AI agents in NBIM: fund transformation

Multi-Agent AI Systems in NBIM Investment Fund Management

Norges Bank Investment Management (NBIM), which manages the world's largest sovereign wealth fund, has integrated AI across all departments over two years. Instead of focusing on isolated projects, multi-agent architectures have been implemented in 171 business processes. Over 50% of employees have learned to code, meeting preparation time has been cut by 80%, and trading costs have been reduced. Training has become mandatory for all, with Scrum replaced by micro-teams of two developers and one business specialist.

Analyzing Over-the-Counter Trades in the Investment Department

A multi-agent system has been deployed in the investment department to evaluate large block trades. Agents simultaneously aggregate data from external sources, internal databases, deal texts, and news. They calculate index effects, seller context, and fair prices.

Case example: Goldman Sachs' offer of a Ferrari stock package worth 30 billion kroner. Previously, data collection took an hour; now it takes minutes. A team of five manages assets worth 2 trillion kroner, generating billions in excess returns annually.

Agent architecture:

• Agent 1 (web search): identifies beneficiaries through holding structures.
• Agent 2 (parsing): extracts terms, deadlines, sanctions from deal text.
• Agent 3 (algorithmic): models impact on index trackers.
• Additional agents: analyze deal history, market expectations, pricing.

The prototype was developed in-house, then refined by one specialist. Time ratio: data collection—20%, analysis—80%. The fund handles about 200 such requests per year.

Media and Social Media Monitoring in Communications

The communications department created the Echo platform to analyze 50,000+ publications annually. The system classifies sentiment, media priority, the fund's role in the material, topics, and individuals. A Snowflake-based chatbot generates reports on social media (LinkedIn, Instagram, YouTube).

A team of two specialists built the solution independently, avoiding expensive external services. Data is stored in the corporate Snowflake.

Main agent components:

• Sentiment classification (positive/negative/neutral).
• Assessment of the fund's role (mention/quote/main subject).
• Source priority (top media vs. blog).
• Content type (news/analysis/investigation).
• Topic and person extraction.

Functionality:

• Timeline of negative spikes with details.
• AI insights on key drivers.
• Echobot: queries like "social media engagement analysis" provide trends and recommendations.

In 2025—50,000 mentions; in 2026—over 5,000 per quarter. The system is cheaper and more accurate than commercial alternatives.

Automating Investigations in Cybersecurity

An AI agent in the cybersecurity department investigates incidents in parallel with humans. From a trillion logs annually, it narrows down to hundreds of thousands of suspicions, delivering a report in 5 minutes.

Scenario: a nighttime anomaly—connection to a suspicious site. The agent gathers context: logs, history, relevance judgments.

Advantages: fixed speed (6x faster than humans), no fatigue on routine tasks. The analyst receives a ready report for decision-making.

Preparing for Meetings with Top Executives

An AI assistant for 3,000 annual meetings with CEOs and chairpersons. A multi-agent system builds plans, researches sources, and checks quality against internal best practices.

Process:

• Upload internal data (hypotheses, notes).
• Planner agent forms the structure.
• 3–5 sub-agents scan news, reports.
• Final agent validates against top examples and fund techniques.

Savings: 10,000 hours per year (3 hours per meeting). Conclusions are traceable—with prompts and sources to prevent hallucinations. Plans: simulating an interviewer with speech synthesis and feedback on negotiations.

Compliance: EVA Agent for Market Alerts

The EVA system investigates suspicions of insider trading and manipulation. Six sub-agents check each alert:

• Deal context.
• Index rebalancing.
• Company news.
• Industry events.
• Timing patterns.
• Contacts with the issuer.

The master agent aggregates, forming an audit trail. Humans intervene in 3 cases: ambiguity, non-automatability, legal responsibility. The same team handles more incidents without false positives.

Key Takeaways

• Implementation scale: AI in 171 processes, 50% of employees code.
• Agent architectures: key to parallel analysis of complex data.
• Resource savings: 80% on meetings, 6x in security, excess returns in trading.
• In-house development: rejecting external tools for custom solutions.
• Traceability: protection against hallucinations through sources and prompts.

— Editorial Team