Contents8
What is an AI agent? The word appears in the announcements of almost every software vendor, to the point of describing just about any tool that contains a language model. For a company deciding where to put AI, though, the definition has consequences: it says what the tool will do without you, and therefore what you will need to monitor. This guide, from our Business AI: private and at the right cost section, starts from the most useful definition we have read, then moves on to what an agent is not and what it costs. If you are looking for where to start with AI rather than a definition, begin with our guide to integrating AI in your business.
In short
What sets an agent apart from a workflow is who decides the next step: your code, or the model.
An agent is justified when the number of steps cannot be predicted; a repetitive task is better handled by a fixed path.
Autonomy is paid for in model calls and possible errors, and is bounded by caps and checkpoints.
Many products sold as agents are not: ask what the tool decides on its own.
What exactly is an AI agent?
This definition comes from Anthropic, the maker of the Claude models, in a December 2024 text, Building effective agents, which sets two families of systems against each other:
- workflows, “systems where LLMs and tools are orchestrated through predefined code paths”;
- agents, “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks”.
Take a quote request received by email:
As a workflow, the path is written down
Extract the name, the company and the need
Create the record in the CRM
Notify the salesperson
A model may read the email at one step, without choosing what comes next
As an agent, the model decides
It is given a goal and tools: the CRM, the catalogue, the mailbox
It chooses whether to look up the customer first, ask a question or stop
The difference therefore lies neither in the quality of the model nor in the interface, but in one question: who decides the next step, your code or the model? In n8n, the documentation for the AI Agent node sums it up this way: you connect a model and one or more tools, “and the agent decides which tools to call to complete a task”.
Agent, chatbot, workflow, RPA: what are the differences?
These words refer to tools that do not behave in the same way. The table sorts them by the previous criterion.
| Tool | What it does | Who decides what happens next | Example |
|---|---|---|---|
| Chatbot | Answers within a conversation | The user, at each message | Online help about your opening hours |
| RPA | Replicates a human’s clicks in an interface | The script, written in advance | Copying orders from a supplier portal |
| Workflow | Chains steps defined in advance | The code | Each invoice received is filed, then entered |
| Workflow with an AI step | Hands a model the reading or writing of a text, at a planned point | The code; the model fills in a slot | A customer email is summarised, then filed |
| AI agent | Chooses its steps and tools to reach a goal | The model, within the set limits | An assistant that cross-checks several sources before answering |
A chatbot becomes an agent when it can act (create a record, change an order) and choose when to do so; as long as it merely answers, it is a chatbot, whatever its model. RPA, for its part, is not autonomous at all: it is a fixed path that goes through the screen for lack of better access to the software, and often one of the last resorts, as our guide to automating software that has no API explains.
Do you always need to go all the way to an agent?
No, and the same Anthropic text says so most clearly. It recommends “finding the simplest solution possible, and only increasing complexity when needed”, which “might mean not building agentic systems at all”. For many applications, optimising a single model call, with document retrieval and examples placed in the prompt, is “usually enough”.
We break the decision into five steps. The breakdown is ours; it follows that recommendation:
Five steps, from the fixed rule to the agent
- A fixed rule, without AI
Above a certain amount, notify the manager.
- A workflow
It chains several rules and several pieces of software.
- A workflow with an AI step
The AI reads free text or writes a first draft.
- An AI that chooses between planned branches
For example by classifying a request.
- An agent
It chooses its own steps and tools.
Each step adds flexibility, but also cost and unpredictability: start with the lowest one that does the job. Our list of questions is in Automation without AI: when a rule is enough. When AI is justified, the model chosen matters as much as the architecture: our approach is to give each task to the model that is just sufficient, with a budget and caps per use.
What an AI agent is not
Not a label
In June 2025, a Reuters wire report carried by Yahoo relayed a phenomenon described by the research firm Gartner: “agent washing”, which consists of relabelling AI assistants and chatbots as “agents” without any real agentic capability. Gartner estimated that, of the thousands of vendors presenting themselves as sellers of agentic AI, only about 130 really are. This is a research firm’s estimate, not a census, but the question it suggests applies to any purchase: what does your tool decide on its own?
Not reliable by default
In 2025, Anthropic, together with the company Andon Labs, entrusted a small automated shop in its San Francisco offices, for about a month, to an agent based on Claude Sonnet 3.7, nicknamed Claudius. The account of this experiment, Project Vend, published on 27 June 2025 by the maker of the model itself, is instructive. The agent set prices “without doing any research”, so that items that could have earned a good margin ended up priced below their cost. For a time it asked customers to pay into an account it had invented. It let itself be talked into granting numerous discount codes by messages on Slack. These are not breakdowns: they are decisions, made by an agent that had control of prices and payments.
From the field
A client learned a more modest version the hard way
On their site, a model published three pieces of content that were entirely made up, which have since been removed. That is why human review before publication is a non-negotiable rule with us, and why we have tooling for an anti-hallucination check.
