Why Agentic AI Is Shifting Attention Back to the Data Layer

The rapid improvement of large language models is changing the way financial institutions approach agentic AI. Model capability remains important, alongside the practical challenges of connecting AI to the right data, providing sufficient context, embedding agents into investment workflows and controlling what they can access and do.

Agentic AI was characterised during the session as an evolution from the standalone LLM. A model can first be grounded using techniques such as retrieval-augmented generation, then given access to tools that allow it to query databases, retrieve prices or run software. Add guardrails and the ability to plan, execute and verify actions in a loop, and the result begins to resemble an agent rather than simply a conversational interface.

Improving model capabilities alone doesn’t resolve many of the practical obstacles to putting that technology to work.

Building around the model

Asked to identify the binding constraint on AI deployment, the panel pointed towards workflow integration rather than model capacity. Financial firms need to combine internal information with external third-party data, provide sufficient context for agents to interpret it correctly, select appropriate models and tools, and connect the resulting output with existing investment processes.

An architecture designed around today’s models may also need to accommodate substantially different capabilities six or twelve months from now, putting a premium on flexibility in the surrounding infrastructure.

Data owners increasingly have a role in providing the context surrounding their information, rather than leaving clients to reconstruct it themselves. Semantic and ontological layers can help explain relationships within datasets and allow AI systems to understand what particular data points represent.

That closely mirrors a recurring theme around making financial data ready for agentic AI: access to information alone has limited value if machines cannot reliably interpret its meaning and relationships.

Bringing investment workflows together

The panel explored how a common data foundation could support workflows that have traditionally looked very different. A discretionary portfolio manager might want an AI system to monitor a particular qualitative characteristic in an earnings call. A systematic researcher could take the same observation and ask whether it has historically predicted subsequent returns across thousands of securities.

Supporting both requires more than an LLM. Clean point-in-time data, properly mapped entities and semantic layers become part of the infrastructure through which investment questions can be translated into repeatable analysis. The same foundation could allow discretionary investors to quantify their investment theses while enabling systematic researchers to capture and test ideas originating from qualitative research.

The growing availability of alternative data adds further possibilities. AI gives investment teams the ability to process far greater volumes and varieties of information, with the quality, structure and context of the underlying data determining how effectively it can be incorporated into research.

Deciding what to own

Agentic AI is sharpening decisions over which parts of the investment technology stack firms should build themselves.

The panel drew a distinction between capabilities closely connected with proprietary investment decision-making and infrastructure that can more readily be sourced externally. Internal knowledge bases, feature construction, signals and bespoke investment processes are areas where firms may want to retain control.

Other components can potentially be handed to specialist providers, particularly where they consume significant resources without contributing directly to investment differentiation. Long-standing data-management problems including security masters, concordance and entity resolution remain unresolved at many firms and don’t disappear simply because better AI models become available.

A hybrid approach combines internal development, vendor partnerships and core external infrastructure. This was also the unanimous choice in an audience poll conducted during the session on the AI strategy expected to dominate financial services.

Governance moves beyond hallucination

Hallucination remains a concern, but agentic systems also raise questions around data provenance, authentication, entitlements and operational control. An agent connected to internal datasets and external tools can potentially act on information at considerable speed, making control over what it can access and execute increasingly important.

Analysts and other employees without conventional software-development backgrounds can now construct sophisticated workflows themselves, creating potential vulnerabilities around credentials, API keys and access permissions. Governance therefore extends beyond validating model outputs to controlling the wider environment in which agents operate.

Measuring what changes

Summarising large numbers of earnings calls or reducing the time required for routine research tasks can produce easily measurable productivity gains. Those measures provide a less complete picture of whether investment research itself has improved.

The panel suggested measures such as whether an analyst can cover a larger universe, whether investment conviction or hit rates improve, or whether AI identifies risks and opportunities that existing processes would have missed. One example discussed during the session involved using an integrated research tool following the US tariff announcements of April 2025 to assess exposures across portfolio holdings in minutes rather than the weeks that a manual exercise could have required.

Frontier models will continue to improve, while individual model capabilities may become easier for competing providers to replicate. Financial institutions still need reliable data, sufficient context around that data, and infrastructure governing how agents can use it.