AI agent vs chatbot: difference explained

An AI agent acts autonomously toward a goal. A chatbot answers questions. That's the short version, but the choice between them affects cost, implementation time, GDPR exposure, and how much work the technology actually removes from your team.

This guide breaks down the difference between AI agent and chatbot in eight dimensions that matter for SMB leaders. You get concrete SEK pricing, decision support from our real implementations, and a clear matrix for "in your situation, start with X". If you first need the definition, read our guide on what an AI agent is.

What is the short difference between AI agent and chatbot?

A chatbot answers questions within a defined topic area, usually via predefined flows or a language model without executing capability. An AI agent combines reasoning with tool use, technically via the MCP protocol, and handles entire cases itself. It can book, order, update CRM, and escalate to humans when needed.

The difference isn't the size of the language model. It's whether the system can act in external systems or only generate text.

Anthropic's distinction between workflow and agent summarizes it well: a "workflow" follows a predetermined chain of steps, while an "agent" itself chooses how to reach the goal based on what it sees in the case. Chatbots are often workflows. Agents are reasoning systems with autonomy.

For SMBs who have been considering AI projects since 2023, the difference has in practice been less clear than it is in 2026. Chatbots with LLM backends (like ChatGPT-connected customer support bots) have started being called "agents" in marketing even when they don't act outside the chat. This guide uses strict definition: act = AI agent, just answer = chatbot.

How do they compare dimension by dimension?

The simplest comparison is eight dimensions that actually matter when making the decision. We rank them by SMB relevance, not technical depth, because business value governs which one fits. The table below shows where they differ most, from what they do to ROI and compliance.

Dimension	Chatbot	AI agent
What it does	Answers questions	Acts toward goal
Use case fit	FAQ volume, repetitive answers	Multi-step cases, combines systems
Implementation time	1–2 weeks	2–6 weeks
Total cost year 1 (SEK)	approx. 17,000–180,000	approx. 28,000–175,000
GDPR / EU AI Act exposure	Lower (limited scope)	Higher (tool use + data access)
Integrations	1–2 systems	3–10+ systems
Autonomy & risk	Rule-based, low risk	Autonomous, requires guardrails
ROI pattern	Cost saver (saves time)	Revenue driver (creates new revenue)

Salesforce has a foundational walkthrough of AI agent versus chatbot from a CRM perspective. What we do differently in the table above is rank by "what you get", not by technical autonomy. For SMBs, ROI pattern is almost always the dimension that drives the decision, not architecture depth.

The year-1 total in row four is derived from the article's own pricing section further down, not a loose market range. It counts implementation plus twelve months of operation: a chatbot lands at 5,000–60,000 SEK to build plus 1,000–10,000 SEK per month, an agent at from 10,000 SEK (one process) to 55,000 SEK and up (several processes) plus 1,500–5,000 SEK per month. Internal hours for process mapping apply in both cases and sit outside these figures.

Note especially rows four and eight. The year-1 ranges overlap more than you'd think: a simple agent on one process can land lower than a custom-built LLM chatbot, because the implementation work and the number of integrations drive the price more than the label "chatbot" or "agent". The real difference is in row eight. An AI agent usually creates new revenue (cases that were previously lost), while a chatbot mostly saves time on existing work. Those are two different investment calculations. When a customer says "we want AI", the first question we always ask is: is it about saving time on cases you're already solving, or capturing cases you're missing today? The answer almost always governs the choice between chatbot and agent.

That an autonomous agent can move the entire cost calculation is not just our experience. Gartner predicts that agentic AI will autonomously resolve 80 percent of common customer service issues by 2029, cutting operating costs by 30 percent in the process. It's exactly that combination, more resolved cases without human input, that makes the agent a revenue driver rather than just a cost saver.

When does a chatbot suffice, when do you need an AI agent?

The rule of thumb we use with clients: if the task can be described in a clear flow diagram and only requires answers within one channel, a chatbot suffices. If it requires systems to be updated, data fetched from multiple places, or entire cases handled from start to finish, an AI agent is needed.

Concrete volume thresholds from our implementations:

• Chatbot suffices: Under 50 FAQ cases per day, rules relatively stable, one communication channel (chat or email), no integration against CRM/ERP needed
• AI agent needed: Over 100 cases per day, multi-step flows (e.g., booking that requires calendar check and CRM lookup), integrations against 2+ systems, seasonal spikes requiring auto-scaling

IBM's threshold reasoning around AI agents confirms this from an enterprise perspective. The difference isn't chatbot vs agent as binary choice. It's when the complexity of the case requires the system itself to combine information from multiple sources and decide what to do next.

For a restaurant taking orders by phone, an agent is needed. The order is multi-step (menu lookup, inventory check, delivery address from CRM, POS insertion, SMS confirmation). For an e-commerce site answering delivery questions and returns, a chatbot usually suffices, if volume is under 100 per day. If you want to see where an agent delivers the most value in a webshop, read our guide on what an AI agent does in an e-commerce.

