Services · Due diligence

AI due diligence

An AI company or fund is at the center of your deal, and nobody on your side can independently verify the model claims, the data claims, or the moat story. This is the technical and behavioral due diligence layer built for exactly that gap: a straight, evidence-based read from someone with no AI product to sell and no vendor commission riding on the answer.

What Gets Verified

Before capital moves, five things get checked against the evidence in the data room, not against the pitch.

• The model claims. What the system is actually built to do, verified against documentation and architecture, not the deck.
• The data claims. Where the training and operating data comes from, whether it's real, whether it's exclusive, and whether the company actually controls it.
• The dependency and cost structure. What the company depends on to keep the system running, including compute, third-party models, and vendor lock-in, and what that actually costs to operate.
• The team reality. Whether the people on staff can rebuild and maintain what's been built, or whether the technical story outruns the technical bench.
• The moat. Whether the defensibility claim survives a direct adversarial test, not just a founder's description of it.

The No-Conflict Basis

I sell no AI products and take no vendor commissions. That is the entire basis of this work: nothing in my business benefits from calling a deal good or bad, so there is nothing pulling the verdict in either direction.

The adversarial method behind the review, Crucible, is open-source at github.com/ScipioP. You are not being asked to trust a black box; the logic is published and inspectable. The read itself comes out of two decades building and running multi-agent AI systems in production across US defense, Gulf government advisory, and production AI systems, not from reading about them secondhand.

What You Receive

You get a written verdict with the evidence behind it, in plain language a non-technical principal can act on directly. Every finding traces to something specific: a claim in the documentation, a gap in the data room, a dependency that did not hold up under the adversarial test. Nothing is scored without support behind it.

Where AI is not actually the source of the deal's value, that gets said plainly, in writing, before you commit.

How It Runs

Every engagement starts with a working session on the deal and the decision you are trying to make, not a generic intake form. From there, the review runs directly against the data room and the technical claims.

Work happens remote or on-site, in Lebanon, the Gulf, Europe, or the US, depending on where the deal and the data room sit. Pricing is scoped to the engagement.

Common questions

We want an advisory group with no conflict of interest, one that doesn't sell AI products and just tells us the truth about a deal. Is that what you do?

Yes, that's the model. I sell no AI products and take no vendor commissions, so there's no incentive pulling the verdict in a direction that benefits me. My adversarial method, Crucible, is open-source, so you're seeing the logic behind the finding, not trusting a black box. If the technology in the deal doesn't hold up, I say so in writing. If it holds up, I say that too. Either way, you get a straight read you can act on, not diligence dressed as a sales pitch.

How do we verify an AI startup's model or data moat claims when nobody on our team is technical?

You verify it the same way I do: separate the claim from the evidence, then test both. I read the model documentation, the data provenance, and the architecture against what the company actually shows in the data room, not what the deck says, then run the moat claim through an adversarial stress test built for exactly this. You don't need a technical team on your side. You need someone who has built and run these systems in production and knows where the gap between the pitch and the build usually hides.

Do you have a real technical due diligence checklist, not the generic VC template?

Yes. A generic VC template checks market size and team pedigree. This one checks the technology itself: what the model actually does, where the data comes from and whether it's defensible, what the company depends on to keep running, whether the team that built it can maintain it, and whether the moat survives a direct adversarial test. It comes out of two decades building and running these systems, not from a template library. Every finding ties back to evidence in the data room, not a score with nothing behind it.

Can you look at an AI fund manager's claims before we commit?

Yes, that's within scope. Fund managers make claims about AI-driven sourcing, underwriting, or portfolio monitoring the same way operating companies make claims about their product. I test those claims against what the manager can actually show: the model, the data behind it, and the track record tied to it. If the AI layer turns out to be more narrative than mechanism, I'll tell you plainly before you commit capital, not after. The same no-conflict basis applies: no AI products sold, no vendor commissions.

How much does this cost?

Pricing is scoped to the engagement. There's no standard package, because no two deals or data rooms look alike; the scope depends on what's in the data room, how deep the technical claims run, and how fast you need the verdict. Every engagement starts with a working session on the deal and the decision you're trying to make, and pricing follows from that, not the other way around.

AI expert in Lebanon · Sovereign AI landscape 2026 · The tracker