of what works well in small businesses today sits somewhere between the two: a fixed process with one or two points where a model makes a judgement call.
Chatbots, workflows and agents compared
| Chatbot or assistant | AI workflow | AI agent | |
|---|---|---|---|
| What it does | Answers questions in a conversation | Runs fixed steps, with AI at some of them | Plans its own steps and uses tools to reach a goal |
| Example | Website chat that answers delivery questions | Every new enquiry is summarised, tagged and added to the CRM | “Chase all invoices over 30 days overdue, checking each customer’s history first” |
| Can change your systems | Usually not | Yes, in the ways you designed | Yes, within the permissions you grant |
| Predictability | Medium | High | Lower |
| Running cost | Low | Low to medium | Higher: many model calls per task |
| Best fit if | People ask the same questions | The process is the same every time | Each case is different and needs some judgement |
If a chatbot is all you need, our guide on whether your business should use an AI chatbot covers the trade-offs.
When an agent is the right tool
Anthropic’s advice is to find “the simplest solution possible, and only increasing complexity when needed”. It suggests agents for “open-ended problems where it’s difficult or impossible to predict the required number of steps”. OpenAI’s guide gives three signs a job suits an agent: it needs complex or context-sensitive judgement, it is governed by rules that have become costly to maintain, or it involves a lot of unstructured information such as emails and documents. It also warns that a simpler, fixed solution may be enough.
In practice, for a firm of five to fifty people, good candidates include:
- Triage. Reading incoming emails or support tickets, working out what each one needs, pulling up the customer’s record and drafting a reply or routing it to the right person.
- Research and preparation. Gathering what you know about a prospect from your CRM, past emails and their website before a sales call.
- Back-office chasing. Checking overdue invoices, looking at each account’s history and drafting appropriately worded reminders for approval.
- Answering the phone. Voice agents that take details, answer common questions and book appointments. We cover their limits in AI phone answering.
- Software development. Coding agents that read a codebase, make changes and run tests are one of the most mature uses, mainly for developers.
Poor candidates are jobs where a mistake is costly and hard to spot, jobs that are already a clean, repeatable process (use a fixed automation), and anything involving decisions about people, such as hiring or credit, where UK data protection law adds extra duties.
Why so many agent projects fail
Gartner’s June 2025 forecast that over 40% of agentic AI projects will be cancelled by the end of 2027 is a prediction, not a measurement, but its reasons match what we see: escalating costs, unclear business value and inadequate risk controls. Gartner also described “agent washing”, vendors rebranding chatbots, assistants and robotic process automation as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI were the real thing.
Gartner’s January 2025 webinar poll of 3,412 attendees gives a sense of how cautious most organisations were at the time. Webinar attendees are a self-selected and mostly large-company audience, so read it as a mood check, not a survey of UK small firms.
- Significant investment19%
- Conservative investment42%
- No investment8%
- Waiting or unsure31%
Self-selected webinar audience, not a representative sample.
Source: Gartner webinar poll of 3,412 attendees, January 2025, reported by SD Times
View data
| Organisations' investment in agentic AI, January 2025 | |
|---|---|
| Significant investment | 19% |
| Conservative investment | 42% |
| No investment | 8% |
| Waiting or unsure | 31% |
Anthropic names the underlying technical problem plainly: “The autonomous nature of agents means higher costs, and the potential for compounding errors.” If each step of a ten-step task is usually right, the chance that all ten are right is noticeably lower. Every step is also a model call, so costs rise with the number of steps. Our explainer on what AI costs to run shows how token bills add up.
For UK firms there is a quieter barrier too. The government’s AI adoption survey 2026 found that agentic AI was less widely trusted than other kinds and more likely to face scrutiny inside businesses, and that limited skills affected 67% of businesses already using AI. Most adopters (81%) said they apply some human oversight to AI outputs, which is the right instinct for agents.
Agents, your data and the law
An agent that reads customer emails or updates records is processing personal data, so UK GDPR applies in full. The Information Commissioner’s Office (ICO) is paying close attention. In October 2026 it reported on its work with major AI developers and opened a call for evidence on how organisations manage data protection risks from agentic AI, covering security, transparency, accountability, automated decision-making, fairness and lawful use of data. It closed on 20 November 2026 and will feed into a forthcoming statutory code of practice on AI and automated decision-making.
Before you switch an agent on, work through our AI agents and UK GDPR checklist. If you sell into the EU, the EU AI Act’s transparency rules may also apply; we explain how in UK AI regulation and the EU AI Act.
How to pilot an agent safely
OpenAI’s guide recommends layered guardrails rather than relying on one, and planning for a person to take over in two situations: when the agent keeps failing (for example, repeatedly misunderstanding what a customer wants) and when it attempts something sensitive, irreversible or high-stakes, such as cancelling orders, authorising large refunds or making payments. Anthropic recommends extensive testing in sandboxed environments with appropriate guardrails. Translated for a small business:
Step 1 of 5
Choose a job
Pick one frequent task with clear success criteria and low cost if it goes wrong.
Step 2 of 5
Try simpler
Check whether a fixed workflow or a plain assistant would do the job first.
Step 3 of 5
Limit access
Give the agent its own account with the fewest permissions and read-only access where possible.
Step 4 of 5
Human sign-off
Require a person to approve anything that sends, pays, deletes or affects a customer.
Step 5 of 5
Measure
Track time saved, error rate and cost per task for a month, then decide.
Agents reach your systems through connectors, and the most common standard for those is the Model Context Protocol. Our explainer on what MCP is covers how to connect tools with sensible limits. And put the rules in writing: which agents are approved, what they may touch and who is accountable. Our guide to writing an AI policy for a small team has a template structure.
Common questions
Is ChatGPT an AI agent?
The chat itself is an assistant. Most major AI products now include agent-style features that browse the web, use connected apps or carry out multi-step tasks, so the same product can behave as either depending on what you switch on and what access you give it.
Will an AI agent replace a member of staff?
Rarely, on current evidence. ONS figures for September 2026 show only 6% of UK businesses with 10 or more staff that use AI said they had reduced headcount as a result. Agents are better thought of as taking over parts of jobs, with a person checking the work.
Should we build our own agent or buy one?
Buy when a product already does the job inside software you use. Build when the task depends on your own systems or processes that off-the-shelf tools can’t reach. Our guide to building or buying software sets out the costs on each side.
How do I spot agent washing?
Ask the vendor what the system can do without a person telling it each step, which tools it can use, what it does when it is unsure, and how you can see a log of its actions. Vague answers usually mean a chatbot with a new label.
Where to start
List the three most repetitive tasks in your business that involve reading information and making a small judgement. Try the simplest fix for each first. If one still needs an agent, our AI agents team can scope a pilot with proper limits, logs and a person in the loop.
Sources
- OpenAI: A practical guide to building agents (15 January 2026, checked 22 November 2026)
- Anthropic: Building effective agents (19 December 2024)
- SD Times: Gartner, more than 40% of agentic AI projects will be canceled in the next few years (June 2025), reporting Gartner’s press release of 25 June 2025
- ICO: ICO secures changes from leading AI developers as scrutiny extends to AI agents (8 October 2026)
- GOV.UK: AI adoption survey report 2026, executive summary (29 October 2026)
- ONS: Business insights and impact on the UK economy (8 October 2026)