There has been growing discussion in the UK about Norges Bank Investment Management. This is the entity that manages the Government Pension Fund Global, Norway’s Sovereign Wealth Fund. Since 1996, the fund has invested in overseas assets to smooth the natural volatility of oil revenues. The growth of the fund protects Norwegians against the day when fossil fuels run out.

Two weeks ago, AI company Anthropic gave an update on how NBIM is using AI. The fund’s managers claim to have saved $100 million in annual trading costs, 213,000 hours of employee preparatory work, and achieved 95% accuracy in automating voting decisions. That’s across around 9,000 companies the fund monitors in 16 different languages.

The improvements came from embedding standard AI into operational workflows.

NBIM used Anthropic’s Claude to reduce the number of trades it sends to public markets. The cost savings add up given that it performs around 49 million transactions a year. The primary mechanism is predictive modelling.

By studying past performance of the fund, technicians built an internal engine to predict future order flows. The model analyses expected inflows, portfolio rebalancing requirements and future inventory needs.

Armed with this knowledge, the fund’s managers can internalise more of the portfolio activity and thereby save the costs of public trading. Internalisation means matching buyers and sellers within the fund rather than routing both orders through external markets.

The next step is to create prediction profiles for how stocks are likely to trade across 60 stock exchanges. Claude advises on how to trade shares and whether to execute immediately or wait. This reduces the extent to which the fund’s own activity moves market prices.

Claude also identified cognitive bias in portfolio managers. This was something we were investigating before the sale of OTAS to Liquidnet in 2017.

NBIM integrated Claude into its Investment Simulator. This builds profiles for each of the portfolio managers by analysing historical records against stock price moves. The system then warns managers when they display psychological biases, such as herd behaviour or panic selling. This system prevents unnecessary trades from taking place, thus preserving capital and lowering trading costs.

NBIM manages over $2 trillion with fewer than 700 people. It prides itself on operational efficiency. It chose to embed AI into its infrastructure rather than giving every employee a chatbot. This automates data extraction, basic coding and compliance workflows and delivered an estimated 20% productivity gain.

In effect, NBIM did away with the need for complex SQL to interrogate its databases. Employees can now query data in natural language and receive immediate answers. This reduces friction and the bottleneck of central teams helping those without requisite SQL skills.

Next NBIM ingests documents and extracts relevant data from the 9,000 companies it monitors. Portfolio managers can go straight to the more valuable analysis stage of the work. They no longer need to spend hours sifting through public statements looking for information.

One of the core principles of AI adoption is to only use the technology where it is needed. NBIM has strict guidelines for how to vote on issues at company AGMs. Rather than have AI take over this process, it uses LLMs to augment research, ingest documents and run compliance screens. Thereafter, a standard rules engine automates almost all of the fund’s global voting.

One interesting aspect of AI adoption is that around half of NBIM’s employees have coding skills. These people build their own custom automation scripts and avoid waiting weeks in an IT department pipeline. This decentralisation of authority is critical to the fund’s ability to manage at scale with so few employees.

Staff training is becoming a critical part of AI adoption. Our work with larger firms at MSBC often requires us to train teams in how to use the preferred AI tool. This might be Claude, but just as often it’s ChatGPT or Microsoft Copilot.

The important lesson from the NBIM case study and our work is that AI can and should be embedded directly into company workflows. This requires central planning and data preparation that takes the bulk of the deployment time. Once the infrastructure and training are in place, firms can safely decentralise usage.

If the best time for Britain to start a sovereign wealth fund was when it discovered oil, the second best time is now. The problem is that it only has moderate block licensing fees and heavy taxes as a means of raising money.

Norway has a state-owned oil company and direct ownership in over 40 producing fields. Over three decades this has allowed it to make a steady stream of investments in overseas companies operating outside of the oil and gas industry. That makes it one of the world’s leading investors in AI, as well as a test case for how best to deploy the technology.