What does it cost to develop an AI agent?

An AI agent differs from a chatbot on one crucial point: it doesn’t just answer questions, it does things. It reads your systems, makes partial decisions, and carries out actions – booking, updating, creating, sending. That makes agents the AI category where the value can be greatest, and where a sloppy build can be the most expensive. In 2026, the price tag for a properly built agent is typically SEK 500,000–2 million.

What’s included in the price

Notice what’s at the top. The agent logic itself – the part that looks impressive in a demo – is a minority item. It’s everything around it that costs, and that determines whether the agent can be trusted.

Why evaluation is the biggest cost item

Traditional software fails predictably: the same bug gives the same error every time. An agent fails unpredictably. It can solve ninety-nine cases perfectly and on the hundredth draw a conclusion nobody foresaw, because it reasons its way there instead of following fixed rules.

You can’t test that away with a few manual spot checks. A serious vendor builds an evaluation harness: hundreds of real scenarios with known correct answers, automatic runs on every change, and measurement points for both the final result and the individual steps along the way. The harness costs money – but it’s also what lets you dare to increase the agent’s responsibility over time. Ask every vendor: “how do we know the agent got better, not worse, after a change?” Anyone without a good answer hasn’t built agents in production.

Guardrails and human-in-the-loop aren’t optional extras

An agent that acts on your systems needs technical guardrails, no matter how good the model is:

• Permission boundaries. The agent may only reach the systems and data the task requires.
• Action limits. Define what the agent may never do on its own – for example send money, delete data, or communicate externally without approval.
• Human-in-the-loop. Critical decisions pass through a human. In the beginning that often applies to most decisions; as outcome data builds trust, the threshold can be raised.
• Escalation and stop. When the agent is uncertain, it should hand off, not guess. And there needs to be a clear emergency brake.

If you see a quote where these items are missing or baked in as a footnote, ask for a revision. This isn’t optional equipment, it’s the prerequisite for production.

A concrete worked example

A property management company wants an agent that handles maintenance reports: reading the report, classifying the case, looking up the property in the real estate system, creating a work order, and booking the right contractor. A PoC for SEK 250,000 shows that classification and lookup work. The agent build comes to SEK 1.2 million, a good third of which goes to the evaluation harness and testing against historical cases. For the first six months, a property manager approves every booking; after that, the agent handles standard cases on its own while unusual ones are escalated. More expensive than a chatbot – but it replaces an entire manual workflow, not just the answers to questions about it.

Work backward from the value

An agent for SEK 1.5 million is cheap if it offloads three full-time positions, and expensive if it saves one hour a week. So start with the process: volume, time spent, cost of error. Factor in operations too – agents often make many model calls per task, so the cost per handled case should be forecast before the decision, not discovered on the first month’s invoice. The wisest approach is to start with a well-defined PoC that tests the hardest step in the chain, before the full agent build is funded. At Weapp we build AI solutions and agents with evaluation and guardrails as part of the standard delivery, and we’re happy to help work out the numbers for your case – get in touch and we’ll take it from there.