What is the difference between an AI agent, a chatbot, RPA and a workflow?
A chatbot talks. A workflow pushes data along a path you drew in advance. RPA clicks through a user interface the way a person would. An AI agent decides at runtime what the next step should be. These are four different jobs rather than four competing products, and most companies end up running two or three of them side by side.
AI chatbot
A conversational surface that answers questions. The 2026 version runs on RAG: it searches the company’s own documents and answers with citations. Input is a human sentence, output is text. The moment the bot also performs an action in another system, it stops being a chatbot and becomes an agent wearing a chat interface.
Workflow automation
A path drawn in advance, triggered by an event, walked step by step by the system. The branches are if-then rules, not decisions. If there is an LLM call in it, that call is one step among many (classify, summarise, extract), not the thing in charge. The same input produces the same output on every run.
RPA
A software robot working on the user interface: it opens windows, types into fields, clicks buttons, reads the screen. You need it when the target system has no API and you cannot get database-level access either. RPA does not understand what it is doing. It repeats the recorded click sequence exactly as you recorded it.
AI agent
A system that gets a goal instead of a script. In a loop of model calls it picks which tool to call next and when to stop. This is the inheritance of the ReAct pattern (thought, action, observation, repeat), which our AI agent fundamentals piece covers in detail. The price of that flexibility is that the number of steps changes from run to run.
The dividing line between the four comes down to two questions. Is the next step decided at runtime, and do you reach the target system through an API or through a screen? Those two questions are the spine of the rest of this article.
Which technology wins for which kind of task?
Not your industry and not your headcount. Six task characteristics decide it: whether the process is deterministic, whether it needs language understanding, whether the input format shifts, whether the target system has an API, whether the next step must be chosen at runtime, and how high the volume is. The table below shows row by row which technology wins and which one fails.
| Task characteristic | Chatbot | Workflow | RPA | AI agent |
|---|---|---|---|---|
| The process is deterministic, every branch can be written down | wrong tool | wins | works | waste |
| Free-text input has to be interpreted | wins | yes, with one LLM step | cannot do it | good if it also acts |
| The input format changes from run to run | partly | yes, with one LLM step | breaks | wins |
| The target system has an API or a database | not relevant | wins | unnecessary | good |
| No API, screen only (legacy, Citrix, terminal) | cannot do it | cannot do it | wins | only with an RPA tool |
| The next step has to be decided at runtime | cannot do it | cannot do it | cannot do it | wins |
| Thousands of runs a day at a predictable unit cost | good | wins | expensive, licensed per bot | risky, the step count varies |
| A person converses with it rather than launching it | wins | wrong tool | wrong tool | yes, with a chat interface |
| A wrong step is irreversible (money, customer message) | limited | wins | good | only with human approval |
| You need a step-level auditable trail | partly | wins | good | only with a separate observability layer |
The sixth row is the important one. It is the only task characteristic that nothing but an agent solves, which makes it the only legitimate reason to build one. Every other row has a cheaper answer. If the next step can be described with a finite set of rules, a workflow is not merely sufficient, it is better: predictable, testable and cheap to run.
What do the four approaches cost, and who keeps them running?
The licence fee is the smallest line item. n8n Cloud Pro is EUR 50 a month for 10,000 executions, an unattended Power Automate bot is $150 a month, the raw model cost of a chatbot at 500 conversations is a few euros. The real cost is who fixes it when it breaks, and how often it breaks. Almost no proposal prices that line.
| Approach | Licence and run cost | Typical build | What breaks it |
|---|---|---|---|
| Chatbot (RAG) | raw model cost of 500 conversations/month: 470–9,900 Ft (EUR 1.2–26), depending on model | 1–3M Ft (EUR 2,600–7,800), 4–6 weeks | stale knowledge base, answers with no source |
| Workflow (n8n) | Cloud Starter EUR 20/month for 2,500 executions, Pro EUR 50/month for 10,000 | 100,000–2,500,000 Ft (EUR 260–6,500) per process | API change, expired credentials |
| RPA | Power Automate Process $150/bot/month, Hosted Process $215/bot/month; UiPath Cloud Basic from $25/month | weeks per process, a licence per robot | any interface change in the target system |
| AI agent | 1,000 documents through an eight-step loop: $232 on Claude Sonnet 5, no caching | 5–15M Ft (EUR 13,000–39,000), 3–9 months | step-count blowup, prompt regression, model swap |
What matters more than the table is that the four bill on different units. n8n charges per execution regardless of how many steps it contains. Zapier charges per step, Make per module operation, RPA per robot per month, and an agent per token, which means the bill grows with the step count. On a six-step process n8n Pro works out at roughly EUR 0.005 per run and Zapier Pro on annual billing at $0.029, a fivefold gap for the same work. Inside an agent loop that gap widens, because the step count swings between eight and twenty per run, so a per-step bill cannot be forecast.
