The Flywheel Part 2: The Demand Side Just Changed

Editor's Note: Re-published with permission from the Insight Innovation Ventures' Substack, this article outlines a framework for investors evaluating companies in the AI era—posing key questions that resonate differently for every market participant.

Part 2 of a three-part series on the flywheel model in insights and analytics. Part 1 defined the mechanism and gave a ten-question audit for telling a real loop from a slide. This post is about what a real loop is worth commercially — and why it shows up first in how a company gets found, sold to, and kept.

A flywheel that nobody can find is a wasted asset. The demand side of this industry just went through a structural change that determines whether a real loop ever gets the chance to compound: who does the finding, and on what evidence.

The Buyer Moved Before the Seller Noticed

Gartner’s projection that AI agents will intermediate 90% of B2B purchasing by 2028, carrying more than $15 trillion in spend, sounded aggressive when it was published and looks conservative now. 6sense’s Buyer Experience Report, based on nearly 4,000 B2B buyers, found that 94-95% of buying groups rank their preferred vendor before any contact with a sales rep, and that the pre-contact favorite wins the deal roughly 80-84% of the time. More than half of B2B buyers now start vendor research inside an AI tool rather than a search engine.

None of that describes the future. It describes the current baseline. By the time a research or analytics buyer picks up the phone, the shortlist is usually already built, and it was built by a process that queries structured, indexed, machine-readable evidence rather than a homepage or a sales deck.

Discovery Became a Retrieval Problem

The practical consequence is that demand generation has quietly become a retrieval-engineering problem. An agent building a vendor shortlist is querying citation graphs, structured documentation, and indexed third-party evidence — the same infrastructure that powers AI search and AI Overviews — rather than crawling landing pages for keyword density. A supplier with excellent positioning language and no machine-parseable evidence behind it is invisible to exactly the process now doing the shortlisting.

This is the same disaggregation we have already tracked at the product layer: the research value chain splitting into ownable layers — data, reasoning, delivery, decision — with each layer exposed as an independently queryable endpoint rather than bundled behind a login and a sales team. What’s changing now is that the same logic is arriving at the top of the funnel. If the product is headless and queryable, the sales motion has to be too, or the product never gets into the query in the first place.

Why This Is Good News for a Real Flywheel and Bad News for a Fake One

This shift rewards exactly the kind of company Part 1 described as having a genuine loop, and it punishes exactly the kind that doesn’t — for a specific, mechanical reason.

An agent assembling a shortlist is not persuadable the way a human buyer is. It doesn’t respond to tone, brand story, or a confident claim on a landing page. It aggregates verifiable signal: documented methodology, disclosed provenance, published calibration or accuracy records, structured technical specifications, and third-party evidence that can be checked. A company that has done the unglamorous work described in Part 1 — disclosing its rights structure, publishing a calibration track record, exposing its retraining cadence — has, without necessarily intending to, built exactly the evidence trail an agent is built to find and weight favorably. A company whose “flywheel” is a slide has nothing for the agent to retrieve, because there was never anything to disclose.

Put differently: the discipline of proving a loop is real and the discipline of being discoverable by an agent are now converging on the same underlying asset — machine-readable, verifiable, disclosed evidence. A company that built the former for defensibility reasons gets the latter as a byproduct. A company that skipped it gets filtered out of the shortlist before a human ever sees the deal.

MCP Is Distribution, Not a Moat — But Absence Is Now a Penalty

Model Context Protocol adoption inside this industry has moved fast enough that it is no longer a signal of sophistication — it is close to a baseline expectation. Research-platform MCP servers now expose panel access, moderated-interview execution, and cross-study retrieval as callable tools an agent can invoke directly, without a dashboard login or a manual export. Analytics and BI platforms have followed the same pattern, republishing governed data as agent-queryable endpoints rather than destinations a human has to visit.

Part 1’s caution still applies in full: an MCP endpoint is a channel, not a moat. It does nothing on its own to make a company’s underlying loop harder to copy. But the commercial calculus has shifted regardless, because the absence of that channel is now a filtering event rather than a neutral choice. An agent cannot shortlist a supplier it cannot query. Treat agent-native distribution as table stakes, the same way a functioning website became table stakes twenty years ago — necessary for consideration, worth nothing by itself once every competitor has it too.

Why a Closed Loop Turns Consulting-Grade Work Into Something That Behaves Like ARR

The harder and more interesting question is what a genuine decision-outcome loop — the rarest of the three loops from Part 1 — does to the economics of highly customized, judgment-heavy engagements that look nothing like a subscription on the surface.

The mechanism runs through switching cost, not through packaging. A supplier with a closed outcome loop is not just delivering a recommendation; it is accumulating a client-specific calibration record — a disclosed, improving history of which recommendations worked, under what conditions, with what confidence. A competitor pitching the same client starts that record at zero. The incumbent’s advantage is not the relationship or the brand. It is a compounding, auditable track record that a new vendor cannot manufacture on day one at any price. That is a switching cost with nothing to do with contract lock-in, and it is precisely the kind of moat that survives even when a competitor has equal or better raw data and an equally capable team.

This also changes what the client is renewing. In a project model, the client renews a relationship and a reputation. In a decision-outcome flywheel, the client renews access to a rising accuracy curve that is measurably, disclosedly getting better specifically because of the history the two parties have built together. Losing the vendor doesn’t just mean re-procuring a similar service. It means abandoning a demonstrated, improving calibration record for an unproven one. That is a retention dynamic that behaves like net revenue retention in a subscription business, even though the underlying engagement is bespoke, advisory, and priced like consulting.

The Overlay That Makes This Real Rather Than Aspirational

None of this works without the same disclosure discipline that makes a flywheel legible to an agent doing vendor discovery. A performance or outcome-linked pricing component only functions as a retention mechanism if the calibration record behind it is actually published, actually falsifiable, and actually tied to a defined cadence — the same requirements from the audit in Part 1, applied to the commercial layer instead of the technical one. A firm that claims a rising accuracy curve without showing the curve is making the identical unearned claim as a firm that calls a static dataset a flywheel. The market is now sophisticated enough, on both the human-buyer side and the agent-mediated side, to tell the difference inside a single evaluation cycle.

Where This Leaves the Series

Part 1 established what a flywheel is and gave the test for a real one. This post has been about what happens once a real loop exists: it gets found differently, it gets bought differently, and it turns even the most bespoke, judgment-heavy engagement into something that retains and expands like recurring revenue rather than resetting with every new project. The next post turns the audit from Part 1 into a full diligence framework — the specific questions worth asking before you accept anyone’s flywheel claim, including your own, as something other than a well-told story.