AI agents: what they are, what they do and what they cost
What is an AI agent, what can it do for your business, and what does it cost? Examples, prices, model costs and the risk of an agent that does too much.
- Published
- Reading time
- 8 min
In short
- An AI agent uses a language model to plan and carry out tasks in your systems, such as looking up an order and queuing a reply for approval.
- With us, a pilot starts at NOK 30,000 and an agent in production with integrations from NOK 130,000. The language model itself often costs only a few hundred kroner a month.
- The biggest risk is letting the agent do more than it should. Give it only the access it needs, have a person approve what it does, and log everything from day one.
An AI agent is software that uses a language model to plan and carry out tasks in your systems, rather than just answer questions. It can, for example, look up an order, draft a reply and queue it for approval. With us, a pilot starts at NOK 30,000, and an agent in production with integrations from NOK 130,000.
What is an AI agent?
An ordinary AI assistant, like ChatGPT or Copilot in the chat window, answers what you ask. You do the rest. An AI agent is given a goal, decides which steps and tools it needs, and carries them out.
In Building effective agents, Anthropic separates two kinds of systems. In a workflow, the model and tools are orchestrated through fixed steps coded in advance. An agent decides for itself how to solve the task and which tools to use along the way. OpenAI's practical guide describes an agent as three parts: a model, tools and instructions.
| AI assistant | Workflow with AI | AI agent | |
|---|---|---|---|
| Who controls the steps? | The user | The code, in fixed steps | The model itself |
| Uses tools and systems | Rarely | Yes, in a predefined order | Yes, chooses for itself |
| Typical example | Drafting text in the chat | Classifying incoming email and routing it to the right queue | Resolving a customer request from start to approved reply |
| Risk | Low | Low to medium | Highest, needs controls |
In practice, many of the best systems we build are workflows with an agent step where one is needed. They're easier to test and cheaper to run than an agent that decides everything itself.
What can an AI agent do for a business?
Agents work best on tasks that have many steps, follow fairly fixed rules and today require someone to copy and paste between systems. Some examples:
- Customer service: look up order and delivery status, draft a reply and queue it for approval before it is sent
- Quotes and sales: pull customer data from the CRM, the price list and earlier quotes, and draft a new quote
- Finance: check incoming invoices against the purchase order and goods receipt, and flag discrepancies
- Case handling: gather the documentation in a case and prepare a basis for a decision that a case officer makes
- Internal knowledge: answer employees' questions with sources from your own procedures, contracts and handbooks, using RAG
- Tenders: monitor new tender notices and assess which ones fit the business
According to Statistics Norway (SSB), 30% of Norwegian enterprises with at least ten employees used AI in 2025, up from 20% the year before. Among enterprises with 100 or more employees the share was 58%. Of those not using AI, 77% say the reason is a lack of skills.
What about Klarna?
Klarna is the example everyone points to. In February 2024 the company announced that its AI assistant handled two thirds of customer service chats in its first month and did the work of 700 employees. In May 2025 the CEO admitted the company had put too much weight on cost and that quality had dropped. Klarna started hiring people for customer service again.
The lesson: measure an agent on quality as well as cost, and keep a person in control wherever it matters to the customer.
How do you build an AI agent?
An agent meant to work in a business is built in roughly this order:
- Pick one process. Start with one task with a clear payoff and a clear answer to what is correct.
- Map data and systems. Which sources does the agent need, and which systems will it read from and write to?
- Give it tools. Each tool is a bounded function the agent can call, such as "look up order" or "draft email". The Model Context Protocol (MCP) is an open standard for connecting AI to such systems. Anthropic launched it in November 2024. It is now run by the Agentic AI Foundation under the Linux Foundation, and ChatGPT, Gemini, Copilot and VS Code are among the tools that use it.
- Write instructions and limits. What should the agent do, what must it never do, and when should it ask a person?
- Test on real cases. Build a test set of historical cases with known answers, and measure how often the agent gets it right.
- Run a pilot with approval. Let the agent propose and a person approve. Give it more freedom where it's proven accurate.
- Operate and monitor. Log everything, track quality and cost, and update when models or systems change.
Can we build an agent in ChatGPT or Copilot ourselves?