For an in-depth walkthrough of how AI agents fit SMBs in different industries, read our pillar article on AI agents for SMBs.

What do they cost?

A chatbot typically costs 5,000–60,000 SEK to build and a few thousand SEK per month. A consultant-built AI agent costs from 10,000 SEK in implementation and 0–5,000 SEK per month in maintenance, depending on scope. The agent costs more because it handles entire cases, not just answers.

Here is how the ranges look for implementations in 2026:

Type	Implementation	Operating per month
Rule-based chatbot	5,000–15,000 SEK	1,000–3,000 SEK
LLM chatbot (custom)	25,000–60,000 SEK	4,000–10,000 SEK
AI agent (one process)	from 10,000 SEK	1,500–3,000 SEK
AI agent (several processes)	from 55,000 SEK	3,000–5,000 SEK

The operating cost of the AI itself is negligible: fractions of a krona to single SEK per case according to the model vendors' price lists. Anthropic's own pricing page shows in a worked example that 10,000 handled support cases cost roughly 37 USD on their efficient Haiku model, meaning around 4 öre per case. Even an agent with tool calls that burns more tokens stays in single SEK per case. What you pay for is the work around it, meaning the build, integrations, and maintenance, not the model calls themselves.

Two concrete cases for reference:

Sannegårdens Pizzeria cost 52,000 SEK in implementation plus 3,500 SEK/month in operating. The AI agent connects the POS with supplier invoices, calculates cost per pizza in real time, and proposes a finished restock order every Sunday. A chatbot could never have done this work, because it requires autonomous action across multiple systems simultaneously and proactive decisions rather than reactive answers. Value: 32 percent less food waste, 9 SEK higher margin per pizza, and 6 hours per week saved on inventory, landing at around 315,000 SEK/year in net effect. Payback: under three months.

NordicRank cost 65,000 SEK for 18 automated processes (report generation, client onboarding, supplier follow-up, invoice handling) plus 4,500 SEK/month operating. Value: 13.4 hours/week of staff time saved, which calculated against labor cost is around 380,000 SEK/year. Payback: four months.

What decides whether the calculation works out is what the manual alternative costs. A handler taking the cases by hand costs salary plus social contributions. Statistics Sweden's wage data puts the average monthly salary at 41,600 SEK for 2024, and with social contributions the hourly cost lands around 300–350 SEK. That's the figure a chatbot or agent is measured against. The more expensive the manual work, and the more volume tied up in it, the faster the automation pays for itself.

For deeper price walkthrough with ROI formulas and all cost factors, read our dedicated cost guide for AI agents. Generally: an AI agent that handles volume + creates new revenue usually pays for itself in 2–6 months. A chatbot pays for itself in 6–12 months through saved time.

How long does implementation take?

Implementation time differs dramatically depending on complexity. A rule-based chatbot is done in a week, while a complex AI agent with multiple integrations takes four to six weeks. The difference is the number of systems to connect and edge cases to test before go-live. Here's typical time from first meeting to live operation:

Chatbot (1–2 weeks)

• Week 1: Discovery + design of flows + building
• Week 2: Test + go-live on the site

AI agent (2–6 weeks)

• Week 1: Discovery + prioritization. We choose ONE process to start with, not ten
• Week 2–3: Design + development. Specify exact flow, integrate against systems, build in test environment
• Week 4: Pilot on 10–20% of volume, measure hit rate and escalation rate
• Week 5–6: Scaling to full volume. Humans take the escalated cases

If you want to build something yourself, alternatives exist. Microsoft Copilot Studio offers a low-code path that can get going in a few days for a chatbot, but still requires significant work for a multi-step agent with integrations against internal systems.

What changed in 2026 is that AI agents are no longer a risk project. The Stanford AI Index shows that demand for agentic AI skills in job postings grew more than 280 percent in a single year, and that 78 percent of organizations now use AI in at least one business function, up from 55 percent the year before. The technology is production-ready. The biggest risk today is not the technology, but not measuring baseline before the agent goes live so you can't show ROI after six months.

How do GDPR and the EU AI Act affect your choice?

Both require you to declare to customers that they're talking with AI. But an AI agent has higher risk exposure because it handles more data and makes autonomous decisions. That means GDPR requirements become stricter and documentation requires more work.

Concrete requirements in 2026:

• Transparency (EU AI Act art. 50): "You're now talking with our AI assistant" usually suffices at the start of the conversation
• DPA agreement with vendor if customer data is processed
• EU data processing: use Anthropic, OpenAI EU region, or Azure Sweden Central
• Logging + retention: typically 30 days for debug, then automatic deletion
• Right to escalate to human: the customer must always be able to ask for a human handler

Sweden's data protection authority IMY is the authority interpreting GDPR application to AI systems. Their guidance during 2025 has clarified that legality doesn't fundamentally differ between chatbot and AI agent, but documentation needs to be more extensive for agents because they make more decisions and handle more integrations.

