Summary
Asset managers are moving from AI copilots to agents capable of executing investment workflows. Scaling that autonomy requires more than powerful models. Firms need trusted investment data, domain-specific context, deterministic controls, governance, auditability, and expert-authored workflows to automate operations without multiplying model, operational, and regulatory risk.
In our last article, we delved into why operating AI at scale is ill-advised if not impossible without a unified data foundation, and how asset management firms gain a decisive edge when they can build AI workflows on a secure system of record, including a governance architecture offering a complete audit trail for AI explainability and auditability. In part 2, we will offer more advice on how to scale agentic AI to win the race to build the next generation of investment operations.
There are some critical approaches you can take to ensure AI agents will reason accurately: a unified data and governance foundation that provides the necessary financial ontology and context, a bias toward determinism, and tools that empower talented human experts to author workflows.
There is a monumental shift underway in asset management. What was once a question of whether you're going to adopt AI had become a question of when. Now the question of when has been replaced by the question of how to scale agentic AI in asset management. Firms are making costly mistakes, like having adoption outpace oversight and buying first while asking questions later.i There is long-term competitive advantage up for grabs for those who get the how right.
It's a big moment in AI transformation. We are plowing ahead, going from chatbots and digital assistants to LLMs and agentic AI autonomously planning and executing multi-step workflows. The human role is shifting from hands-on operator to supervisor of outcomes. As discussed in part 1, AI tools are entirely dependent on being pointed to standardized, centralized systems of record, the single source of truth for investment data. We are asking AI agents to perform advanced reasoning, which demands a certain understanding of how financial market operations work. This calls for data infrastructure that understands the investment management domain.
Blank-slate platforms like Snowflake and Databricks provide powerful data infrastructure and bundle capabilities like data governance. But they require firms' engineers to expend a lot of time building from scratch investment-specific schemas and the custom transformation logic required to normalize raw data.
An AI agent that autonomously executes reconciliations needs to grasp the legitimate reasons why differences arise, different FX rate application conventions, and different settlement times. Settlement conventions differ by asset class and jurisdiction; trades can appear differently depending on whether they’re booked on trade date or settlement date, and corporate actions require retroactive adjustments across positions and P&L records. To research data breaks and recommend adjustments for complex OTC instruments, the reconciliation agent must understand which discrepancies warrant investigation and which are normal parts of an investment lifecycle. An investment-domain-aware reconciliation agent reduces manual effort and allows reports to be generated in hours instead of weeks.
Boston Consulting Group found that a traditional asset manager with a cost of 15 to 20 basis points that reshapes their organization to deploy AI at scale could reduce expenses by 3 to 6 basis points, perhaps a 25% to 30% cut.ii Our production metrics show that deploying governed AI over a trusted data foundation yields up to 65% time savings and 64% error reduction in automated operations.
With governed data infrastructure in place, AI agents can handle the manual drudgery of data mapping, pipeline building, and exception diagnostics. This shifts the human's role from mind-numbing data entry to high-value oversight, analysis, and strategic decision-making. However, deploying that AI without an investment-native, unified data foundation with full governance can reverse those efficiency gains. As a firm scales agentic AI, errors come with much higher stakes, rippling across the workflows through NAV calculations, investor reports, and regulatory filings.
Bias toward determinism is an architectural and product philosophy designed to eliminate live model risk and the inherent unpredictability of LLMs in financial workflows where the tolerance for error is near zero. If you think generative AI is inherently probabilistic, you are right. LLMs are predictive models trained to generate plausible outputs, but they do not naturally calculate or reason with absolute mathematical accuracy. In investment operations, we cannot afford hallucinations. Getting a slightly different result day-to-day for the same query is unacceptable. If we continue down a probabilistic road by training models to become more accurate, this is what results:
"The increase of accuracy in LLMs only enhances the syntactic or probabilistic plausibility of output, rather than their epistemic validity. In other words, a model could become better at predicting what sounds correct based on training data, without any grounding in whether the content is actually true, justified, or verifiable."
When generating sensitive investor reports, the formatting, calculations, and charts must be flawless to prevent fibbing to investors. So, instead of having an LLM dynamically write and assemble the presentation, the LLM is used to generate a deterministic process to extract the correct numbers. Those figures are then seamlessly fed through a standardized template engine or mail merge. A reconciliation agent can use probabilistic reasoning to investigate a reconciliation break, for example. But once it determines that a validated rule applies, the actual operational action should be executed using deterministic logic, instead of improvised by an LLM every time. The human being and a probabilistic model collaborate to perform the workflow planning, but deterministic AI systems execute the tasks.
This bias towards determinism also ensures that agents' behavior can be explained, audited, and tested, because it is frozen in a historical record. Additionally, this dramatically reduces token consumption, costs, and model drift in production.
The sum of the above, a unified investment system of record, domain-specific financial context, deterministic controls, and embedded governance, which nets out as AI-ready investment operations, needs one more crucial element. People. Asset management firms' human (investment management) domain experts play an essential role in scaling agentic AI.
We are not replacing human judgment with machine judgment. AI's value comes from doing things that are inefficient for a human to do, namely making sense of impossibly large datasets. Your teams play an active role but not a manual role, using their ingenuity and experience to author autonomous agents, review outcomes, and make higher-value decisions on escalated exceptions. This empowers the firm's leaders and provides a more seamless process than business users continually translating requirements through engineering teams. This important element requires a flexible AI orchestration layer with structured authoring environments, pre-approved templates, and validated workflow patterns.
For a mid-sized asset manager with $500 billion in AUM, a complete AI transformation could capture 25% to 40% of their total cost base in efficiencies. At the same time, only 17% of institutional investors feel their current systems are ready to leverage AI.iv How is your firm moving to the next phase of AI transformation? Getting the how wrong results in remediating activities that stall momentum and frustrate those AI is supposed to help. Getting the how right positions the firm to advance agentic AI in asset management and ultimately build the next generation of investment operations better. The upside of AI transformation is so monumental that managers that do it right are poised to gain real separation from the pack.
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 Thomson Reuters, August 3, 2026. https://tax.thomsonreuters.com/news/ais-hidden-price-tag-millions-spent-returns-still-a-mystery-and-mistakes-can-make-the-bill-worse/
ii BCG, August 6, 2026. https://www.bcg.com/publications/2026/reimagining-asset-management-with-ai
iii Computer Law & Security Review Volume 61, July 2026, https://www.sciencedirect.com/science/article/pii/S2212473X26000520#sec0001
iv McKinsey On Investing, November 2025. https://www.mckinsey.com/industries/private-capital/our-insights/mckinsey-on-investing
No spam. Just the latest releases and tips, interesting articles, and exclusive interviews in your inbox every week.
