Private markets have a problem that cannot be solved by more information.
The marginal cost of information retrieval has fallen to roughly zero. Generative AI has flooded the diligence surface with synthetic polish. The signal-to-noise ratio has collapsed, and the tools built for the old ratio no longer scale.
Anybody running a deal team already feels the shape of it. Diligence is getting more detailed, not less. Secondaries compress the clock to the point of absurdity: a portfolio of 150 companies arrives, a response is wanted inside 24 hours, and the underlying portfolio companies are frequently not available to ask. Supply chain exposure has walked into the diligence room and sat down. LPs increasingly want the evidence trail behind a model output, not just the output. And the same fund now has to report several different ways to several different investors, applying bespoke exclusion criteria at a scale that no longer fits in an analyst’s week.
The instinct is to point probabilistic AI at the gap. The instinct is wrong.
The trap: probabilistic AI makes it worse
Generative AI deployed as a diligence tool does not close the gap. It introduces a new systemic risk on top of the old one.
An LLM is a probabilistic engine optimised for fluency and prediction. Pointed at a cap table or a stress test, it will produce the plausible-sounding answer, and plausible-sounding is precisely what a fatal flaw looks like on the way past. These systems are optimised for user satisfaction, which is not the same objective as being right.
They are built to give you the most likely answer. The median. The consensus. In investing, the median answer is not much of an answer. And when you ask something the model genuinely cannot establish, it answers anyway — same fluency, same confidence, same tone it uses when it is completely correct. It cannot say “do not trust me on this one.” Withholding the next token is not the thing it was built to do.
This is not a bug. It is structural. You cannot audit a statistical guess. You cannot explain a neural network’s weights to an investment committee, and you certainly cannot explain them to a judge.
The market has learned this at pilot scale. Through 2026, deal teams have been walking back their reliance on unsupervised AI conclusions — hallucinated citations that survived into IC memos, answers that read equally assured whether they were right or wrong, and an accountability gap that becomes very concrete the first time a missed change-of-control clause surfaces post-close.
The most dangerous person in an AI-enabled firm is not the one refusing to use it. It is the one using it for everything, including the moments where confidence has to be earned rather than generated.
The alternative: a diligence stack
To scale judgment, you have to move from probabilistic guesswork to deterministic verification. That means separating jobs that are currently collapsed into a single prompt.
When I started askOdin the mandate was narrow: build the verification infrastructure for the next decade of private capital. Not another commodity SaaS. Not a chatbot wrapper. A deterministic compiler for investment logic.
RUNE Protocol™ is the translator. It strips narrative polish and persuasive rhetoric out of unstructured pitch materials and renders what is left as strict business logic. It does not guess. It compiles.
RAVEN Protocol™ is the cross-examination. It triangulates claims across documents and surfaces the contradictions instead of smoothing them over — which matters because in a heterogeneous data room the finding is almost always the disagreement between two files, not the consensus across five.
The architectural mechanics of RAVEN’s triangulation engine are protected under U.S. Provisional Patent No. 63/994,876 and are not publicly disclosed.
JUDGE Protocol™ is the circuit breaker. When a unit economic model violates basic arithmetic, or a cap table presents a terminal governance risk, the runtime halts rather than narrating its way past the problem.
U.S. Prov. Patent No. 64/017,488 | IPOS §34 National Security Clearance (Issued 2026-03-26).
The interesting case is the inverse one. If a deal presents wildly anomalous unit economics that nonetheless obey the rules of business physics, a deterministic engine does not discard it as noise. It marks it as a structural outlier worth underwriting. Determinism is what separates a genuine category creator from a well-told story — a probabilistic system regresses both to the same mean.
The output is the Defensible Audit Log™: a machine-verifiable record of the analysis. Every claim tied to a page and a line. Every finding sourced. Something you can put in front of an investment committee, an LP, or a regulator without having to say “the model felt strongly about it.”
LLMs optimize for persuasion. askOdin compiles for physics.
The seven questions for any AI vendor
In our SSRN working paper we proposed seven questions for any AI vendor handling deal documents. Here is the list, and why each one earns its place.
| Question | Why it matters |
|---|---|
| Where is my data stored? | Determines jurisdiction, and therefore which regulator has a say |
| Who can access it? | Personnel, process, and permission boundaries — including tenant boundaries, not just staff |
| How long is it retained? | Does it fit your own governance policy, or quietly override it? |
| How is it deleted? | The end of the lifecycle is where policies are usually vaguest |
| Is the processing scoped? | Whole documents to a model, or only the fields required? |
| Are there separate abuse-monitoring logs? | The model platform may keep logs your vendor does not control and cannot delete |
| Is the judgment auditable? | Even with the data deleted, is the conclusion reproducible and sourced? |
The last three are the ones most AI vendors cannot answer cleanly. The seventh is the one that decides whether the tool is usable in a regulated process at all: a vendor can retain nothing and still hand you an answer you cannot reproduce, cannot source, and cannot defend.
Then one more, which cuts across all seven:
For each control you just named, is it implemented, or is it written down? A stated policy is not an implemented control, and an implemented control is not an evidenced one. Anyone can produce a document. Ask for the artifact showing the control executed.
The honest assessment
We answered all seven about askOdin in public, including the one where we come off badly. It belongs here too, because an article telling you to interrogate your vendors is worth nothing if it exempts the vendor writing it.
Our 30-day retention ceiling on raw documents is a documented policy. It is not yet an automated control. The purge job that enforces it is scheduled work that lands with our SOC 2 Type I process; today, deletion is something we carry out deliberately rather than a cron job you could point an auditor at.
The shape of that gap matters more than the fact of it. We are in controlled rollout with pilot engagements, and there is no institutional deal file — no LP data, no portfolio company financials — sitting in the system waiting on that job. What is in the system is founder material from Crucible, which is free and live, and the structural record derived from it carries founder names. Those founders are data subjects. They hold erasure rights, and they are owed exactly the control an institution would demand of us later. The job is in build now.
Staff non-access to customer documents is a procedural control, not an architectural one. We have a policy that people do not read your material; we have not built a system that makes reading it impossible. If a vendor tells you their engineers cannot reach your data, ask what enforces that. Sometimes the honest answer is a rule, and a rule is fine — as long as nobody dresses it up as mathematics.
The full set of answers, and which of our controls are implemented versus in progress, lives on the security page and in our note on the retention fork.
What good looks like
The firms that win the next decade will not be the ones holding the most data. They will be the ones with the infrastructure to systematically verify the physics of their deals.
A stack that preserves contradictions instead of resolving them quietly. A process that produces a provable record rather than a convincing story. A runtime where a violated constraint stops the analysis instead of being narrated past.
Private capital AUM now runs well past $13 trillion. Capital scales smoothly; human verification does not. When volume rises, funds fall back on narrative and on the pattern-matching of whoever is in the room, and every firm ends up improvising its own internal logic standard — none of which are auditable across firms, across LPs, or across years. That is the gap this infrastructure exists to close, and it is the same argument we made in The Diligence Crisis, now with the vendor questions attached.
Public markets got a governing standard a century ago. Private capital never did. That is what this layer is for.
The tools of persuasion have been democratised. The premium now sits on the infrastructure of truth.
Venture capital is the last unaudited asset class. askOdin provides the infrastructure to close the gap.
This essay builds on our SSRN working paper, “The Last Mile of AI: Judgment Infrastructure, Defensible Audit Logs, and the End of Information Retrieval” (DOI 10.2139/ssrn.6664200). Related reading: The Diligence Crisis · The Fork in Data Retention · The Memo Is the Decision