Data sourced from Dakota Marketplace, the global LP and GP intelligence platform trusted by thousands of investment professionals. Learn More | Book a Demo
Most firms calling themselves "AI-ready" mean they've turned on a chatbot. At Dakota's 2026 Data Summit in Philadelphia, senior data and distribution operators from firms including PIMCO, Apollo, Blackstone, Lazard, and Nomura made clear that's not what the term means at all.
The panel converged on a specific, three-part framework, and a harder truth underneath it: the AI tool is the least important decision a firm will make. The data underneath it is the only part that's durable.
The Three-Part Framework
Asked what "AI-ready" actually requires, the room landed on three concrete components, not a vague commitment to modernization:
-
Cloud storage with auto-metadata tagging on every file, so unstructured content, emails, notes, recordings, documents, can be found and used, not just stored.
-
A YAML file for every relational dataset that defines basic semantics: what "AUM" means, who counts as "the client," and which definition wins when multiple versions exist across the organization.
-
A skills layer that adds business-specific context so the model understands "we," meaning the firm's departments, internal language, and structure, not just general knowledge pulled from the open internet.
None of the three is about picking a smarter model. All three are about the data the model will draw on.
Why the Order Matters: Trunk Before Branches
The panel's sharpest point was about sequencing. Most firm logic exists as branches: exceptions, workarounds, and team-specific definitions layered on top of each other over years. The work of becoming AI-ready is bringing that logic into a single trunk first, so every team is drawing from the same clean source.
Standardizing definitions matters more than selecting a vendor. A firm can have access to every AI tool on the market and still get unreliable output if "qualified opportunity" means five different things across five teams. The number of tools available doesn't fix data that isn't semantically clean underneath.
The Model Is Replaceable. The Data Is Not.
Firms at the Summit described running ChatGPT and Claude side by side, not because one outperforms the other, but because the two serve different purposes and the cost of running both is trivial next to the upside. The principle underneath that choice: get the data right, and the model becomes swappable.
Clean data, clean connectors, and an agentic layer get a firm roughly 80% of the way to strong output with any major commercial model. That's the practical argument for treating data governance, not model selection, as the real hedge against a stack that changes every quarter. Nobody in the room could predict which tool they'd be running in three months. The firms with clean data underneath don't need to.
Dakota's own framing of this, shared at the Summit, puts the stakes in sharper terms than most firms have priced in:
"A database you chose wrong, you replace at renewal. A model you trained wrong, you carry forward, because it has learned, deeply and durably, from whatever you gave it."
That's the shift from a procurement decision to a training decision. Choosing what data a model learns from isn't an annual vendor evaluation anymore. It's choosing the ceiling on what that system will ever be able to see.
What This Means for Distribution Teams
A few practical implications follow directly from the framework above:
- Answer discipline beats answer confidence. The most disciplined firms at the Summit train their models to return information only from defined sources, and to explicitly decline to answer when the underlying data doesn't support it. A model that fills gaps produces confident wrong answers, which cost a distribution team more than no answer at all.
- Publish what the tool can and can't see. Teams that don't know what data their AI tools have access to can't calibrate how much to trust the output. An internal changelog of covered and uncovered data sources solves this cheaply.
- Start with the business question, not the tool. The most common failure mode the panel named was dashboards built with every available metric and zero adoption, because no one defined what success looked like before building.
- Treat "data" as broader than rows and columns. Emails, files, voice, and meeting notes are data now. Infrastructure built only for structured tables misses most of what a distribution team actually generates.
The Real Question to Ask
Once a firm accepts that AI-readiness is a training decision, one question matters more than any other: what share of the relevant universe, whether that's the LP market, the GP market, or a firm's own client base, is the model actually being trained on?
A model trained on a fraction of the market doesn't give partial answers. It gives complete, confident answers about the slice it can see, and silence on everything outside it. The gap isn't a coverage gap. It's an intelligence gap, and most teams don't know how large theirs is until a competitor's model surfaces something theirs never could.
Book a demo to see how Dakota's daily-refreshed LP, GP, and private company data can serve as the clean foundation underneath your firm's AI stack.
