AI Agent or Chatbot – What's the Right Level?
The difference is simple to describe: a chatbot answers questions, while an AI agent also acts – it uses tools, retrieves data, and performs tasks in your systems. That step raises both the value and the requirements: agents need guardrails, access control, and ongoing evaluation. Choose the level based on the case's value and the risk when something goes wrong.
AI agent has become the sales term slapped on everything – and companies risk buying the wrong ambition level in both directions: an agent when a chatbot would have been enough, or a chatbot when the value was in having someone actually carry out the work. Fortunately, the difference is easy to define, and the right level can be reasoned out.
The difference in one sentence: answer or act
A chatbot answers. It receives a question, optionally retrieves knowledge from your content, and formulates a response. If something goes wrong, the result is a bad answer – embarrassing, but reversible.
An AI agent acts. It has a goal, access to tools – systems, APIs, databases – and the ability to work in multiple steps: look up, decide, execute, verify. If something goes wrong, the result can be an incorrect change in a business system – not always reversible.
That difference, the risk of an answer versus the risk of an action, is the core of the whole decision.
The step costs – in money and in requirements
Going from chatbot to agent isn’t a feature upgrade, it’s a step up in responsibility. Three things drive the cost:
- Integrations. The agent needs secure, permission-controlled access to every system it acts on – often the biggest build item.
- Guardrails. Limits on what the agent may do, caps on amount and scope, blocked case types, escalation paths, and full logging.
- Evaluation. Agents fail in unpredictable ways, so testing and ongoing evaluation is the single largest cost item in serious agent projects – and it continues after launch.
In kronor: a platform chatbot costs from SEK 5,000–20,000 a month and a custom-built chatbot SEK 400,000–1,500,000, while an agent that acts on your systems typically lands at SEK 500,000–2 million depending on the number of systems, risk requirements, and testing depth. It depends – but on predictable things: more systems, more sensitive actions, and higher control requirements move you up the range.
The ladder: three levels with criteria
| Level | Characteristics and when it's right |
|---|---|
| 1. Chatbot | Answers based on your content, does nothing in the systems. Right when the need is information: FAQs, terms, product questions, internal knowledge search |
| 2. Assisted agent | Prepares actions, a human approves. Right when actions have real consequences but the manual flow is too slow: case handling, record maintenance, quote preparation |
| 3. Autonomous agent | Carries out actions itself within hard limits. Right when volume is high, the actions are reversible or low-risk, and evaluation has shown stable quality over time |
The criteria for stepping up are the same at every level: measurable quality at the current level, reversible or risk-limited actions at the next, and a log that makes every decision auditable after the fact.
A scenario: the same case at three levels
A customer emails: the shipping address on order 4711 is wrong. The chatbot correctly tells the customer how to change the address themselves – the customer does the work. The assisted agent looks up the order, prepares the address change, and queues it for the handler, who approves with one click – handling time drops from minutes to seconds. The autonomous agent verifies the customer, changes the address within policy – the order hasn’t been picked, the address is within the same country – logs the action, and confirms with the customer. Outside the rules, it escalates to a human.
All three are reasonable solutions. Which one is right depends on the volume, the cost of a mistake, and how mature your evaluation is – not on what sounds most impressive in a steering group.
How to choose the right level
Start with the need: list your ten most common cases and mark which are information questions and which require an action in a system. If the actions are few or rare, a chatbot will do for a long time. If the actions are many, choose the case with the lowest cost of error and the highest volume, and build an assisted agent there first. Autonomy is something the agent earns with data, not something it launches with.
At Weapp we build both chatbots and AI agents – and we’re happy to help with the level analysis before any quote gets written. Get in touch and we’ll look at your case types together.
Frequently asked questions
Is an AI agent always better than a chatbot?
No. The agent is more expensive to build, requires more governance, and introduces new risks because it actually does things in your systems. If the need is to answer questions based on your content, a chatbot is the right level – buying agent capability that goes unused is paying for risk without benefit.
What is meant by an assisted agent?
An agent that prepares actions but lets a human approve before they're carried out: it looks up the case, proposes the solution, and drafts the response, but the handler presses the button. It's often the best middle step – large time savings without giving up control of the outcome.
What guardrails does an AI agent need?
Minimal permissions in every system, hard limits on which actions it may perform and on amounts or scope, logging of every step, blocks on sensitive case types, and a defined path to a human. On top of that, an evaluation suite that runs continuously, not just before launch.
What does a chatbot cost compared with an agent?
A chatbot costs from SEK 5,000–20,000 a month as a platform service, or SEK 400,000–1,500,000 custom-built. An agent that acts on your systems normally costs SEK 500,000–2 million, where testing and evaluation is the largest line item. The integrations and risk requirements drive the difference.
Can we start with a chatbot and upgrade to an agent later?
Yes, and that's often the right path – provided the chatbot is built with the upgrade in mind: a tool layer that actions can be attached to, a permission model that can be extended, and logging from day one. Then the agent step becomes an addition rather than a rebuild.