Summary
Once asset owners establish a unified, governed data foundation, they can move AI beyond isolated productivity tools and into enterprise investment operations. Trusted investment data allows AI agents to analyze performance, investigate exceptions, monitor exposures, automate reporting, and support cross-asset workflows while preserving governance, lineage, and human oversight. The result is not simply more AI, but scalable, explainable AI embedded across the investment lifecycle.
For pension funds, endowments, insurers, and sovereign wealth funds, the next phase of AI transformation is about scaling AI investment operations by moving into multi-step use cases. This is an inflection point when a gulf could widen between those with rapidly maturing AI operations and those grappling with hallucinations or inaccuracies, data quality issues, and underwhelming results.
In our last blog post, we talked about why a unified data layer is an unequivocal prerequisite for AI at scale. Now, let's delve into the elements that make up a transformational AI program that can be a competitive differentiator for asset owners: sound autonomous agent reasoning that mitigates operational risk, a flexible AI operations layer allowing internal leaders to create their own agentic workflows, automated data quality and governance backstops, and agents that have a native understanding of investment management.
The endgame of AI transformation should look like this:
Boston Consulting Group analysts wrote: "Build governance from day one. A senior committee must ensure that auditability, human intervention, and other key issues are considered at the design stage. Retrofitting is much harder and often slows organizations down."i
One of the most potent factors that enables all the above is investment domain expertise. Institutional investors that come out on top in AI transformation will also come out on top in operational efficiencies, in speed to reacting to volatile market movements, and in driving alpha. The key to all of this is a live total portfolio view: one view of positions, exposures, collateral, performance, and liquidity. But AI can only produce a total portfolio view if it understands the financial ontology of markets. The big general AI models do not natively understand the language and logic of finance. Moreover, prompt engineering can only go so far in teaching general models because modern institutional workflows are simply too complex, especially given the blending of private and public asset classes in modern portfolios.
We recommend adopting AI infrastructure that natively speaks all the dialects of finance fluently. We're talking about infrastructure that encodes context at the workflow level, as a versioned validated rule. In agent-based modeling and process automation, domain context is a prerequisite to prevent catastrophic operational errors.
"While deep learning models are often seen as ’black boxes’, LLMs’ ability to generate human-like outputs opens doors to explainability. This characteristic facilitates the provision of both results and their underlying explanations, thereby enhancing the comprehensibility of the reasoning processes within LLMs, and increasing trust and transparency in their financial applications... By integrating domain-specific data and parameters, LLMs can be trained to focus on particular aspects of financial markets, such as risk assessment for bonds or trend prediction in stock markets. This approach enhances the analytical capabilities of LLMs, allowing them to generate insights that are finely tuned to the complexities of different financial environments." — A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challengesii
An AI infrastructure layer that cannot understand specialized financial terminologies and sector-specific jargon will struggle to maintain positions across dozens of fund structures, vehicles, regulatory regimes, and asset classes, each arriving with different data formats, different reporting timelines, and different systems of record maintained by different external managers and administrators. For a CIO to grab a total portfolio view to, for example, identify total private-credit exposure, including exposure held indirectly through hedge funds and private-market vehicles, the agent will need a unified data layer and investment domain expertise to look through funds and structures to determine underlying geographic, sector, issuer, borrower, counterparty, credit exposure, credit quality, manager, or portfolio — and show where each number came from. An AI reconciliation agent needs to grasp the legitimate reasons why differences arise across FX rate application conventions and settlement times. Meanwhile, humans stay in the loop, using their ingenuity and experience to author autonomous agents, review outcomes, and make higher-value decisions on escalated exceptions. Investment domain fluency also helps contain risk vectors in operationalizing enterprise AI.
AI model risk is the danger of agents making mistakes, producing biased, outdated, or flat-out errant outcomes. A domain-aware AI foundation helps agents master the cross-asset interplay of exposures, valuation methodologies, and regulatory classifications. Another pivotal method to contain AI risks is a bias toward determinism.
Agents operating in live production environments must not be permitted to improvise calculations or decision-making paths. Because LLMs generate probabilistic responses, relying on a single, unchecked output in live production makes effective risk control nearly impossible.iii In investment management, we cannot get a different answer from an agent every time we ask the same question. AI can use LLM capabilities and a well-defined ontology with a clear set of business rules, at the same time. The combination produces answers that are deterministic where they need to be and auditable throughout.
The centralized, domain-aware AI architecture orchestrates the auditable data foundation that modern regulatory regimes require. This ensures that every model interaction, query, accessed data point, action taken, and operational outcome is recorded in immutable logs, connecting AI behavior directly to institutional risk and controls.
Further, with the right data and AI orchestration foundation in place, an asset owner can put agent development safely in the hands of business users inside its own investment department.
The best people to design agentic workflows for institutional investors are their own experts across treasury, accounting, trading, and reporting. They have boots-on-the-ground knowledge of their day-to-day functional requirements and nuances, and they know what edge cases look like. To enable seamless business user authoring, they need AI architecture with structured templates, validated workflow patterns, and a library of investment domain-specific skills.
For example, a risk leader at an insurance asset owner would ideally be able to create their own AI risk control agents, like a capital impact agent to keep an eye on how asset changes might affect their risk-based capital thresholds.iv Ideally teams can also choose from pre-made core operational agents that execute common functions, like a P&L explainer to examine what exposures caused a loss. This approach to AI architecture offers full control over model selection, access, human oversight, auditability, and governance.
There are some decisions of how pension funds, sovereign wealth funds, insurers, and endowments are adopting AI: build or buy, which models to use, and operating model ownership. However, in institutional investing, we see a few non-negotiable components that make up the secret formula to AI operationalization: a unified data platform so AI is built on a system of record, embedded AI governance controls that enable complete explainability and auditability, and a bias toward determinism. There is also a secret formula for AI optimization if your firm is seeking to lead the pack: data infrastructure with investment domain ontology, business users empowered to author AI agent workflows, and rigorous human oversight and testing.
Rochelle Glazman
Rochelle is responsible for enabling go-to-market and growth strategies across sales, marketing, product, and client engagement. Before taking on this role, Rochelle was a Senior Pre-Sales Consultant, engaging with clients and prospects across the financial services industry. Prior to joining Arcesium, Rochelle spent over five years at BlackRock Aladdin servicing institutional asset managers and leading several implementation projects across North and South America. She graduated from Vanderbilt University with a degree in economics.
Sources:
i BCG, August 6, 2026. https://www.bcg.com/publications/2026/reimagining-asset-management-with-ai
ii Yuqi Nie∗, Yaxuan Kong∗, Xiaowen Dong, John M. Mulvey†‡, H. Vincent Poor, Qingsong Wen, Stefan Zohren, June 15, 2024. https://ora.ox.ac.uk/objects/uuid:439fb47b-91a8-4f22-82f0-34407a08e8de/files/s6q182n742
iii Yuqi Nie∗, Yaxuan Kong∗, Xiaowen Dong, John M. Mulvey†‡, H. Vincent Poor, Qingsong Wen, Stefan Zohren, June 15, 2024. https://ora.ox.ac.uk/objects/uuid:439fb47b-91a8-4f22-82f0-34407a08e8de/files/s6q182n742
iv NAIC, June 30, 2026. https://content.naic.org/insurance-topics/risk-based-capital
No spam. Just the latest releases and tips, interesting articles, and exclusive interviews in your inbox every week.