Not a colleague
On the morning of 1 April, the same account relates, Claudius claimed it would deliver products itself “in person”, wearing a blue blazer and a red tie. Anthropic’s conclusion is blunt:
If Anthropic were deciding today to expand into the in-office vending market, we would not hire Claudius.
An agent has no responsibility: it executes within the limits it is given, and a human answers for the result, as we explain in our article on the human role alongside AI.
When is an AI agent really useful?
Anthropic gives the clearest answer: agents suit “open-ended problems where it’s difficult or impossible to predict the required number of steps, and where you can’t hardcode a fixed path”. Workflows, for their part, “offer predictability and consistency for well-defined tasks”.
Four signs indicate that an agent may be worth what it costs:
- the path changes from one case to the next, in an order you do not know in advance;
- the input is unstructured: free text, a photo, a scanned document;
- an error can be caught, because a human validates before the action commits the company;
- the volume justifies the effort of setting up and monitoring.
Our accounting data-entry agent combines two of these signs: its input is a photo of a receipt sent on Telegram, and only one human action remains, approving with a button the record it has prepared. It has been in internal use since July 2026, among the tools we run ourselves.
What an AI agent costs, beyond the subscription
Autonomy has a price, and Anthropic’s guide says so: agentic systems “often trade latency and cost for better task performance”. An agent that hesitates and calls three tools instead of one consumes that many model calls, and so that many tokens. Its autonomous nature implies “higher costs, and the potential for compounding errors”, hence the recommendation to test it extensively “in sandboxed environments”, with suitable guardrails.
Key takeaway
Three settings keep an agent within its limits
- A maximum number of turns, which Anthropic calls “stopping conditions (such as a maximum number of iterations)”.
- A budget per task, with a clean stop when it is reached, rather than a bill discovered at the end of the month.
- Checkpoints, where the agent pauses to get human feedback, especially before an irreversible action.
How the bill is calculated (input and output tokens, context sent back at every turn) is detailed in our guide to the cost of AI in business. And if you wonder why so many AI projects are slow to pay off, our analysis of the return on AI in business shows that the technology weighs less than the organisation around it.
Where to start
Start from the task, not the tool. Describe what comes in, what must come out and who checks it today: the lowest step that does the job often becomes obvious by itself. If the answer is an agent, the remaining job is to turn it into a reliable tool for a whole team, which is the subject of our guide to the AI agent in production, from demo to product. If you are torn between several projects, our digital and AI maturity audit helps you know what to automate before you spend, with a written GO or NO-GO for each.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers within a conversation: the user decides what happens next, message after message. An AI agent receives a goal and tools, and chooses the steps to reach it itself: look up information, create a record, send a message. A chatbot that can act in your software, and decide when to do so, becomes an agent, with the controls that this requires.
Can an AI agent make up an answer?
Yes. A language model produces a plausible answer, not a verified one, and an agent chains several of these answers together: an error at the start spreads to what follows. A serious agent therefore cites its sources, answers “I don’t know” when it is not sure and has anything that commits the company validated. That is the rule of our AI assistant on your documents, which cites its sources.
Can an AI agent work unsupervised?
Not on decisions that commit you, at least not yet. In the second phase of Project Vend, published in December 2025, Anthropic gave its agent new models and better tools; over the course of that phase, the weeks with a negative margin largely disappeared. Is it ready, then, to run a vending machine in your company? Anthropic’s answer is “Not quite”: its agents still needed a great deal of human support.
Do you need an AI agent to automate a process?
Not necessarily. A process that follows rules you can write down is automated with a fixed workflow, which is cheaper and gives a consistent result. AI is justified at a specific step, when you need to read free text or settle an ambiguous case, and an agent only when the path itself changes from one case to the next. Our decision framework is in Automation without AI: when a rule is enough.
Sources
- Building effective agents (new tab)anthropic.com
- Documentation for the AI Agent node (new tab)docs.n8n.io
- Reuters wire report carried by Yahoo (new tab)tech.yahoo.com
- Account of this experiment, Project Vend (new tab)anthropic.com
- Second phase of Project Vend (new tab)anthropic.com
Jim · founder of Brain Plus AI
Also in business ai
Practical guide10 min
An AI agent in production: from demo to product
An AI agent demo impresses quickly. Turning it into a tool a team uses every day takes an owner and written rules: what four public cases show, and what we run ourselves.
Practical guide10 min
The cost of AI in business: tokens, routing and forecasting
What you really pay when you use an AI through an API: input and output tokens, the context sent back with every call, an agent's turns. And a method to forecast the bill before you receive it.
Stepping back11 min
AI return on investment: why it is taking so long
The figure of 95% of AI projects with no return went round the world. What it really measures, why the other studies say roughly the same thing, and the practical lessons to draw from them for an SME.