In short
- Automation applies a fixed rule. A chatbot answers a question. An AI agent reads a case, chooses what to do and prepares an action in your tools.
- With us, an agent never pays, does not commit the company without you, and hands over to you in the cases set out in advance.
- If your task always follows the same rule, automation is enough. If it means reading and judging each case, an agent can help.
The word “agent” is everywhere. Self-service automation tools use it too: Zapier presents its Agents as “AI teammates”, and Make offers AI Agents. The need for clear definitions is real. According to Eurostat, in 2025, among EU businesses with 10 or more people that did not use AI but had already considered it, 70.89% cited a lack of expertise. It was the first reason given.
Here are some simple definitions, by example.
Three tools, three ways of working: the fixed rule, the conversation, the prepared action
Automation applies a fixed rule. “When this happens, do that.” Zapier puts it simply: a trigger is the event that starts the automation, and an action is what it does next (Zapier help). Example: each new order creates a row in a spreadsheet. It is predictable: as long as the case follows the rule, the result is the same. But automation understands nothing. If the case falls outside the rule, it gets it wrong or it stops without knowing why.
A chatbot answers a question. It is a conversation window, on a website or in a messaging app. Example: a customer asks for your opening hours, and the chatbot answers. It talks. Most of the time, it does not act in your tools.
An AI agent chooses its approach and prepares an action. It reads a document with no fixed layout: an email, a PDF invoice, a quote request. It analyses the case and chooses what to do. Then it prepares something in your software: a draft reply, a customer record, an accounting entry. Properly supervised, it submits this to you before acting, at least until it has proved itself.
Anthropic, the publisher of Claude (the AI model our agents will use by default), makes a similar distinction. In what it calls a “workflow” (a sequence of steps fixed in advance), the model and the tools follow “predefined code paths”. In an agent, the model “dynamically directs its own processes and tool usage” (Anthropic, 19/12/2024).
| What it does | Example | Its limit | |
|---|---|---|---|
| Automation | applies a rule | filing each order in a spreadsheet | cannot handle the unexpected |
| Chatbot | answers in a conversation | giving your opening hours | talks, but rarely acts in your tools |
| AI agent | reads, chooses, prepares or acts | preparing the reminder for an unpaid invoice | can make mistakes, so it must be supervised |
The boundaries are shifting: at Make, for example, agents can launch automations. Remember the question that matters: who prepares, who decides, who carries out?
What an AI agent actually does in an SME: payment reminders, emails, quotes, accounting documents
We are preparing four standard agents, which will be adapted to each business.
- Payment Reminders and Invoices. It tracks due dates and prepares reminders in the tone you choose. It suggests matching each payment received to its invoice, that is, indicating which invoice the payment settles. Each week, it lists the customers at risk.
- Emails and Support. It sorts incoming messages (enquiry, complaint, invoice, prospect). It creates the ticket, that is, the tracking record for the request. It suggests a reply drawn from your own documents and passes everything to the right person.
- Quotes and Prospects. It qualifies each enquiry (need, budget, deadline, decision-maker) and completes the record in your CRM (the software where you track your customers). It prepares a draft quote from your catalogue and your pricing rules. It also prepares reminders for quotes left unanswered.
- Accounting Documents. It collects supplier invoices and extracts the amounts and the VAT. It suggests the account in which to file each expense (the coding) and the bank reconciliation, that is, the link between each movement on the account and its supporting document. At month end, it prepares a clean batch for your accountant.
Take a typical scenario, for example an SME of 30 employees that receives its quote requests by email and through a form. An assistant copies them into the CRM before preparing each quote. With an agent, she finds the record already filled in and a draft quote ready. She reads it, corrects it if needed, and she is the one who sends it. If a request falls outside the catalogue, the agent does not set a price: it flags the request to her.
What it does not do: decide for you, pay, commit the company
A well-designed agent has written limits. With us, they appear in each agent's rulebook, which you approve before any development.
- It does not decide for you. Sending a quote, granting a discount, promising a lead time, issuing an invoice or a credit note: a human always approves.
- It never pays. It can suggest a reconciliation, never make a transfer.
- It does not commit the company. It does not send on its own a document that binds you to a third party. It deletes no data without your agreement.
- It hands over in the cases set out in advance. A customer in dispute, an unreadable document, a legal threat, a person who wants to exercise their rights over their personal data, a language it does not cover: it stops and hands over to you rather than guess.
- It sees only what it needs. Each connector (the link between the agent and one of your software tools) gives minimal access, read-only whenever possible.
Why a human stays in control: autonomy levels 0 to 3 and stopping in the cases set out in advance
An agent can make mistakes. Anthropic says as much: agent autonomy means “higher costs” and the risk of “compounding errors” (source). Hence the safeguards.
The EU AI Act defines an AI system as a system “designed to operate with varying levels of autonomy” (Regulation (EU) 2024/1689, Article 3). Autonomy is not an all-or-nothing switch. With us, it is set agent by agent and action by action, in four levels. These levels are our method, not a requirement of the Regulation.
| Level | The agent… | You… |
|---|---|---|
| 0: shadow | proposes, does not act | do the work; its proposals are compared with yours |
| 1: systematic approval | prepares each action | approve, correct or refuse each action |
| 2: approval by exception | acts alone on safe cases, defined in writing | approve the cases outside the rules and check a sample |
| 3: autonomous, supervised | acts alone | read the report and receive the alerts |
Three rules govern these levels:
- An agent moves up a level only on the basis of measurements. We look at its agreement rate, that is, the share of its proposals that your team accepts, and at its errors. We also need your written agreement.
- It will step back down on its own. After a serious alert, it will automatically return to the level below.
- Some actions never go beyond level 1. Sending a binding document or deleting data always goes through you.
Before going live, the agent works for a few weeks in shadow, for a period defined with you. Each action will be recorded in a log: what it read, what it proposed, what the human decided, who, when. In the event of a malfunction or a repeated error, the agent will report it to us itself: this automatic reporting is being prepared.
Does your need call for simple automation or an agent? Five questions to decide
- Does the task always follow the same rule? If so, automation is often enough.
- Do you need to read free text to know what to do? An email, a scanned invoice, a request written by a customer: this is an agent's territory.
- Do cases vary so much that a rule misses some? If you often correct your automation by hand, an agent can take over.
- Would a mistake be costly or commit the company? Then you need human approval, whatever the tool. With us, an agent includes it from the start.
- Does the task come up often enough? If it happens only a few times a month, keep it manual.
Anthropic gives developers the same advice: look for “the simplest solution possible”, even if that means not building an agent at all (source). If automation is enough and someone on your side knows how to set it up, a self-service tool may do the job. We will tell you.
Going further: method, deployment, price
- Our agents: the four standard agents in preparation, on quote.
- Our method: from the audit to going live, and the two moments when you decide.
- Supervision: what it will include, with a dashboard, alerts, a monthly report, and agents that will report their own problems.
- Data and security: where your data is processed, and the choice of deployment: your cloud account, your servers or a dedicated server with us.
- The audit: we look at your tools and your tasks, and we tell you whether an agent is worth it or whether a simpler solution is enough. It costs €490 excl. VAT, deducted from the quote if you sign it within 30 days.