AI agent, chatbot or automation: which does your process need?

Use automation when the inputs are predictable and the steps never change. Use a chatbot when the job is answering people who ask questions. Use an AI agent when the inputs are messy, the next step depends on what the system reads, and the work has to be carried out across several tools.

Most real processes are not one of the three. They are a mix: fixed rules for the predictable parts, an agent for the parts that need reading and judgement, and a person who approves anything that moves money, carries a signature or reaches a customer. The useful question is not "which technology?" but "which part of this process needs which one?"

What is the difference between an AI agent, a chatbot and automation?

The cleanest line is who decides the next step. Anthropic, which builds the Claude models, puts it this way in its guide Building effective agents: workflows are systems where models and tools are "orchestrated through predefined code paths", while agents are systems where models "dynamically direct their own processes and tool usage".

In practice the three differ like this:

	Automation	Chatbot	AI agent
Who decides the next step	The rule someone wrote	The person asking	The model, within limits you set
What it is good at	Same steps, same format, high volume	Answering questions from a known body of information	Reading unstructured input and acting on it across tools
What it touches	Systems, through fixed connections	A conversation	Inboxes, files, task boards, other systems
How it fails	Breaks loudly when the input changes	Gives a confident wrong answer	Takes a plausible wrong action
Running cost	Low and predictable	Moderate	Higher, and variable unless capped

A chatbot talks. Automation does. An agent reads, decides and does, which is why it needs the most care.

When is plain automation enough?

More often than vendors suggest. If a form always arrives with the same fields, if a report always pulls the same numbers, or if a file always goes to the same folder, a rule will do the job faster, cheaper and more reliably than a language model.

Anthropic's own advice is to find "the simplest solution possible" and add complexity only when it is needed, noting that agentic systems "often trade latency and cost for better task performance". Gartner makes a sharper point: in a June 2025 forecast it predicted that over 40 percent of agentic AI projects will be cancelled by the end of 2027, and said many use cases presented as agentic today do not need an agent at all.

The test is simple. If you can write the process down as a flowchart with no box that says "it depends", start with automation.

When do you actually need a chatbot?

When people come to you with questions and the answers already exist somewhere: opening hours, programme details, policy questions, the status of an application. A good chatbot answers from your own material and hands over to a person when it cannot.

What a chatbot does not do is the work behind the question. If a visitor asks to change a booking and the chatbot replies "please email us", the process has not moved. Many organizations buy a chatbot hoping to save staff time and find the inbox is just as full, because the requests still need someone to act on them.

One rule now applies to every chatbot facing the public in the EU. Since 2 August 2026, Article 50 of the EU AI Act requires that people are told when they are interacting with an AI system, unless it is obvious. If you run a chatbot on your website, say so on screen.

When is an AI agent worth it?

When three things are true at once:

• The input is unstructured. Emails, PDFs, scanned invoices, contracts, messages written in five different ways.
• The path is not fixed. What happens next depends on what the input says. Anthropic describes agents as suited to "open-ended problems where it's difficult or impossible to predict the required number of steps".
• The work spans several tools. Read the email, find the contract in the drive, check the amount, prepare the payment, update the task board.

That is the shape of most back office work, which is why it is where we start. In our own group of companies, agents sort the mail, prepare reports, draft paperwork and line up payments, and a person approves anything that moves money or carries a signature. You can read how that runs in the back office case.

The cost of that flexibility is risk. Anthropic warns that autonomy means "higher costs, and the potential for compounding errors". That is why every agent we deploy works inside a fixed budget, logs every action and stops for human approval at the points that matter.

How do you decide? Four questions

Ask these about each step of the process, not about the process as a whole.

1. Does the input always look the same? If yes, automation. If it arrives as free text, attachments or scans, you need a model to read it.
2. Is the next step always the same? If yes, automation, even if a model does the reading. If it depends on what was read, that step is agent territory.
3. Is someone asking, or does something need doing? Questions from people point to a chatbot. Work that needs doing points to automation or an agent.
4. What happens when it is wrong? If a mistake is cheap and easy to undo, let the system act. If it costs money, reputation or a legal commitment, the system prepares and a person approves.

Run a typical invoice process through those questions and you get a hybrid. A model reads the incoming invoice and pulls out the supplier, amount and due date. A rule checks it against the supplier list and the budget. An agent handles the exceptions: the invoice that does not match a contract, the supplier asking about a late payment. A person releases the payment.

What breaks when you pick the wrong one?

Each mismatch has a recognisable symptom.

• An agent where automation would do. Costs rise, results vary from run to run, and nobody can explain why the same input produced two different outcomes. The fix is to take the predictable steps out of the agent and turn them into rules.
• A chatbot where an agent was needed. Customers get polite answers and nothing happens. Staff time is not saved, it is just moved.
• Automation where an agent was needed. The rules work on the clean cases and every exception lands in someone's inbox. Over time the exception queue becomes the real process.

Where to start

Take one process with real volume and a clear owner. Write down each step and run it through the four questions above. Automate what is fixed, put an agent on what needs reading and judgement, and decide in advance where a person approves. Then measure what gets corrected in the first weeks. That is how we scope work in agentic implementation.

If you have a process in mind and are not sure which of the three it needs, we are happy to look at it with you.