For venture, private equity, and corporate development

AI due diligence: is the AI real, and what does it depend on?

Every target now arrives with an AI story. Some are products, some are pilots, and some are a slide. We match your deal team with an engineer who has shipped the kind of system the target says it has, for a short, fixed-scope technical read, so you know which one you are underwriting.

Tell us about the deal Start with the 48-hour report

Why this needs its own read

Conventional technology diligence is good at what it was built for: architecture, code quality, security, scale. It was not built to tell you whether a model is doing the work or a person is, whether a claimed accuracy figure would survive a realistic test, or what happens to gross margin when a model vendor changes its prices.

Those are now the questions that decide whether an AI company is worth what it says. The gap between claim and system even has a name, AI washing, and regulators have started bringing cases over it. What a read like this usually turns up is less dramatic than fraud and more common: a real product with one irreplaceable engineer, an unmeasured quality claim, or a cost structure that belongs to someone else.

We see the pattern from the outside every day. Of the 47 New York AI companies whose job boards we read each morning, 37 are hiring people to build on AI agents and 12 are hiring anyone to fine-tune a model. Most AI companies are application companies. That is a perfectly good business, and it is valued differently from one that owns its models, which is why it matters that the pitch and the payroll agree.

How it works

Hiring signal report

Before you commit · 48 hours · no access needed

An outside-in read built from the target’s own public job board: what they hire for, which models and tools they name, whether they recruit people to build models or to build on them, how long roles stay open, and how posted pay compares with the market. It ends with the questions to put to management. It needs nothing from the target, so you can have it before the first management meeting.

The read

Confirmatory diligence · 5 to 10 business days

A specialist who has shipped the kind of system in question works through the six areas below with the target’s technical team: architecture walkthrough, code and evaluation review where access allows, vendor contracts, and interviews with the people who built it. You get a written report, a red-flag summary on one page, and a call to go through it. The report is the specialist’s own professional opinion and carries their name.

The first hundred days

After close · optional

The findings become a plan: what to fix first, which roles to hire, and what to stop. If you want help executing it, the same network supplies the engineers, fractional leaders, and consultancies to do the work.

The signal report draws on the same pipeline as our public NYC AI Hiring Index and AI skills glossary, pointed at a single company. It works for any company that posts roles through Greenhouse, Ashby, or Lever, wherever it is based.

What the read covers

What is actually in production

The deck says AI-powered. We establish what that means today: which features run on models, for how many customers, at what volume, and which parts are a pilot, a roadmap item, or a person doing the work by hand behind an interface.

What it depends on

Which models and vendors the product rests on, on what contract terms, and what a price change, a deprecation, or a policy change would do to margins. Whether the company has the rights it needs to the data it trains or grounds on. How much of cost of goods sold is model spend, and how that moves with usage.

Who built it, and what happens if they leave

Usually the most valuable finding. We look at who holds the knowledge, how much is written down, and whether the team could rebuild or extend the system without its one or two key people. We judge engineers for a living, so this is the part we are best placed to do.

How they know it works

Whether output quality is measured or assumed: evaluation sets, regression testing, monitoring in production, and a record of failures caught before customers found them. A team that cannot show you its evals is telling you something.

What it costs to keep

The realistic engineering and infrastructure cost of maintaining the system for the hold period, the technical debt that will come due, and what it would take to replace a component if a vendor relationship ended.

Regulatory and governance exposure

Where the product makes or informs decisions about people, which rules apply and what has been documented. We flag the exposure; your counsel advises on it.

What you receive: a written report organized by those six areas, a one-page summary of red flags ranked by what they could cost you, the questions we could not get answered, and a call with the specialist. If the read finds nothing wrong, the report says so in a paragraph. The scope and fee are fixed up front, so nobody has a reason to pad the findings.

Independence and conflicts

We publish data on AI companies, place people into them, and assess them for investors. A diligence opinion is only worth having if those activities cannot touch each other, so the rules are written down. They are part of our wider disclosures.

We step aside where we have a stake

We will not arrange a read on a company where Neuronify, its principals, or a fund affiliated with them is an investor, a bidder, or a lender, or where we hold an active search or placement mandate. If that is the case we tell you before the engagement letter, and we decline.

The specialist is independent too

The engineer who does the work confirms in writing that they hold no financial interest in the target or its direct competitors and have not worked for either in the past two years.

What we learn stays with you

Every read is under NDA. The target’s name and anything learned in diligence never enter our publishing, our data, or our conversations with other clients. We do not arrange reads for two bidders on the same deal.

Our publishing does not bend

Our public data on New York AI companies is built by rule from the companies’ own job boards. A company being a diligence subject, a client’s portfolio company, or neither has no effect on how it is listed or counted.

Who it is for

Venture funds without a technical partner for AI, or with one who cannot be expert in every kind of system a pipeline throws up. Private equity firms underwriting an AI narrative in a software or services target, where the question is often how much of the claimed efficiency is real; see also AI partners for private equity. Corporate development teams acquiring for capability, where the asset is the team and the diligence is mostly about whether it stays and whether it can repeat what it built.

Buying AI rather than investing in it? The same questions apply to a vendor. Our guide to choosing an AI consulting firm and the project scoping pack cover that side.

AI due diligence: common questions

What is AI due diligence?

A technical assessment of a company’s AI claims, carried out for an investor or acquirer before a deal closes. It establishes what is genuinely in production, what the product depends on, who built it, how its quality is measured, what it will cost to maintain, and what regulatory exposure it carries. It sits alongside financial, legal, and commercial diligence.

How is this different from ordinary technical due diligence?

General technology diligence reviews architecture, code quality, security, and scalability, and those still matter. AI adds questions a generalist reviewer is rarely equipped to answer: whether a model is doing the work or a person is, whether performance claims survive a realistic evaluation, how exposed margins are to a model vendor’s pricing, and whether the training data was the company’s to use. We cover the AI questions and work alongside your existing technology diligence provider.

Who does the work?

A specialist from our vetted network who has built and operated the kind of system the target describes, matched to the deal: a retrieval system is read by someone who has run one in production, a voice product by someone who has shipped voice. Every specialist has at least two production AI deployments we have verified. The report is their professional opinion and carries their name. Our part is the match: verifying their production work, confirming their independence from the target, and scoping the read with you.

How long does it take, and what does it cost?

The hiring signal report takes 48 hours. The read takes five to ten business days from the first technical session, depending on access. The fee is fixed and agreed before work starts, in a written proposal that states the specialist’s fee and our match fee separately. Diligence reads sit at the lower end of our engagements, which start around $25k.

Can you work within a live deal timeline?

Yes. Diligence runs on deal timelines and the read is scoped for them. Tell us the timeline when you first get in touch. If we cannot staff the right specialist inside it, we will say so the same day rather than accept the work and compress the quality.

Do you assess early-stage companies with little in production?

Yes, with a different emphasis. At seed and Series A there is less system to inspect, so the read weighs the team more heavily: whether the founders and early engineers have built this kind of thing before, whether the technical plan is plausible at the budget, and whether the evaluation approach would tell them if it was not working.

Is the report investment advice?

No. It is the specialist’s technical opinion, for the use of your deal team alongside your financial, legal, and commercial diligence. It does not recommend whether to invest or at what price, and it may not be relied on by third parties without the specialist’s written agreement.

Have a deal in front of you?

We take a small number of reads at a time. Tell us the company, the timeline, and what the AI story is, and we will come back the same day with whether we can staff it, who would do the read, and whether any conflict rules us out.

Tell us about the deal Talk to us first