Research Methodologies

October 28, 2026

7 min read

The Flywheel Part 3: The Diligence Framework

The Flywheel Part 3: The Diligence Framework

What makes an AI flywheel claim credible? Explore the due diligence questions investors, acquirers, and research leaders should ask.

Editor's Note:  This article first ran on the Insight Innovation Ventures Substack and appears here with permission. It was written for investors sizing up AI-era companies, and the questions it asks land differently depending on which side of the table you sit on.


Part 3 of a three-part series on the flywheel model in insights and analytics. Part 1 defined the mechanism and the four-factor test for a real loop. Part 2 covered what a real loop is worth once buyers and agents can see it. This post turns the framework into a diligence process — the specific questions worth asking before anyone accepts a flywheel claim as real.

Regulators, acquirers, and LPs have all arrived at the same conclusion from different directions this year: an unsubstantiated AI claim is no longer a marketing exaggeration. It is now a legal, financial, and deal-structure liability, and the burden of proof has shifted onto the company making the claim.

The Market Stopped Taking the Claim on Faith

The SEC’s Cyber and Emerging Technologies Unit has spent the past two years building a body of “AI washing” enforcement precedent — cases against Delphia, Global Predictions, Presto Automation, and Nate Inc., among others — all resting on the same underlying theory: representations about AI capability are material to investors, and overstating them is securities fraud regardless of whether AI is the company’s core product or a claimed feature. The founder of Nate Inc. was charged criminally alongside the civil case, for raising more than $42 million on the claim that an AI shopping app processed transactions automatically when the work was, in fact, being done manually by outsourced human labor. Securities litigation tied to AI claims is on pace to double in 2026.

Stopped Taking Claims on Faith

Private equity has converged on the identical instinct from the buy side, arriving at it through diligence practice rather than enforcement. A recent survey found that roughly 80% of investment professionals had personally flagged overstated AI claims inside a CIM. Grant Thornton’s current AI-diligence practice tests every claim across the same four dimensions on nearly every deal now: whether AI is embedded in the product in a way that drives pricing power and retention, whether the claimed advantage is demonstrable rather than aspirational, whether the underlying data infrastructure is actually clean and owned, and whether governance exists before it’s forced by a buyer.

Independent AI-diligence consultancies have built out formal six-question protocols — testing data ownership, model dependency versus wrapper status, switching costs, regulatory exposure, talent concentration, and whether the claim maps to real P&L rather than a demo — specifically because deal teams reading a data room cannot reliably tell a moat from a wrapper on their own.

Convergence of Scrutiny

None of these three constituencies — regulators, acquirers, or LPs — is applying a novel standard. They are applying ordinary scrutiny to a claim that spent two years getting a free pass because nobody had yet built the vocabulary to test it. That vocabulary now exists, and it maps almost exactly onto the mechanism this series has spent two posts establishing.

The Diligence Protocol

The ten-question audit from Part 1 was built to be asked by anyone evaluating their own claim. This version is built to be asked by someone else — an acquirer, an investor, a client procurement team, or an LP — who has no incentive to accept a comfortable answer. The structure runs in four stages, each one closing off a different way a fake flywheel survives a superficial look.

4 Stage Diligence Protocol

Stage one: Does the input actually belong to the company? Request the specific contractual chain proving the company owns, rather than licenses, aggregates, or scrapes, the data behind the claim. A wrapper on a third-party model with rented or borrowed data is not a flywheel claim — it’s a subscription with better branding, and the diligence question that surfaces this fastest is simple: what breaks if the underlying vendor changes price, changes policy, or gets acquired ? If the honest answer is “most of the product,” the company does not have the asset it’s claiming to have.

Stage 1: Ownership

Stage two: Does new data actually improve the system, on evidence rather than assertion? Ask for the calibration or accuracy record itself, not a description of one. A company with a real loop can produce a time series showing the metric moving — error rate falling, confidence scores tightening, coverage expanding — tied to a specific retraining or re-validation cadence it can name. A company without one will describe the mechanism in the present tense and produce nothing dated. This is the single fastest way to separate the two populations, and it is exactly the gap PE diligence teams report finding in roughly four out of five CIMs that lead with an AI claim.

Stage 2: Evidence

Stage three: Is the loop legally defensible, not just technically real? A technically real loop built on data the company cannot prove it has the right to use, retain, or train on is a liability wearing a moat’s clothing. This is where the diligence question intersects directly with the regulatory environment: can the company produce consent documentation and provenance records at the level of the individual record, and has it disclosed — accurately — what it trained on and under what terms? Every enforcement case brought so far, from Delphia through Nate Inc., has rested on the gap between what was claimed and what could be substantiated when asked directly. A diligence process that doesn’t ask the company to substantiate its provenance claim in writing is not testing the claim; it’s recording it.

