Due Diligence
How to Test Due Diligence Platform Scalability: 5 Questions to Ask
Five questions to ask before selecting a due diligence platform, or to pressure-test the one you're already using.
Why Reliable AI in Due Diligence Depends on a Structured Data Layer
Reliable AI in due diligence depends on more than the model. A structured data layer gives agents clean, queryable data points across your whole portfolio.
Dasseti
Jun 22, 2026
Allocators are under growing pressure to put AI to work on manager due diligence, and the natural first move is to point it straight at the documents on file: compliance manuals, completed DDQs, and policies. The assumption is that a capable enough model can read those documents and answer questions about them as well as a person would.
Whether that works depends less on the model than on how the underlying data is held. Any AI agent can only hold so much information in front of it at one time, so pointed at a stack of source documents, it reasons over the few it manages to take in and nothing beyond them. Hold the same information as structured data instead, with the key responses and figures pulled out of those documents and stored as discrete data points the tool can query, and it can work across every manager you oversee in a single pass.
That intermediate layer, the structured data layer, is what decides whether AI gives you answers you can rely on.
Why AI gives unreliable results on raw due diligence data
AI tools can only reason over the documents held in context at once, and a million-token context window comes to about two thousand pages. That sounds like room to spare until you measure it against a live program - a single compliance manual can run two or three hundred pages, and that is only one document for one manager. Once you add all the supporting files the context window can fill after a handful of managers, so the agent works from whatever slice happens to fit.
That slice is where the trouble starts. You have no reliable way to know which documents the agent read and which it never reached, so any answer it returns rests on partial context rather than the full picture. It might miss the one manager whose valuation policy lapsed, or treat a superseded document as current. The output reads as finished, but you cannot put your name to it, because the data underneath was incomplete before the model started reasoning. Even a strong model returns a weak answer here, because it never had the whole record in front of it.
What a structured data layer changes for due diligence AI agents
The difference comes from holding the data in a structured layer that sits between the documents and the agent. Rather than leaving the answers buried in files, the platform keeps them as firm-level and fund-level profiles, with the responses from a questionnaire and the figures from a long report extracted into structured data the agent can read directly.
Now the agent does not have to open a three-hundred-page manual to find one policy. It queries a single data point. That structured data takes a fraction of the space the source files do, so all your managers stay within reach, and the agent works from data you have already reviewed and approved.
What a structured data layer lets a due diligence team do
With data structured this way, the questions due diligence teams actually ask become ones an AI tool can answer across all your managers in one pass.
Compliance manuals and policies:
- Surface which managers are missing a current valuation or business continuity policy across the whole portfolio
- Compare a manager's stated policy against the answer they gave in the DDQ, using policy-to-response comparison
- Find where a manual contradicts the version a manager submitted the year before
Completed DDQs:
- Pull every manager who changed a material answer since the last cycle, in a single query
- Run a red flag scan by asset class without opening each questionnaire on its own
- Line up how different managers answered the same question, side by side
Fund terms and track record:
- Pull the funds in a given strategy that fall within a fund-size and fee range, then line up their terms across managers
- Compare a fund against the prior vintage you reviewed, on terms, strategy, and track record, rather than opening both research documents and reading them line by line
Which due diligence teams will get AI to work
The teams that hold their due diligence data in a structured layer now are the ones who will get something usable out of the AI tools their firms are steering them toward. The teams aiming those same tools at fragmented documents will keep getting answers they cannot trust, because there is no structured layer underneath for the model to reason over.
See how Dasseti COLLECT turns your due diligence documents into a structured data layer.