Operating effort grows in the opposite order. A workflow can usually be maintained by the process owner on the visual canvas. RPA needs someone who re-records the steps after every update to the target system. An agent needs continuous prompt maintenance, an eval set and cost monitoring, and that is human work. Hungarian market rates for it run 16,000–24,000 Ft per hour (EUR 42–62), against 120–160 EUR per hour for a senior at a German agency (software-entwickeln-lassen.com, 2026). That gap is the whole nearshore argument, and it applies to the maintenance years, not just the build. Our own automation maintenance runs 30,000–150,000 Ft (EUR 78–390) a month, and the larger half of that is tuning rather than servers.
One common misreading belongs here. The licence for self-hosted n8n is genuinely zero, but running it is not. On the Hungarian market managed n8n on a VPS starts at 40,000 Ft (EUR 104) a month and on-premise at 80,000 Ft (EUR 208) a month (n8n.hu, 2026-08-14), which is more than n8n Cloud Pro itself. Self-hosting can be justified by data residency or compliance. On cost alone it is not, until very high volume.
Four edge cases where the choice is not obvious
The table settles the clean cases. Most projects are not clean cases. In the four situations below we argue the decision through with the client during discovery, and each one has a single concrete question that ends the argument. These recur at companies of 50 to 500 people across the EU.
1. No API, but the input is free text
A fifteen-year-old ERP that has to receive data taken from supplier emails. RPA can do the data entry but cannot read the email. The LLM reads the email but cannot reach the ERP. The deciding question: is there a read-only database replica or a file-based import? If there is, the answer is a workflow plus one LLM extraction step, and you save the RPA licence entirely. If there is not, RPA becomes the executing arm and the intelligence sits in front of it. That is the larger half of our systems integration work.
2. Support, where 80% of questions are the same twelve questions
Almost everyone wants to build an agent here, when the real question is what share of tickets requires an action in another system. Below 20%, a RAG chatbot is the answer and you escalate the rest to a human. Above it you need the agent, because parcel tracking, changing a delivery address and issuing a credit note cannot be answered, only performed. Measure the ratio instead of guessing it. One month of categorised tickets is enough.
3. Five hundred invoices a day in varying formats
The classic document processing task, and here the decision is a cost question. A single structured extraction call per document is orders of magnitude cheaper than an agent loop that reads, validates, looks the record up in the ERP, corrects it and posts it. The agent, though, handles the missing partner code and the amount that does not match. The deciding question: what percentage of items fails validation? Below 5%, one-shot extraction plus human exception handling wins. Above it, the agent loop earns its price back.
4. The rulebook changes every month
A pricing or approval logic that is deterministic but gets rewritten every quarter. Technically that is a workflow, but the real question is who maintains it. If the business side wants to edit it themselves, the visual canvas is worth more than clean code, even when it is uglier from an engineering point of view. If the rule has to be audited, version-controlled code is the requirement, and we argue for a code-based stack instead of low-code.
The expensive mistake: an agent where a workflow would have done
The money is not lost to picking the wrong tool. It is lost to oversizing. Build an agent on a deterministic process and you pay the higher build price, the multiplied run cost and a class of failure that does not exist in a workflow at all. The numbers on this are fairly brutal.
95%
of enterprise generative AI pilots produced no measurable P&L impact
MIT Project NANDA, July 2025
40%+
of agentic AI projects will be cancelled by the end of 2027
Gartner, 2025-06-25
27×
more expensive to run the same 1,000 documents through an agent loop than through one call
own modelling, Anthropic price list 2026-08-14
The concrete example is ours. Processing a thousand documents with a single structured extraction call on Claude Sonnet 5 costs $8.70. The same work through an eight-step agent loop, where context accumulates step by step, costs $232. Same model, same task, twenty-seven times the bill. What sets the price of an agent is not the model, it is how many rounds it runs, and that is exactly the number most proposals never put in front of the buyer.
Gartner even gave this a name. What they call agent washing is the practice of relabelling an existing chatbot, assistant or RPA robot as an agent: of the thousands of self-described agentic AI vendors they estimate around 130 are real (Gartner, 2025-06-25). The buyer-side mirror image is ordering an agent for a task that never needed reasoning. Anthropic’s own engineering guidance has said the same since December 2024: for most tasks the predefined path beats the autonomous agent.