Most of the EU AI Act's rules start to apply on August 2, 2026. For most SMB implementations, both chatbots and agents land in the "limited risk" category that only requires transparency. But if your agent makes decisions about credit assessment, employment, or health-related matters, you land in the "high-risk" category that requires substantially more compliance work.

It's worth knowing that the regulation has proportionality built in for smaller players. Fines are weighed against company size, the documentation may be filed in simplified form for small companies building high-risk systems, and several exemptions are written for exactly that size. What actually happens at supervision, and which reliefs lower the risk, we cover in our guide to EU AI Act sanctions and exemptions.

Practically: if you already run manual customer service with staff who know data protection law, the step to chatbot or agent is smaller than it seems. It's mostly about translating existing processes to the AI context and adding clear communication that AI is involved. The biggest pitfall we see in the implementation phase is not legal, but unclear data access policy. Many SMBs give AI systems broader permissions than they would give a human junior employee. That's a risk that becomes obvious only at an audit or customer complaint.

When is hybrid the right answer?

The hybrid model is often stronger than either-or. Chatbot up front for simple FAQ cases, AI agent that takes over when the case requires action or combines information from multiple systems. This pattern is becoming increasingly common in 2026.

Concrete architecture:

• Chatbot tier 1: Handles 60–80% of incoming cases directly. Delivery questions, opening hours, returns, invoices
• AI agent tier 2: Takes over when the chat contains keywords like "book", "change my order", "combine deliveries", or when the customer asks for something requiring CRM update
• Human tier 3: Escalation for complex complaints, crisis situations, or negotiations

That tier 1 and tier 2 together take around 80 percent of the volume is not a figure we made up. Gartner's forecast that autonomous AI will resolve 80 percent of common customer service issues by 2029 points the same way, and matches what we measure in hybrid setups: most incoming cases are repetitive enough to be solved without a human, and only a smaller tail requires tier 3.

This is the same logic as the "agent supervisor" architecture described by Anthropic and IBM. An orchestrator that directs cases to the right level. The difference from chatbot-only is that tier 2 doesn't just answer but acts. The difference from agent-only is that you don't need to build 100% autonomy from day one.

Hybrid is often also cheaper overall. Chatbots solve volume cheaply, agents solve complexity. Trying to get an agent to handle EVERYTHING (even mundane FAQ) becomes both more expensive and less reliable than letting a rule-based chatbot take the simple cases it can.

Another practical advantage of hybrid: you can roll out the chatbot on day 7 and start collecting operating data while the agent is built in parallel. When the agent is live in week 5, you already have 4 weeks of data about which cases the chatbot handles and which actually need the agent's autonomy. That makes scaling more data-driven.

What does Eteya recommend based on its implementations?

Here's our recommendation matrix based on our implementations at SMBs. It's not scientific but empirical, built on what has actually worked for customers of different size and industry. Start from your case volume and how many systems each case touches, and the matrix usually points the right way:

You have under 50 FAQ cases per day, no integrations needed: Chatbot. Implementation cost 5,000–15,000 SEK, operating 1,000–3,000 SEK/month. ROI on saved time within 6–9 months. You start simple and can upgrade later.

You have 50–150 cases per day, one or two integrations (CRM or calendar): AI agent for one process. This is the "sweet spot" for SMBs. Implementation from 10,000 SEK, operating 1,500–3,000 SEK/month. ROI within 2–6 months through combination of saved time and new revenue.

You have over 150 cases per day, several systems, seasonal spikes: AI agent for several processes or hybrid. Here you earn the most from automation. Implementation from 55,000 SEK, operating 3,000–5,000 SEK/month. ROI within 3–6 months. This is often restaurants, booking services, and e-commerce over 5 million EUR turnover.

You need both but are unsure: Hybrid with pilot. We start with an agent on one specific process (like phone orders), connect a simple chatbot for FAQ, measure pilot data for 4–6 weeks, and scale based on what the data says.

According to a Google Cloud study of 3,466 executives across 24 countries, 88 percent of early agent adopters see positive returns on at least one use case. We see the same pattern in our cases when three things align: clearly defined process, baseline measured before go-live, and 2–3 months of incubation period after rollout.

Three anti-patterns we've seen fail:

1. "We want to automate EVERYTHING". Too broad scope. Choose ONE process first, measure, expand
2. No baseline measured. The customer doesn't know what they saved after six months. Measure volume, cost, customer satisfaction BEFORE go-live
3. No one owning guardrails. The agent is left on autopilot. Someone must review escalated cases weekly during the first quarter

What's happened with our first clients after 12 months is interesting. Those who started with chatbots are now upgrading to agents as FAQ volume has grown. Those who started with agents are expanding to multi-process. No one has gone back from agent to chatbot. That says something about how the technology has matured during 2025–2026: once you have a working agent, you don't want to lose the autonomy.

Frequently asked questions