In February 2024, Klarna turned on an AI assistant built with OpenAI and watched it handle two-thirds of the company's customer service chats within a month. By late 2025, that same assistant was doing the work equivalent to 853 full-time employees. Then Klarna started hiring humans again.
That reversal is the best short answer to a question a lot of teams are still asking in 2026: what's the actual difference between an AI agent and an AI chatbot, and does it even matter which one you build?
It matters more than the marketing copy suggests.
What a chatbot is built to do
A chatbot, even a modern one powered by a large language model, answers one message at a time. You ask a question, it matches your intent to an answer, and the conversation ends there. Most chatbots pull from a knowledge base or a set of scripted flows: FAQ pages, order status lookups, appointment booking. They read, they respond, and they don't decide anything beyond which pre-approved answer fits best.
This isn't a knock on chatbots. For a support team fielding thousands of "what are your hours" and "where's my order" questions, a well-built chatbot resolves the easy majority and frees humans for the harder cases. Klarna's original assistant, for what it's worth, started closer to this end of the spectrum before its scope expanded.
What an AI agent is built to do
An AI agent works in a loop instead of a single pass. It observes the request, reasons about what needs to happen, picks a tool or action, checks the result, and decides what to do next, repeating until the task is actually finished: refund a payment, update a CRM record, cross-check a shipping status against a delivery API, then write back to the customer with a resolution instead of a deflection.
The language model underneath an agent and a chatbot is often the same one. What changes is the architecture wrapped around it: memory across steps, access to external tools, and the freedom to chain several actions together without a human clicking "next" at every stage.
Where the line actually sits
Three things separate the two in practice. A chatbot displays information; an agent reads from and writes to real systems, including databases, payment processors, and calendars. Context is another split. An agent holds memory across a multi-step task, so it doesn't ask you to repeat your account number four times, while a chatbot mostly treats each message as its own event. Then there's autonomy: a chatbot follows a script or retrieves the closest matching answer, while an agent decides which of several possible next steps actually gets it closer to solving the problem.
None of this makes agents strictly better. They cost more to run per interaction, since planning and tool calls burn through more tokens than a single lookup, and a wrong action from an agent (an incorrect refund, a bad CRM update) causes more damage than a chatbot giving a slightly off answer. Gartner's research points to a similar pattern: far fewer of the products marketed as "AI agents" actually meet that bar architecturally than vendors claim.
What Klarna's story actually teaches
Klarna's assistant is worth studying because it shows both ends of this trade-off inside one company. The 2024 launch leaned agentic: it read account and transaction data, resolved refunds and cancellations end to end, and cut resolution time from 11 minutes down to under 2. That's agent behavior, not chatbot behavior.
The 2025 course correction is the part most case studies skip. CEO Sebastian Siemiatkowski said cost had become too dominant a factor in how the system was built, and quality suffered for it. Klarna began rehiring humans for support roles it once assumed the AI would fully own. The lesson isn't that agents fail. Autonomy without the right guardrails, escalation paths, and human checkpoints creates a different kind of risk than a chatbot ever could.
Choosing between the two
For a linear, low-risk, high-volume task like answering pricing questions or walking someone through a password reset, a chatbot is usually the right call. It's cheaper to run, faster to deploy, and the failure mode is mild.
For a task that spans multiple systems, needs a decision made with incomplete information, or requires follow-up over several steps, an agent earns its higher cost. Teams that get this wrong tend to either over-engineer a simple FAQ bot into an agent it never needed to be, or under-build a system that customers actually need to act on their behalf.
This is the exact conversation SolveMotive has with clients before a single line of code gets written: which parts of a workflow genuinely need autonomy, and which parts just need a fast, accurate answer. Getting that scoping right upfront saves months of rebuilding later.
The short version
A chatbot answers. An agent acts. The model behind both might be identical, but the architecture, the cost, and the size of the mistake each one can make are not. Klarna's own arc, from scripted support to a near-autonomous assistant to a hybrid model with humans back in the loop, is probably the most honest case study available for what getting that balance right actually looks like.
If your team is trying to figure out where your own product sits on that spectrum, that's worth working out before committing engineering time to either direction. SolveMotive works with teams on exactly this kind of build-vs-buy-vs-hybrid decision.
Top comments (0)