Respond or execute: the table that settles it.
The confusion comes from marketing: anything containing a language model gets sold as an "AI agent." The distinguishing criterion is actually simple and binary: can the tool act in your systems without a human prompting it? If not, it's a chatbot. If yes, it's an agent.
| Chatbot | AI Agent | |
|---|---|---|
| Trigger | A human writes a message | A business event: email, form, data change, schedule |
| Output | Text in a chat window | Actions: CRM updated, email sent, document processed, task created |
| Autonomy | None: it waits | Scoped: it decides within its perimeter, escalates beyond it |
| Tool access | Optional, often limited to a knowledge base | Structural: CRM, invoicing, ERP, inbox, internal APIs |
| Works while you're away | Responds to visitors, nothing more | Processes, qualifies, follows up, consolidates, alerts, continuously |
| ROI measurement | Tickets deflected, satisfaction | Hours recovered, processing speed, volume absorbed without hiring |
| Setup budget | €2,000 to €5,000 (RAG on your documents) | €4,000 to €8,000 (scoped), €8,000 to €18,000 (complex) |
The rest of the article breaks down every line. But if you only remember one sentence: a chatbot is an interface, an AI agent is an operational team member.
What a chatbot does very well.
The chatbot isn't a discount agent. It's a different tool, with a specific field where it excels.
Answering recurring questions, all the time. Hours, deadlines, terms, the status of a simple request: a well-built chatbot absorbs 50 to 70% of first-level questions, nights and weekends included. Every question absorbed is time given back to your team.
Answering from your documents, not from the internet. The current generation of chatbots relies on RAG (retrieval-augmented generation): the model looks up the relevant passages in your documentation (procedures, contracts, product sheets) before answering, and cites its sources. The answer reflects your reality, not an average of the web.
Qualifying a conversation before a human. The chatbot captures the request, asks the two missing questions, structures the context. The human arrives on a clean case instead of a "hi, I have a problem."
Its limit is structural, not technological: nothing happens until someone writes to it. A chatbot doesn't follow up on a quote, doesn't process an invoice, doesn't update a CRM at 2am. That's not its role.
What an AI agent does more.
An AI agent works in a loop: it perceives a situation (an incoming email, a filled-out form, a data change), reasons with your business's context, chooses an action, executes it in your tools, checks the result, and repeats until the task is done or it hits its limit. At that point, it escalates to a human, with the case already prepared.
A concrete example, the same one we see at our clients: a lead fills out your form at 10:40pm. The agent enriches it (company, size, sector), scores it against your fit criteria, creates the record in the CRM, routes it to the right salesperson, sends a first personalized reply, and schedules a follow-up in 3 days if there's silence. The salesperson arrives in the morning to a complete case. No human was involved.
So the difference from a chatbot comes down to three capabilities:
- The trigger. An agent doesn't wait to be spoken to: it watches business events and gets to work on its own.
- The tools. An agent has access: reading the CRM, writing to invoicing, sending an email, creating a task. That's what turns a response into a result.
- The guardrails. A serious agent knows its boundaries: what it decides on its own, what it escalates, what it isn't allowed to do. That framing is what makes autonomy safe in production.
For a deep dive into how this works, the use cases, and the deployment method, see our guide AI Agent 24/7 for SMBs.
The 4 levels between chatbot and orchestrated agents.
In practice, there aren't two boxes, there's a scale. Each level adds autonomy, value, and a need for extra framing. Knowing where to set the cursor is 80% of the decision.
The scripted chatbot
Decision trees, pre-written answers, buttons. Reliable and predictable, but rigid: any question outside the script falls flat. Suited to very defined paths (simple appointment booking, short FAQ). Low cost, capped value.
The RAG chatbot, connected to your documents
A language model answers from your knowledge base, with citations. Understands freely phrased questions, answers within your context. This is the right level for first-line support and internal assistance. Setup: €2,000 to €5,000.
The tooled agent, that acts in your systems
The model receives tools: reading an order, changing an appointment, creating a record, sending an email. Triggered by a business event, it executes end to end and escalates sensitive cases. This is the level that changes a process's economics: €4,000 to €8,000 for a scoped perimeter.
Orchestrated agents
Several specialized agents cooperate: one qualifies, another drafts, a third verifies, an orchestrator arbitrates. Reserved for high-volume, high-stakes processes, once a single agent hits its limits. €8,000 to €18,000 depending on scope.
The progression is natural: many of our systems start at N2 on a response perimeter, then gain tools and move to N3 once trust is established. The interface stays, the capacity to act gets added.
Chatbot or agent, function by function.
The choice becomes obvious when you apply it to a real process instead of the abstract. Three situations we run into every week.
Answering vs resolving
The RAG chatbot answers: return conditions, timelines, product documentation. The agent resolves: it looks up the actual order, triggers the return, updates the status, notifies the customer. Frequent question equals chatbot. Request that touches your systems equals agent. The combination of both covers 70 to 90% of level 1.