The German market data points the same way. According to the ifo Institute (2026-06-05), 54.5% of German companies now use AI, but only 18.7% of those users build their own system, so roughly four in five run something off the shelf. That is the right default. Custom work earns its place where the boxed product stops: your own data, your own process, your ERP or WMS.
When do you need to combine more than one of the four?
A production enterprise system rarely contains a single technology. The pattern that works is layered: the chatbot is the surface, the workflow is the backbone, RPA is the last metre into a system with no API, and the agent sits only at the one point where the next step genuinely cannot be predicted.
A typical document pipeline at our end looks like this. n8n watches the mailbox and moves files, a structured LLM call extracts the data, a deterministic validation step checks the partner code and the amount, and only the items that fail go into an agent loop, where the model searches, reconciles and proposes. Human approval sits in front of any movement of money. That layering is what pulls the twenty-seven-fold cost gap down to the share of exceptions.
One compliance item attaches to the chatbot layer. The transparency obligations under Article 50 of the EU AI Act apply from 2 August 2026, and the Digital Omnibus package that entered into force on 27 July 2026 left them in place, unlike the high-risk deadline, which slipped to 2 December 2027 (European Commission, digital-strategy.ec.europa.eu, page updated 2026-08-03). In practice this means telling the visitor they are talking to AI. One line in the widget, and always more expensive to retrofit than to design in.
Human approval, incidentally, is not an optional extra. For money movement, outbound customer communication and edits to live data we build it in by default. At the low-code layer the n8n 2.0 approval step provides it; in code, an interruptible LangGraph graph does. How the two layers fit together is the subject of our n8n versus LangChain comparison, and the full twelve-month cost of the three classic layers is broken down in our chatbot, n8n and custom agent TCO piece.
Summary and frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers questions: the input is a sentence typed by a person, the output is text. An AI agent is given a goal and picks its next tool call at runtime, then performs an action inside another system. If something is being written into an ERP or a CRM behind the conversational surface, that is an agent that happens to have a chat interface.
Will AI agents replace RPA?
No, because they solve different problems. RPA exists so you can reach a system that has no API, through its screen. An agent exists so the next step can be decided when the input varies. The combination we see most often in production: the agent decides, the RPA script is the arm that reaches into the legacy system.
Is n8n enough for a customer-facing production system?
For internal automation and for wiring systems together, yes. For a critical customer-facing core, no: there is no version-controlled code, no unit and integration tests, no CI/CD and no permission-level control. n8n 2.0 (January 2026) added an AI Agent node and a human approval step, but memory resets between runs by default.
What does an AI agent cost compared to a chatbot?
Our ranges: a chatbot or simple automation is 1–3M HUF (EUR 2,600–7,800) in 4–6 weeks, a RAG knowledge base 3–8M HUF (EUR 7,800–20,800) in 2–4 months, a custom AI agent 5–15M HUF (EUR 13,000–39,000) in 3–9 months. The run-cost gap is wider still: the same 1,000 documents cost $8.70 with one call and $232 through an eight-step agent loop on the same model.
How do I know whether I actually need an agent?
One question decides it: can you draw every branch of the process in advance? If yes, you need a workflow, even if it contains an LLM call. If the next step depends on the incoming data and cannot be captured by a finite set of rules, an agent is justified. Anthropic's own engineering guidance says the same thing: for most tasks the predefined path is better.
Do I have to tell customers they are talking to AI?
Yes. The transparency obligations under Article 50 of the EU AI Act apply from 2 August 2026, and the Digital Omnibus package that took effect on 27 July 2026 did not postpone them. If a visitor is talking to a chatbot, you have to say so. The compliance deadline for high-risk systems did move, to 2 December 2027.
Can you build an AI agent in n8n?
As a prototype, yes. The n8n 2.0 AI Agent node handles tool use and vector database connections. For a production, customer-critical agent core you want a code-based framework (LangGraph, Pydantic AI, the OpenAI or Claude Agent SDK), because that is where durable state, tracing, rollback and testability live. A good split: n8n as the glue, code as the agent core.
If you have a specific process and cannot tell which column it falls into, a one-week discovery closes that question. Our process automation page shows what we deliver across the four layers, and our SME automation guide starts with the smaller processes that pay back fastest. If you would rather have a number, ask for a quote and we answer within 24 hours.
