The Data Foundation AI Agents Need

Photo of author

A system of record that supports your investment process matters more than the AI model you choose

Request a demo →

TL;DR: Every fund can rent the same frontier models, but they are not all getting the same value from AI. What makes an AI agent trusted by an investment team is the foundation it runs on: a governed system of record organized around how the fund operates. Over time that foundation becomes the fund's institutional memory, a unique asset that turns a system of record into a system of reasoning for the whole firm. Anthropic, OpenAI, Salesforce, and Gartner point to a similar conclusion, and the foundation should come before any decision to build or buy AI tools.

Most investment teams are now deciding how to put AI agents to work inside their funds. An analyst can connect Claude or ChatGPT to a few data sources and have a working agent by the afternoon, and many firms have started building their own applications and artifacts, encouraged by how quickly an AI coding assistant turns an idea into a prototype.

However, a standalone artifact in an analyst's chat history is a long way from an institutional system. What investment teams need is a context-aware agent that securely accesses client data, informs position sizing, and helps an investment committee answer questions about the portfolio on the fly. It has to be accurate, properly permissioned for the right data, auditable, and supported for years. Nearly all of those requirements are decided by the layer beneath the agent rather than by the agent itself. Before a fund decides whether to build or buy its next AI tool, it should ask a more basic question: does it have the right foundation for that tool to run on?

Every fund essentially rents the same frontier models

A multibillion-dollar hedge fund and a retail investor analyzing a position can use a frontier model from one of the three leading labs and work with closely comparable raw capabilities, so access to the world’s best models alone does not provide an edge. As we argued in How AI Is Transforming Investment Research, durable edge still comes from proprietary research, your institutional knowledge, and process discipline.

No matter how much it knows of the outside world, a general-purpose model cannot know how your fund thinks and operates without the proper context. What strategy are you running, and how do you think about sizing, risk, and diversification? Capturing that takes much more than one very long prompt, and since most funds use broadly similar external data (like market and alternative data), the difference is in how that data is organized around your own portfolio and the record of decisions that built it. That record is the fund's institutional memory and one of its most valuable assets: what the team believed, what it did, why it did it, and how each investment played out. With that context, an agent understands how your fund invests rather than how the average investor would.

In addition, general LLMs tend to make things up, or hallucinate, when they try to answer a question they have little information on. Ask a frontier model why your team bought NVDA and it will confidently reconstruct your thesis from market commentary, public analyst notes, and price action instead of your history of investment memos, similar themes in your book, and the sell-side research note in your email. When models hallucinate, they lose the trust of the investment team, and worse, they can lead the team to conclude that AI is just “not there yet.”

“A good foundation is a system of record organized around your specific investment process. Before you even take the step to build an app, ask whether you have the right foundation.” - Greg McCall, Co-Founder & President

This is why AI agents need a solid foundation before they can be useful to an investment team, and we are not the only ones who think so.

The companies building AI have reached a similar conclusion

When Anthropic introduced the Model Context Protocol in 2024, it wrote that even the most sophisticated models are “constrained by their isolation from data,” and designed MCP to connect them to the systems where data lives. OpenAI's Frontier platform, launched in February 2026, is built around a business context layer that links systems of record so agents work from the same information people do. In August, Anthropic and Salesforce expanded their partnership so Claude can reason over the customer data, workflows, and business rules companies have spent decades building in Salesforce, and Marc Benioff summed up the logic: “Probabilistic intelligence alone doesn't run a company, and deterministic systems don't reason.” Andrew Ng has made the same case for years with a simpler line: “data is food for AI.”

The research on adoption agrees. Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. Asset managers say the same about their own industry: in FE fundinfo's 2026 survey, which polled 200 senior asset managers in the UK, Luxembourg, and Switzerland, 99% use AI in some form, yet 64% said it will only deliver meaningful value once the quality and structure of their underlying data improve.

What a good foundation looks like for an investment team

For a fundamental manager, the foundation is a system of record built around the way the fund generates ideas, constructs portfolios, and manages risk. It also codifies the process, so the fund can repeat what works and catch its blind spots. In practice it has four traits.

It holds all the data needed to make investment decisions. First, it needs internal data such as your notes, models, price targets, positions, exposures, attribution, and the dated history of each thesis. Next, all of your external data such as your alt data sources, investor presentations, broker research, expert networks, and management call transcripts have to be centralized in one place. On this foundation, an agent can reconstruct the thinking behind why a position is sized the way it is. If a PM asks why the team added to NVDA in March, the agent answers from the note, the price target, the broker research and call transcripts the team was reading, and the exposure as of that day. Our guide to research management systems covers this in more depth.

It speaks the fund's language. When a PM asks for YTD attribution, the system already knows which portfolio, benchmark, period, and factor model that means, so the agent answers the question that was asked instead of guessing at definitions.

It knows who is asking. Entitlements by role and team are part of the architecture, and an agent acting for a user inherits exactly that user's permissions. Managers who worry about IP leaking into AI raised the same concern about the cloud a decade ago, and the answer is the same: keep the data in a governed system, whether you run it yourself or use a trusted partner, and let the model see only what each user may see.

It calculates deterministically. Numbers come from verified analytics with a traceable calculation path, and the model retrieves, explains, and synthesizes them. The same question returns the same numbers inside the platform or from Claude, ChatGPT, or an internal agent.

Foundation first, then build versus buy

With tools this capable, the first instinct at many funds is to build, and for a small internal utility that is reasonable. A mission-critical investment system is a different project: making it scalable and deterministic, earning the trust of PMs and the risk committee, and then evolving and supporting it as markets change takes sustained investment long after the prototype works. Building an app is also a different job from building the system of record it depends on, which is the logic behind the Anthropic and Salesforce partnership. Every firm has to decide whether its core business is managing money or developing software, and all things being equal, buying is faster, cheaper, and less risky than building.

That decision is easier once the foundation is defined. Our next two articles take on build versus buy directly, first for the research and portfolio analytics system, then for the AI interface on top of it.

How EDS approaches the foundation

EDS was built by investors and engineers as an investment system of record, on a platform in use at fundamental funds since 2012. It brings research, portfolio construction, factor risk, and attribution together and configures them to each firm's methodology, which is what we mean by investment process management. Fusion AI and the EDS MCP layer sit on that foundation, so an analyst working in Fusion AI, Claude, or ChatGPT gets the same governed, entitlement-aware numbers, because the AI synthesizes while EDS calculates. Our AI earnings preview walkthrough shows it in a real workflow.

Whatever AI tool you evaluate next, ours included, start by looking at the foundation it will run on.

Request a demo →

Ready to Strengthen Your Investment Process?

Indicates required field
Name