Capturing vs qualifying
The chatbot captures the site visitor and asks the first questions. The agent takes over outside the chat window: enrichment, scoring, CRM creation, routing, first personalized reply, follow-up on day 3. The chatbot alone generates conversations, the agent alone generates cases ready to sign.
This is where the chatbot doesn't play
Invoice processing, reporting consolidation, monitoring, follow-ups: no human writes a message in these processes, so a chatbot has no foothold. This is the agent's exclusive territory, triggered by incoming documents and changing data. It's also where the gains are the most measurable.
Three questions to decide.
No need for a 40-criteria grid. Three questions are enough to set the cursor at the right level.
- Does the need come from a conversation or an event? If a human asks a question, start on the chatbot side. If an email arrives, a document lands, data changes: that's an agent.
- Is an answer enough, or does something need to get done? Information alone equals RAG chatbot. Action in your tools (CRM, invoicing, scheduling) equals tooled agent.
- Does the volume justify the autonomy? Ten occurrences a month get handled by hand. A hundred occurrences a week justify an agent: that's where the hours recovered start counting in days.
And if the answers point in both directions, that's normal: most businesses end up with both, chatbot up front, agent behind the scenes. The costly mistake isn't choosing one or the other, it's buying a tool before mapping the process. Our full method is detailed in How to Automate a Business Process in 2026.
That leaves the question of the engine: which AI model powers your chatbot or agent? Claude, GPT, Gemini, Mistral each have their strengths. Our comparison Which AI to Choose in 2026 lays out the criteria, and our guide How Much Does an AI Agent Cost breaks down the budgets line by line.
How AUTOMATE ALL builds the right level.
We don't sell chatbots or agents off a catalog. We map the business first, then we build the level of autonomy the process justifies. Understand first. Build next. Stay always.
- Audit & Discovery (1 to 2 weeks): process mapping, real volumes, identification of the relevant autonomy level, costed projected gain.
- Strategy & Architecture (1 week): documented architecture, reasoned model choice, defined guardrails, quote before any commitment.
- Build & Deploy (variable): iterative development, shadow mode for agents, team training, code and prompts 100% client-owned.
- Optimize & Scale (ongoing): active monitoring, monthly reviews, progressive autonomy increase from N2 to N3 once the numbers justify it.
The initial audit is priced at a fixed fee based on your size: Small business €2,500 (2 weeks), SME €7,500 (3 to 4 weeks), Mid-market €15,000 (5 to 6 weeks). Fixed fee, excl. VAT, response within 24 business hours. You leave with a concrete action plan, whether we work together next or not.
"No promises. No jargon. A precise diagnosis, clear architecture, delivered on time."
R. Thiébaut, CTO, SaaS scale-up
To see the six agent families we deploy most, head to our page Custom AI Agents.
AI agent vs chatbot: what people ask us.
What's the main difference between an AI agent and a chatbot?
A chatbot answers questions in a chat window: it waits to be spoken to and returns text. An AI agent acts: it's triggered by a business event (email, form, data change), reasons with the company's context, executes actions in your tools (CRM, invoicing, ERP), and checks the result. The difference comes down to one word: action.
Can a chatbot become an AI agent?
Yes, it's a natural progression. A chatbot connected to your documents (RAG) becomes an agent as soon as it's given tools: looking up an order, changing an appointment, triggering a refund. The conversational interface stays, the capacity to act gets added. Many of our deployments follow this two-stage path.
How much does a chatbot cost vs an AI agent?
A RAG chatbot connected to your documentation costs €2,000 to €5,000 to set up. An AI agent on a scoped perimeter costs €4,000 to €8,000, and €8,000 to €18,000 for a complex multi-source agent. Recurring cost ranges from €120 to €2,500 a month depending on complexity. The chatbot costs less, the agent delivers more value because it executes work.
Is ChatGPT an AI agent?
Not in the business sense. ChatGPT is a conversational assistant: it answers when you write to it, it isn't connected to your tools, and it doesn't work while you're away. A business AI agent runs in the background on business triggers, reads your real data, and executes actions in your systems. ChatGPT can serve as an agent's reasoning engine, but the engine alone doesn't make the agent.
Is an autonomous AI agent reliable in production?
Yes, provided its autonomy is framed: defining what it decides alone, what it escalates to a human, and its hard action limits. Deployment goes through a two-week shadow mode where the agent runs alongside the human before the handover. Our agents run in production 24/7 with 0 hours of manual supervision thanks to this framing and continuous monitoring.
Do you have to choose between a chatbot and an AI agent?
Not necessarily: the two combine very well. The chatbot up front captures requests and answers recurring questions from your knowledge base. The agent behind the scenes executes: it qualifies the lead, updates the CRM, triggers actions. The right starting point depends on your volume and which process eats the most time, which is exactly what an audit determines.