AI agent development for Danish companies

What an agent does that a chatbot does not

A chatbot answers. An agent acts: it looks up an order, checks stock, drafts a reply, books a meeting, updates a record. That makes it more useful, and it makes mistakes more costly. Our AI agent vs chatbot article explains the difference. In a Danish company the first agents we usually build are internal: sales assistants that prepare account notes, back-office agents that match invoices to orders, and support agents that gather facts before a person replies.

Our AI agent development service covers the full build, and the agent security checklist is the document we work from.

Permission levels we set

We keep credentials in your accounts, separate the agent's identity from staff identities, and set spending limits on model usage. Every tool call is logged with inputs, outputs and time.

GDPR, the AI Act and Datatilsynet

An agent with access to customer or staff data is a processor of personal data, and GDPR applies. Datatilsynet is the Danish regulator. You need processor agreements with us and with the model and hosting providers, a documented purpose, data minimization and a retention rule for logs. For high-risk processing, a data protection impact assessment comes first.

• GDPR Article 22 limits decisions based solely on automated processing that have legal or similarly significant effects on people. We keep a person in the loop for those.
• The EU AI Act sets transparency duties and classes some uses, such as employment and credit decisions, as high-risk. We check where your use case falls and tell you if it needs a different approach; confirm the current dates and scope with your lawyer.
• India has no EU adequacy decision, so our access to personal data relies on standard contractual clauses and a transfer assessment. We develop against synthetic data where possible.

Testing an agent in Danish

Before launch we build an evaluation set from your real cases: 50 to 100 tasks written in Danish and English, with the expected outcome of each. We run the agent against the set after every change, count correct, wrong and unsafe results, and review the failures with you. The test covers wrong tools, misread Danish phrases, invented data, and attempts to push the agent outside its scope, such as instructions hidden inside an email it reads.

After launch the agent runs in shadow mode for a period: it proposes actions and a person compares them with what they would have done. Only when the match rate is acceptable to you do we move to live approval.

How the work runs

1. ScopeOne task, one team, one set of systems. A fixed quote within one working day of the call.
2. PrototypeA working agent on sample data for you to try within the first weeks.
3. EvaluationTest set, safety checks, and a list of what the agent will refuse.
4. Shadow runAgent proposes, people decide, results compared.
5. Go live and monitorApproval screens, cost alerts, weekly review of logs.

Calls fall between 09:00 and 14:30 Danish time in winter and 15:30 in summer, hours when our team in Pune is also at work.

Where agents go wrong and how we limit it

Agents fail in predictable ways. They pick the wrong tool, repeat a step in a loop, trust text that was planted in a document, or fill a gap with an invented fact. None of these is rare, so the design assumes them. Each agent gets a step limit, a cost limit per run and a list of approved tools. If the model is unsure, it stops and hands the case to a person with the facts it gathered.

• Hard limits on steps, time and spend per run, with an alert when a run hits the ceiling.
• Input from emails and web pages is treated as data, never as instructions.
• Answers must cite the record they came from, so a reviewer can check them in seconds.
• A kill switch your team can use without calling us.

For routine, rules-based work a plain automation is cheaper and more predictable than an agent. We say so when that is the case, and point you to our AI automation service.

Pricing

Projects start at $1,500, and extra work costs $30 an hour. A single-purpose agent with two tools is near the start. More tools, longer workflows, approval screens, an evaluation suite and several languages increase the scope. Model fees are paid to the provider on your account. The AI agent cost guide gives ranges, and the Denmark AI page sets out our general approach to processors and data.