The problem

At Altruist, maintaining accurate, reconcilable financial data is foundational. As a custodian, we process large volumes of transactions, holdings, pricing, and cash data, and every account must be correct, consistent, and explainable at all times.

Daily reconciliation requires that yesterday’s positions plus today’s transactions equal today’s positions. In practice, financial data is inherently messy: late or corrected transactions, asynchronous pricing updates, and corporate actions introduce edge cases that static rules cannot always anticipate.

When positions fail to tie out or AUM movements lack explanation, teams rely on manual SQL analysis, cross-dataset comparisons, and institutional knowledge to investigate. This investigative burden is inherent to financial infrastructure and represents the problem space we sought to improve.

Traditional approach

Historically, reconciliation and data quality required heavy manual effort, supported by three full-time engineers. Improvements in alerting and pipeline reliability reduced the workload to two, but new edge cases continued to surface outside existing checks. While these improvements reduced noise, they did not eliminate the need for human reasoning, exposing the limits of static alerts and motivating a new approach.

Our approach

While alerts and pipeline improvements reduce noise, they do not eliminate the need for complex investigation. Financial data can appear correct yet still require contextual reasoning to resolve inconsistencies. The challenge was accelerating the investigation of unknown issues. To address this, we built three AI agents that augment existing controls with automated investigation, explanation, and remediation.

  • Reconciliation Agent: Ensures account-level correctness by validating position roll-forwards and investigating gaps. 
  • Net Flow Alert Agent: Flags and explains unusual AUM movements not driven by markets or expected transactions. 
  • Check Reversal Agent: Resolves complex check reversal tickets by generating targeted corrections across transactions, positions, and cash balances. 

Together, these agents do not remove financial complexity, but they materially reduce the time and effort required to manage it.

Implementation

We integrate Claude AI agents into our financial data pipeline to accelerate investigation and remediation.

Custodial data flows through ETL pipelines into our Enterprise Data Warehouse (EDW), the system of record powering billing, performance reporting, and trading. Data quality has traditionally relied on automated alerts and manual investigation, with engineers identifying root causes and applying fixes directly in our EDW, often through time-intensive workflows.

AI agents enhance this process by targeting the most expensive steps: investigation and remediation. The Reconciliation, Net Flow Alert, and Check Reversal Agents analyze warehouse data, explain inconsistencies, and generate targeted SQL fixes, reducing time-to-resolution while preserving our EDW as the single source of truth and keeping humans in the approval loop.

Measurable impact

The introduction of three AI agents reduced manual investigation by 20-25 hours per week while improving consistency and confidence across account and portfolio-level data:

  • Reconciliation agent: Saves ~2 hours per day during high-priority reconciliation windows.
  • Net flow alert agent: Saves ~2 hours per day by explaining unusual AUM movements.
  • Check reversal agent: Reduces resolution time by ~3 hours per complex ticket.

These agents shorten time-to-resolution, reduce reliance on institutional knowledge, and create a scalable, repeatable approach to resolving new and unpredictable data issues, thus amplifying existing controls rather than replacing them.

Key takeaways

  • AI agents deliver the greatest impact when focused on narrow, well-defined operational problems.
  • Combining deterministic financial logic with AI reasoning drives reliable, explainable outcomes.
  • Claude Skill enables AI to move from experimentation to production-grade financial operations.
  • The true value extends beyond automation to speed, consistency, and confidence.

What’s next

  • Expanding agent coverage to corporate actions and pricing validation.
  • Adding feedback loops to continuously improve agent accuracy.
  • Deeper integration with workflow orchestration and alerting systems.
  • Continually enhancing these agents to improve overall team efficiency.