Why PE Firms Are Moving Past Hebbia to Defensible AI Due Diligence

When Hebbia raised its $130 million Series B in 2024 at a reported $700 million valuation, the coverage was everywhere. The pitch was compelling: an AI platform that could ingest entire data rooms and answer questions the way a top analyst would, but in seconds. For private equity deal teams drowning in VDR files, CIMs, and management presentations, it sounded like exactly what they needed.

Plenty of PE firms took demos. Some purchased licenses. And what many of them discovered over the following months created a pattern that is now showing up regularly in conversations with operating partners: the answers came fast, but the verification work did not go away. In some cases, it got worse.

This article is not a takedown of Hebbia as a company. Their funding validates that the market for AI in diligence is real and growing. The issue is architectural. The category of tools that Hebbia sits in — semantic search layered over a vector store — has fundamental limitations that matter specifically when the work being automated is high-stakes, legally significant, and auditable.

The Conviction Gap: why generic RAG fails diligence

Most enterprise AI tools built in the 2023-2024 wave use Retrieval-Augmented Generation, or RAG, as their core architecture. The pattern is well-understood: chunk documents into 500-token segments, embed those chunks into a vector store, retrieve the most semantically similar chunks when a user asks a question, and pass those chunks to a language model to synthesize an answer.

For many use cases, this works well. Internal knowledge bases, HR policy Q&A, IT helpdesk automation — these are forgiving environments where a slightly wrong answer can be corrected on the next query. Diligence is not a forgiving environment.

The problem with RAG in a VDR context is structural. A purchase agreement for a mid-market acquisition might be 180 pages. A covenant — say, a change-of-control provision with an 18-month escrow holdback — may span multiple sections, cross-reference definitions on page 12 from the operative clause on page 94, and be modified by an amendment rider attached at page 165. Standard RAG chunking will never evaluate all three sections together in a single retrieval pass. The model will return the most semantically similar chunk, which might be the covenant header, but miss the amendment that changes the escrow period entirely.

This is what creates the Conviction Gap. The analyst receives an answer that looks complete and reads confidently. But because they cannot see the exact clause the model is referencing — or confirm that the model evaluated the full document and not just the top-ranked chunk — they have to go back to the source themselves. The AI accelerated the first pass but created a new verification step that did not exist before.

What defensible diligence actually requires

The firms that have solved this problem — and there are not many yet — have converged on three non-negotiable requirements for AI diligence tooling.

First, the AI must evaluate the entire document set, not a retrieved subset. Underlying's Iterative Source Decomposition architecture breaks complex diligence questions into hundreds of micro-queries that run in parallel across every page of the data room simultaneously. The system evaluates page 1 through page 180 of that purchase agreement at the same time, not just the three chunks that scored highest in a vector similarity search. This is the technical foundation that makes covenant extraction actually reliable.

Second, there must be a mandatory human review gate before any AI output enters a deliverable. This is not optional. The only way to guarantee that a hallucination does not end up in a signed IC memo or a due diligence report that goes to an LP is to force a human to approve every extracted fact before it flows downstream. Underlying's UI presents every finding with source-linked evidence, and the approval workflow is enforced in the product — analysts cannot skip it.

Third, every decision — both AI-generated and human-approved — must be logged in an immutable audit trail. When a deal closes and something is contested 18 months later, the firm needs to be able to produce a complete record of what the AI extracted, what the reviewer approved, and when. Without this, the firm has no defensible record of its review process.

Hebbia vs. Underlying: a structural comparison

The differences are not about feature counts or UI polish. They are architectural, and they matter at the exact moment when a deal gets complicated.

On source linking: Hebbia returns answers that reference documents generally. Underlying hyperlinks every number and claim directly to the source paragraph and page — click any generated text to view the highlighted PDF clause.

On hallucination risk: Hebbia's RAG architecture produces black-box answers with no enforced verification step before the output reaches the analyst's working document. Underlying's Mandatory Review Gate forces the analyst to approve or flag every extracted finding before it enters the deliverable.

On audit trails: Hebbia does not create an immutable record of what the AI extracted versus what a human reviewed. Underlying logs every decision with timestamps, reviewer identity, and source citations — exportable for legal review.

On data security: Underlying processes all documents ephemerally — ingested into secure, tenant-isolated memory streams, analyzed, and wiped once the workflow completes. No document caches are stored on system infrastructure. Underlying also supports Bring Your Own Key, where API credentials are stored in the client's own AWS Secrets Manager.

How PE operating partners use Underlying in a real diligence window

A mid-market buyout team faced a compressed 48-hour diligence window covering 140-plus vendor MSAs, management schedules, and commercial slides. The deal included a complex supplier concentration risk profile that had not surfaced in the initial commercial diligence.

Parallel ISD agents parsed all change-of-control provisions across the full document set simultaneously. The system flagged three hidden liabilities containing buyout exit fees buried in supplier MSA riders — provisions that would have been missed entirely in a manual first-pass review given the time constraint. Each flag was presented in the review gate UI with the exact clause highlighted in the source document. The deal team approved, flagged, or requested additional review on each item before any finding entered the diligence memo.

The outcome was not just speed — it was completeness. The team had a defensible record that every contract was reviewed, every change-of-control clause was evaluated, and every finding had a human sign-off.

What this means if you are currently evaluating Hebbia

Hebbia's pricing is opaque — enterprise deals require contacting their sales team directly, which makes it difficult to benchmark value before committing. More importantly, the core architectural constraints described above are not features that can be patched in a future update. They reflect a fundamental design philosophy: Hebbia is built for broad document search and exploration. Underlying is built specifically for the work that gets audited.

If your firm's diligence outputs end up in signed closing documents, LP reports, or regulatory filings, the Conviction Gap is not a minor inconvenience — it is a liability. The verification work that a black-box AI creates does not disappear. It transfers from the AI to your analysts, and it accumulates as deal complexity increases.

Dotnitron deploys Underlying for PE and advisory teams that need diligence automation which is fast and defensible. The 30-day pilot model lets your deal team run actual transactions through the platform before committing — with a full refund if the system does not remove the bottleneck it was deployed to solve.