AI in the capital markets has moved past the pilot phase. In 2026, production-grade agentic systems are operating inside live market workflows: triggering pre-approved trades, adjusting risk models, and automating compliance reporting. But individual productivity gains don’t automatically become business results. Without the right data architecture, controls, and cost discipline, firms end up “running in place.”
In the third edition of this TabbFORUM report, one message comes through clearly: success at scale depends less on how fast a model answers than on how well the firm’s infrastructure can govern the data behind it and act on the result.
What’s Inside the Report
This TabbFORUM report looks at how capital markets institutions are taking AI from experimentation to execution at scale, including:
- Compute crisis and “sticker shock”: Why consumption-based AI spend compounds quickly, and how firms are applying unit economics such as “cost per idea” and “compute per outcome” before scaling.
- Front office and “action risk”: Where firms draw the line between AI-generated insight and client-impacting decisions, and why many reject fully autonomous AI in trading and investment.
- Middle-office systemic risk: How interconnected multi-agent systems turn model risk into systemic risk, and how ensemble agents, validator agents, and exchange kill switches help contain it.
- Real ROI: Why the clearest returns today come from middle- and back-office operations, document automation, and faster software delivery.
- Enterprise infrastructure and data gravity: Why data movement is a hidden cost of AI, and how unified data fabrics, portable architectures, and purpose-built cloud environments support predictable scale.
- 2026 mandate: Why competitive advantage will go to firms that pair architectural and budgetary discipline with clear human accountability.
Key Takeaways: Data Architecture Drives AI ROI
- Stop treating every AI use case as a new project: Traditional pipelines need repeated ETL cycles and data duplication before information can be used, and that compounds compute costs.
- Keep models away from raw source systems: Agentic AI can’t safely automate post-trade processing, reconciliation, or document extraction without strict guardrails. To scale AI safely, firms need a governed data layer and agentic harnesses that enforce data entitlements, access policies, and audit trails.
- Create one copy of data and use it many times: Each copy, pipeline, and reconciliation cycle uses infrastructure capacity, adds latency, and increases energy demand. A unified data layer that queries data at the source reduces that load and keeps total cost of ownership in check.
























