for PE / VC deal teams

AI claims survive pitch meetings. Most do not survive an audit.

MIT's NANDA initiative measured it: roughly 95% of enterprise AI pilots never touch P&L. When a target says AI, your committee has one question. Which side of that number is this company on? A deck cannot answer that. Reading the actual system can.

I open the codebase, turn every AI claim into something falsifiable, and deliver verdicts your committee can defend.

01 / the offer

Technical AI Diligence

A fixed-fee engagement scoped to one deal or one portfolio company, run pre-close or inside the first 100 days post-close.

For the deal team

An AI thesis with a build, kill, or rescue verdict on every initiative: what is actually wired to P&L, and what is theater.

For the company CEO

A 100-day plan: stop this, fix that, build this first. Written to be executed, not filed.

I get repo access and time with whoever built the system. You get a claims file that stays useful after the wire clears: reopen it at any board meeting and re-check every claim. Fixed fee, scoped before work starts, priced in the intake conversation.

02 / operator proof

Both sides of this table.

PE rollup, operated as co-CEO

Co-CEO of a PE-backed rollup: $1.5M platform acquisition to roughly $54M in enterprise value. I have been the buyer, and I have lived the operating cadence your portfolio companies run on.

Healthcare SaaS, built end to end

As fractional AI co-founder, I built a multi-tenant, HIPAA-aware post-acute care coordination SaaS from first commit to production: architecture, agents, evals, compliance. Shipped software in a regulated domain, not a slide deck about one.

A live auditable research system

130 tickers under continuous AI coverage in my own research system. Every claim sits in a falsifiable claims file with sources. Drift monitoring flags when the world changes. A locked eval harness scores the system, not my memory of it. Your diligence runs on the same method.

03 / methodology

The rubric is published.

Inspect the method before you pay for it. Hold the deliverable to it afterward.

Claims inventory
Every AI capability the company asserts, restated as a falsifiable claim with a named source.
P&L wiring
Which claims connect to revenue or cost in the data, and which are demo theater.
Code-level verification
The actual repositories, read directly: architecture, vendor exposure, what breaks at scale.
Eval discipline
Whether outputs are measured against a locked evaluation set, or nobody can say if the system works.
Data and workflow readiness
Whether the data layer and operating workflows can carry the AI roadmap being sold.
Drift and dependency exposure
What happens when models, vendors, or prices change, and whether anyone would notice.

Findings land as an auditable claims file, not a deck: every claim falsifiable, sourced, and re-checkable long after the engagement ends.

Sample claims-file entry (illustrative format; the full PRBot teardown replaces this)

Claim: "The recommendation engine drives a third of revenue," per the CIM.
Evidence: Order attribution over the trailing 90 days: 4.1% of completed checkouts touched a recommendation surface.
Falsified by: Re-run the attribution query at close and at day 100; the claim fails if attributed revenue stays under 10%.
Verdict: Kill the claim, rescue the pipeline. The revenue claim is theater. The event pipeline underneath it is sound and feeds two items in the 100-day plan.

Want the method at full depth? Read the sample teardown: my own product, PRBot, run through this exact rubric, warts included.

Read the PRBot teardown

04 / expansion path

Where this goes next

Some funds keep me on after diligence as an AI Operating Partner across the portfolio. That engagement is application-only, and it opens after a diligence engagement, not before. By then we are not negotiating trust. The claims file already earned it.

05 / start

Start a diligence conversation

This door is for fund-side readers: operating partners, deal teams, independent sponsors. Building your own company? Your door is the main site. Head to willmitchell.com