CalSTRS board hears AI education, approves sandboxed investment proofs of concept

CalSTRS staff outlined efforts to govern and test artificial intelligence across the pension system, telling the board on Sept. 24 that the organization is positioning itself as an early, responsible adopter while putting guardrails in place.

At a board education session, speakers described an enterprise AI policy covering investments and operations, an expanded inventory of AI use cases and a closed “sandbox” environment for investment testing that managers said will use only CalSTRS data and remain on the system’s secure infrastructure. "This will only be closed to our investment data... It's secure. It's only on our location. It doesn't leave anywhere," a presenter said when explaining the sandbox.

The investment branch has identified more than 40 generative AI use cases and selected six for proof‑of‑concept work, including a real‑estate example that uses AI to summarize partner reports and census or demographic data to surface emerging trends for staff analysis. Presenters stressed that AI outputs will be used to inform, not to make, final investment decisions: staff said the tools "help us gather and summarize" information faster but require validation by analysts and managers.

Board members asked for tailored AI training for directors and raised governance and environmental concerns, with one director asking how data‑center energy use would align with CalSTRS’ sustainability goals. Presenters said CalSTRS is working with major platform providers and enterprise compliance, legal and ISO groups to monitor energy and security implications and to ensure responsible procurement and use. "If you don't have good data in there, we can't expect right information out," one presenter summarized, underscoring the importance of data quality and governance.

Staff also noted they are drafting investment‑specific guidelines and a governance framework and are piloting a private AI solution for pension testing that they said is not connected to public, open models and is designed to keep sensitive data private. The board was told the rollout is being staged through a controlled working group and additional training and oversight measures will follow.

The board requested a future director‑focused training session on AI and was told broader innovation updates will return in November, including further governance work and vendor engagement plans. The presenters emphasized that users and managers retain responsibility to validate AI outputs before using them in decision documentation.

The presentation did not propose formal board action; it served as education and a status update on governance and pilot work ahead of broader adoption.