AI agent, chatbot or workflow automation: which one do you need?

Which one you need depends on the task. If it follows the same steps every time, you need workflow automation. If customers or colleagues ask questions in their own words and the answers already exist in your company information, you need an AI assistant. If the task also means deciding what to do next and acting in your systems, you need an AI agent. A scripted chatbot is enough only when the questions are few and predictable.

For most small businesses the order matters more than the choice. Start with a workflow built on rules, and add AI at the steps where the input is messy.

Plain definitions

The terms overlap in marketing material, so it helps to pin them down.

Scripted chatbot

A chat window that follows a decision tree. The visitor clicks buttons or types a keyword, and the bot replies with a prepared text. It cannot handle a question nobody scripted.

AI assistant

A chat or email assistant built on a language model. It understands questions in free text and answers from the information you give it: services, prices, opening hours, procedures. It can collect details and pass the conversation to a person. Many people call this a chatbot too, which is where the confusion starts.

AI agent

An assistant that can also act. It decides which step comes next and uses tools to carry it out: it looks up an order, creates a ticket, books a slot or updates the CRM. You set the limits of what it may do on its own.

Workflow automation

Software that moves data between your tools by fixed rules. Something triggers it, such as a submitted form, a new invoice or a date, and the same steps run every time. No conversation is involved, and often no AI at all.

Two questions separate the four. Who decides the next step: a script, fixed rules or the model? And does the system only answer, or does it also act?

Comparison at a glance

Aspect	Scripted chatbot	AI assistant	AI agent	Workflow automation
Input	Buttons and keywords	Free-text questions	Free text, documents, events	Structured data and triggers
Who decides the next step	The script	The model, within your information	The model, within limits you set	Fixed rules
Takes actions	Rarely	A few: takes details, hands over	Yes: books, creates, updates	Yes, the same ones every time
Predictability	High	Medium	Lower without guardrails	High
Good for	A few fixed questions	Customer questions answered from your own information	Multi-step tasks with varied input	Repetitive processes with clear rules
Main risk	Dead ends that frustrate visitors	Wrong answers if it is not limited to your information	Wrong actions	Quiet failures if nobody gets an alert

When each one fits

A scripted chatbot fits when visitors need one of five or six things and a menu gets them there faster than typing: opening hours, a price list, a link to the booking page.

An AI assistant fits when the same questions arrive every week in different words, the answers are written down somewhere, and enquiries wait overnight or over the weekend. It is also useful inside the company, for finding an answer in procedures and contracts.

An AI agent fits when answering is not enough and the request has to be completed. The customer wants to reschedule, not to read about rescheduling. An agent needs systems it can connect to and clear limits. Our page on AI agents and assistants shows what both can do.

Workflow automation fits when the work is moving information, not holding a conversation: an enquiry from the form into the CRM, a confirmation to the client, a reminder three days later, a report on the first of the month.

Why most small businesses should start with a workflow

There are four practical reasons.

1. Rules are predictable. A rule does the same thing every time, so you can test it once and rely on it. A model has to be tested on many real examples and watched afterwards.
2. Rules cost less. Our indicative prices start at €800 for a workflow and €1,300 for an AI agent or assistant, excluding VAT. A rules-only workflow also has no AI usage to pay for.
3. A lot of lost time is in moving data. Typing the same client into three systems, copying figures into a report and chasing confirmations take hours that nobody counts. Workflow automation removes them without changing how customers reach you.
4. An agent needs a workflow underneath. An agent that books a slot or updates a record relies on connected systems and clean data. Build those first and the agent has something reliable to act on.

Add AI where the input is messy

Messy input is anything a rule cannot read: an email written in free text, a scanned invoice, a question typed by a customer. That is where an AI step earns its place.

Take enquiries as an example. The ones from your website form go straight into the workflow, because the fields are already structured. The ones that arrive by email pass through an AI step that works out what the person wants and extracts their details, and then they join the same workflow. For PDFs and scans, that AI step is document processing.

If you are not sure which of your processes qualify, the free checklist of 50 processes to automate is a quick way to build a shortlist.

Guardrails for customer-facing assistants

An assistant that talks to your customers speaks for your company. Three guardrails should be in place from the first day.

• Limit it to company information. The assistant answers from the content you approve and stays on the topics you agree. When the answer is not there, it says so and does not improvise.
• Hand over to a person. For complaints, sensitive topics or anything it does not know, it collects the details and passes the conversation to a colleague with the context attached. Customers should always have a way to reach a human.
• Log conversations. Keep a record you can review. It shows you which answers to correct and tells you what customers actually ask.

Two more are worth adding. Decide exactly which actions an agent may take alone and which need approval. And test on real questions from your inbox before launch, not on the ones you imagine customers ask.

Decision checklist

Go through these in order and stop at the first one that describes your task.

• Nobody can describe the steps or the correct answers yet. Write them down first. Nothing below will work until you do.
• The task follows the same steps every time, with structured input. Use workflow automation.
• The steps are fixed, but one of them involves free text, a document or a photo. Keep the workflow and add an AI step there.
• People ask questions in their own words, and the answers are written down. Use an AI assistant limited to that information.
• The request must be completed, not only answered, in a system you can connect to. Use an AI agent with defined limits and a hand-over to a person.
• There are only a handful of fixed questions. A scripted chatbot or a good FAQ page will do.

Next step

If you are unsure where your process falls, book a free 30-minute automation audit. We map one process with you and tell you what can be automated, with rules, AI or both, and roughly what it costs.