Scaling AI Assistants Across Swedish Government Offices Through Rapid Experimentation and Business-Led Innovation

government_of_sweden · Government · 2024 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Government knowledge work and policy analysis

The Government Offices of Sweden needed to help civil servants analyze large volumes of documents and internal knowledge faster while preserving accountability, transparency, public-sector governance, and human responsibility for decisions.

Constraints:

• The organization has roughly 6,500 people and about 100 embassies worldwide.
• The assistants operate in public-sector workflows with accountability, transparency, GDPR, and sensitive-information concerns.
• The initiative deliberately focuses on cognitive enhancement of civil servants rather than automated decision making.
• Model and cloud choices affect which sensitivity levels can be handled.
• Adoption depends on early adopters and core-business ownership, not an IT-only technology push.

Architecture

Sweden's Government Offices run an innovation-incubator model that rapidly prototypes assistants, kills weak pilots, scales useful ones, and centers civil servants as accountable users. Assistants are built in Intric as packages of prompts, knowledge, data, and RAG databases sent to selectable language models, with civil servants trained in prompting and core-business teams deciding which assistants should scale to department, ministerial, or office-wide use.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  A["Civil servants"] --> B["Intric assistant platform"]
  C["Innovation incubator"] --> D["AI pilots"]
  D --> B
  B --> E["RAG knowledge base"]
  B --> F["Language model providers"]
  C --> G["Center of excellence"]
  G --> H["Capability council"]
  H --> A

Components

Data flow

Civil servants identify a knowledge-intensive workflow, structure and upload knowledge sources into an Intric assistant, encode the manual analytical framework as prompt/context, optionally use RAG over uploaded material, and review the assistant's output before it moves further in the public process. Pilots that work are candidates for broader rollout through a center of excellence and capability council.

Why this architecture

The architecture keeps AI behind accountable civil servants because public processes require clear responsibility and transparency. It uses a cloud assistant platform and multiple model choices because the organization wants rapid experimentation without a large GPU team, while preserving future flexibility for sensitive-information handling.

Models used: Anthropic models · OpenAI models · Llama · Sweden-hosted open-source models · Intric assistants

Scale

• The Government Offices have roughly 6,500 people.
  “we're you know roughly 6 and a half thousand people”transcript, stated
• The initiative had at least 30 assistants by spring.
  “We have 30 in in the spring and we've killed some of them that hasn't worked and and some that has worked we've built upon.”transcript, stated
• The company-analysis assistant reduced an analytical cycle from 24 weeks to 6 weeks.
  “And now they're doing the same analytical work with higher quality in 6 weeks”transcript, stated
• The company-analysis workflow covers about 40 state-owned companies with 850 billion kronor turnover and 122,000 employees.
  “it's actually 40 companies >> and they turn around 850 billion croners and they employ 122,000 people.”transcript, stated
• The environmental-permit example involved roughly 25,000 expected pages of information.
  “So, they were expecting somewhere around 25,000 pages of information about these”transcript, stated

Alternatives considered (roads not taken)

IT-led technology push

Rejected because: The early digitalization strategy was criticized as too technology-oriented and was turned toward business focus.

One fixed language model

Rejected because: The team wants model agnosticism so sensitivity, storage, cost, and capability tradeoffs can be handled by use case.

Automated decision-making first

Rejected because: For current assistants, the government keeps civil servants accountable and avoids transferring agency.

Patterns

human-in-the-loop-gate (governance)

Assistant outputs remain behind civil servants, who retain accountability before results move through government processes.

rag-retrieval (retrieval)

Intric assistants package uploaded knowledge sources and RAG databases for civil-servant analysis workflows.

model-routing (model-ops)

The platform can choose Anthropic, OpenAI, or Sweden-hosted open-source models depending on sensitivity and storage constraints.

eval-driven-development (evaluation)

The incubator rapidly prototypes assistants, kills those that fail or are too risky, and scales those that work.

context-engineering (context)

The company-analysis assistant encodes a long manual analytical and judgment framework into the system prompt.

Tradeoffs

Adoption strategy

Chose early adopters and workshops over PowerPoint persuasion.

Rationale: The team believes practical demonstrations and early adopters convert civil servants more effectively than presentations.

Model deployment

Chose cloud and selectable providers over in-house GPU operation.

Rationale: The government offices lacked GPU expertise and personnel, and wanted easy access and faster experimentation.

Automation boundary

Chose cognitive enhancement of civil servants over automated decision making.

Rationale: This keeps responsibility and accountability in the public process while still giving assistants more retrieval and tool capability.

Gotchas (what bit them)

Public-sector assistants can lose accountability if agency transfers too early.

Root cause: Government processes require clear responsibility for information handling and decisions.

Resolution: Keep assistants behind accountable civil servants while governance matures for more tool use.

Complex assistants require prompt and RAG literacy.

Root cause: Users need to know how the RAG database works and when to trust it.

Resolution: Train civil servants in the craft of prompting and knowledge-source structuring.

Scaling many assistants requires organizational ownership.

Root cause: Successful pilots raise questions about department, ministerial, or office-wide scope and responsibility.

Resolution: Create a center of excellence and capability council with core business in the room.

Evaluation

The initiative evaluates assistants through rapid pilot outcomes, risk screening, business usefulness, early-adopter workshops, measurable workflow reductions such as company analysis time, and decisions about which pilots become department, ministerial, or government-office-wide capabilities.

Metrics: Assistants created and killed · 24-to-6-week company-analysis cycle reduction · Early-adopter workshop conversion · Risk under AI Act and sensitivity constraints · Department/minister/government-office scaling level · Potential 20-30% civil-servant efficiency gain

Outcomes

• The Government Offices built at least 30 assistants by spring.
• The company-analysis assistant reduced a 40-company analytical cycle from 24 weeks to 6 weeks.
• The initiative established a phase-two center of excellence and capability council approach.
• The team adopted a business-led, early-adopter model rather than an IT-only push.
• Assistants remain accountable to civil servants rather than making automated decisions.

The source reports roughly 6,500 employees, at least 30 assistants, a 24-to-6-week workflow reduction, 40 companies with 850 billion kronor turnover and 122,000 employees, and a target of potentially hundreds of assistants.