Stage 3: Defensibility

Stage four: Would the loop survive the company losing its current team or its current vendor? Talent concentration is the diligence risk buyers consistently underweight. A flywheel that depends on two or three specific people who built the retraining pipeline and hold the tribal knowledge of what the data actually means is not institutional — it’s a key-person dependency dressed up as infrastructure, and it evaporates on exactly the timeline that matters most to a buyer: the first eighteen months after close.

Stage 4: Durability

Reading the Answers Together

None of these four stages is diagnostic alone. A company can pass stage one and fail stage three — genuinely own its data but have no defensible provenance trail, which converts a real technical asset into acquired legal risk. A company can pass stages one through three and fail stage four — a real, well-governed, well-evidenced loop that only exists because of two irreplaceable engineers. The point of running all four is that a flywheel claim is a conjunction, not a single fact: ownership and evidence and legal defensibility and institutional durability, together, or the claim doesn’t hold at the price being asked.

The Conjunction Equation

This is also where the four-factor equation from Part 1 earns its place in a live diligence process rather than staying a framing device. Unique signal, rights and trust, workflow embedding, and closed-loop learning were presented there as a multiplication because any one factor near zero collapses the whole claim — and the four-stage protocol above is simply that equation operationalized into four things a deal team can actually go verify with documents, dated metrics, and direct questions to the engineering team, rather than four things a management deck can assert.

What This Means in Practice

Operationalizing the Theory

For an acquirer or investor, the practical implication is that AI-diligence workstreams belong inside standard technology and commercial diligence on every deal where a flywheel claim materially supports the valuation — not as an optional add-on triggered by an unusually AI-forward pitch, but as a default line item, the same way quality-of-earnings review became a default rather than an exception. For a company making the claim, the implication runs the other way: the evidentiary bar has already moved, whether or not the company has noticed. A flywheel claim that can’t produce a dated calibration record, a documented rights chain, and evidence of institutional (not individual) capability is now a diligence finding waiting to be written up, not a differentiator waiting to be believed.

What This Means in Practice

The Series, Closed

Three posts, one throughline. Part 1 established that a flywheel is a specific, testable mechanism — not a synonym for having data and using AI — and gave the four-factor equation and ten-question audit for telling a real one from a slide. Part 2 showed what changes commercially once a real loop exists: it gets found differently by agent-mediated discovery, and it converts even bespoke, judgment-heavy engagements into something that retains like recurring revenue because the client is renewing a compounding, unreplicable track record rather than a relationship. Part 3 has shown that the market has independently arrived at the same skepticism from three directions — securities regulators, PE diligence teams, and AI-diligence specialists — and built a four-stage protocol for testing the claim under adversarial rather than friendly scrutiny.

The industry spent the last two years rewarding anyone who could describe a flywheel convincingly. That window is closing. What gets rewarded next is the much smaller set of companies that can prove one.

Comments

Comments are moderated to ensure respect towards the author and to prevent spam or self-promotion. Your comment may be edited, rejected, or approved based on these criteria. By commenting, you accept these terms and take responsibility for your contributions.

Insight Innovation Ventures

Insight Innovation Ventures

3 articles

author bio

Disclaimer

The views, opinions, data, and methodologies expressed above are those of the contributor(s) and do not necessarily reflect or represent the official policies, positions, or beliefs of Greenbook.

More from Insight Innovation Ventures

The Flywheel Part 2: The Demand Side Just Changed
Artificial Intelligence and Machine Learning

The Flywheel Part 2: The Demand Side Just Changed

How is AI changing research supplier discovery? Explore why evidence and visibility now determine whether a growth flywheel can compound.

The Flywheel Part 1: What Actually Compounds in the Insights Industry
Artificial Intelligence and Machine Learning

The Flywheel Part 1: What Actually Compounds in the Insights Industry

What creates a durable AI advantage? Explore why insights companies need more than proprietary data to build a true flywheel.

ARTICLES

Challenges for Smaller Service-Led Suppliers
Research Methodologies

Challenges for Smaller Service-Led Suppliers

Each firm may have its own formula for success, but pressure to conform may be increasing

Nelson Whipple

Nelson Whipple

GRIT Senior Research Director at Greenbook

The Neighborhood Never Got Safer, So We Built Bigger Fences
Research Methodologies

The Neighborhood Never Got Safer, So We Built Bigger Fences

Why is market research fighting data fraud one company at a time? Explore the need for industry-wide solutions to protect research quality.

Karine Pepin

Karine Pepin

Co-Founder at The Research Heads

How to Choose a Market Research Firm: A Buyer’s Checklist
Research Methodologies

How to Choose a Market Research Firm: A Buyer’s Checklist

Learn what to ask a market research firm, which red flags to watch for, and how to compare proposals before choosing a partner.

Ashley Shedlock

Ashley Shedlock

Content Producer at Greenbook

Only the Attentive Survive
Research Methodologies

Only the Attentive Survive

Two decades of Research Industry Trends Reports continue to help researchers navigate change

Nelson Whipple

Nelson Whipple

GRIT Senior Research Director at Greenbook

Sign Up for
Updates

Get content that matters, written by top insights industry experts, delivered right to your inbox.

67k+ subscribers