The guide

The article collected advice on designing an AI assistant for a concrete work task. It began with the user’s role, intended outcome, available sources and the instructions that shape how the assistant responds.

What this means for the organisation

An assistant that suggests text requires different controls from an agent allowed to take actions. Permissions, approval, logging and stop criteria should increase with the consequence of error. Testing should also include ambiguous and unwanted requests.

The guide began with role, task, desired result, sources and instructions. This sequence distinguishes a writing assistant from an agent allowed to act. As the action scope grows, approval, logging and stop rules must also become stricter, and tests need to include ambiguous or unwanted requests.

NIST AI RMF structures risk management around govern, map, measure and manage. It provides a vendor-independent frame for reading ProsessPilotene’s guidance and model tests. The demonstrations show particular tasks; the framework also requires organisations to assess context, consequences, measurement and responsibility throughout the lifecycle.

A pilot needs a bounded dataset, expected answers, difficult counterexamples and a named decision owner. Errors should be classified by consequence rather than merely counted. Model and prompt versions should be retained so the organisation can explain changed results before giving an agent wider scope to act.

Sources

Primary source: ProsessPilotene, “Slik lager du din egen AI-agent (AI-assistent),” 25 May 2025 (Norwegian).

NIST, AI Risk Management Framework 1.0, 26 January 2023.

Further reading

How do you build your first agent?

Join the discussion

Which experience from your own organisation is most relevant to this story? Share what you learned.