Yes, for simpler tasks. ChatGPT has custom GPTs, and Microsoft Copilot Studio lets you build agents in a graphical interface connected to Microsoft 365. That is often enough for internal assistants that answer from documents.
When the agent will write to business systems, handle personal data, connect to an ERP or be used by customers, you need code, tests, access control and logging. At that point it's an ordinary software project, and we build it as one.
What does an AI agent cost?
The cost has two parts: development and ongoing operation. The prices below are our own estimates, excluding VAT.
| Scope | Price |
|---|---|
| Pilot: one process, one integration, tested on real data | from NOK 30,000 |
| Agent in production: several integrations, logging, approval steps and monitoring | from NOK 130,000 |
| Larger AI system: several agents, search in your own data and access control | from NOK 500,000 |
| Operation and further development | from NOK 6,000 a month |
What drives the price up is the number of integrations, how tidy the data is, security and privacy requirements, and whether the agent only proposes actions or actually carries them out.
What does the language model itself cost?
The model cost is often lower than people expect. A worked example: a customer service agent handles 1,000 conversations a month, and each conversation uses around 10,000 input tokens (instructions, context and tool results) and 1,000 output tokens.
| Model | Price per million tokens (in / out) | 1,000 conversations a month |
|---|---|---|
| Claude Sonnet 5.5 | USD 2 / 10 | about USD 30 |
| GPT-6.1-sol | USD 2 / 10 | about USD 30 |
| Gemini Flash | USD 0.75 / 3.75 (doubles from 1 January 2027) | about USD 11 |
| GPT-6-luna | USD 0.10 / 0.50 | about USD 1.50 |
Ten times as many conversations means a bill ten times as large, but with a mid-sized model it is still a few hundred dollars a month. Caching repeated context can cut the input cost sharply. The prices are taken from the providers' pages and change often.
Copilot Studio is priced differently. Microsoft charges per credit, USD 0.01 per credit or USD 200 a month for 25,000 credits, and an action carried out by the agent costs 5 credits. In Microsoft's own example, a customer service agent with 900 customers a day uses 7,200 credits daily, which comes to about USD 2,160 a month on pay-as-you-go.
The risk: when the agent is allowed to do too much
An agent that can act can also make mistakes with consequences. OWASP has two risks at the top of its 2026 list that hit agents directly:
- Prompt injection. Language models don't distinguish between instructions and data. An email or document the agent reads can contain text that makes it do something other than what you asked. The more tools the agent has, the bigger the damage.
- Excessive agency. The agent has more functions, broader access or more autonomy than the task needs.
What we do in practice:
- The agent gets only the tools and access the task needs, preferably read-only.
- Actions that send something out, change data or affect people go to approval.
- Everything is logged: what the agent received, which tools it called and what it decided.
- Content from outside is treated as data, never as instructions.
Agents that talk to customers must also meet the transparency requirements of the EU AI Act: people must know they're talking to a machine. If the agent is used to assess job applicants, credit or access to services, the high-risk requirements apply from 2 December 2027, with human oversight and logs kept for at least six months.
Checklist: before you build your first agent
- Pick one bounded process. It should have a clear payoff and a clear answer to what the right outcome is.
- Measure today's situation. Time per case, error rate and volume let you see whether the agent actually helps.
- Map data and access. Clarify which systems the agent needs, and give it as little access as possible.
- Decide where a person must approve. Anything that is sent out, changes data or affects people should be approved before it happens.
- Build a test set with known answers. Use historical cases to measure accuracy before the agent is let loose.
- Log everything from day one. Inputs, tool calls and decisions are needed both for debugging and for documentation.
- Check the rules. Review data protection, data processing agreements and whether the use falls under the EU AI Act.
- Track the cost. Set a budget for model usage and monitor spend per case.
We build AI agents and automation with logging, approval steps and access control as standard, and we're happy to start with a bounded pilot.
Sources and further reading
- Building effective agents, Anthropic
- A practical guide to building agents, OpenAI
- Model Context Protocol
- OWASP Top 10 for LLM Applications 2026
- AI assistants at work: a practical guide, regjeringen.no (in Norwegian)
- The use of AI has taken off in the past year, Statistics Norway (in Norwegian)
