Name in original language
Konsekvenshjälpen
Initiative overview
Konsekvenshjälpen (“Impact Assessment Helper”) is a public-sector AI assistant which is responsible for guidance and training on regulatory impact assessments under the new Impact Assessment Ordinance (SFS 2024:183). We launched the service as a web-based prototype in 2025 to complement written guidance on the web and to provide on-demand, situation-specific support.
The assistant supports civil servants and inquiry teams across government throughout the impact assessment workflow: framing the problem and objectives, planning the work, proposing and comparing policy options (including the “do nothing” option), identifying affected stakeholders, selecting suitable analytical methods, suggesting data sources, and reviewing drafts for clarity and compliance. The goal is not to replace professional judgement, but to lower the threshold for asking the right questions early and consistently.
Technically, Konsekvenshjälpen combines retrieval-augmented generation with a tightly scoped system prompt. Retrieval is used to fetch relevant passages from the Agency’s curated guidance, while the model is instructed to base answers on (and refer back to) the ordinance and Agency materials. The deployed solution uses an OpenAI vector search component for retrieval and Anthropic Claude for response generation.
The Agency designed the tool with transparency and responsible use in mind: it is explicitly positioned as a drafting and reasoning aid, users are warned about hallucinations and asked not to input personal or sensitive information, and the team performs ongoing monitoring and updates. The Agency for Financial and Public Management expects the approach to be transferable to other jurisdictions by swapping the knowledge base and legal framework while retaining the same human-in-the-loop model of use.
Results, outcomes and impacts
Early user testing with experienced impact-assessment practitioners showed strong perceived usefulness. In immediate post-answer surveys, 96% of responses were rated relevant and usable, and 93% of respondents said they would recommend the service to others. ESV also observed that the tool can help first-time AI users adopt AI in analytical work. A full evaluation after five months of use will soon be concluded. Operationally, low marginal costs are estimated: for a medium-use scenario (200 active users per month, with approximately five questions each), the total annual cost for 2026 is estimated at up to SEK 120,000. Expected impacts include more consistent application of the methodology, fewer overlooked issues, and clearer regulation for affected stakeholders.
