Picture the last slide of a twelve-week pilot review. An AI agent has read four thousand supplier invoices from last spring, matched them to purchase orders, and flagged the exceptions with an accuracy the slide puts at 94 percent. The executive who sponsored it is delighted, and says so. Then the manager who runs accounts payable asks three questions. Which system will it read from in the live month, not the sample? Who looks at the six percent it gets wrong, and how fast, during the five days of the close? And who gets the call when it stops working at eleven at night? The room goes quiet. Nobody in it has answered those questions, because the pilot was never built to answer them.
The scene is invented, but the stall it describes is not. In early October the consultancy BearingPoint reported that only 13 percent of the companies it surveyed were on track with their AI initiatives, and fewer than a third had got beyond pilot projects, as Reuters reported the study.1 A week and a half later the data and AI firm Further, which sells help with exactly this problem, published a survey of 515 American AI decision-makers. In it, 37 percent said they had paused, scaled back or abandoned an underperforming AI initiative in the past year.2 A paper from the Futurum Group, built on survey data from 600 enterprises that build their own AI applications, found that 53 percent still could not point to a return.3 These are different surveys with different questions, and none of them is disinterested. They point the same way.
My argument is that much of this stall is a buyer problem, and that Geoffrey Moore described it more than thirty years ago. A pilot needs two approvals that look alike and are not. The first is the sponsor's yes: an executive who wants the future and will tolerate a rough tool to reach it. The second is the owner's yes: the people who will run the work every day and live with the results. Most pilots are designed to win the first, and the first does not produce the second. It often gets in the way of it.
A qualification before going further. I am not claiming that sponsors are wrong, that every stalled project failed for this reason, or that Moore's model is a law of nature. Some pilots stall because the technology does not work. The claim is narrower. When the technology works and the project still stalls, look at who said yes.
Two kinds of yes
Moore was a technology marketing consultant, and Crossing the Chasm, first published in 1991 and revised most recently in 2014, was written for companies selling new technology to businesses.4 He took the familiar adoption curve, with innovators and early adopters at the front, then the early and late majorities, then the laggards, and argued that one of its joins was not smooth. Between the early adopters and the early majority sits a gap. Companies that thrive selling to the first group fall into it when they try to sell to the second, because the two groups are buying different things.
The early adopters, whom Moore called visionaries, buy a break with the past. They want a technology that lets them leap ahead of rivals, they expect to be first, and they will put up with missing pieces and bugs to get there. The early majority, his pragmatists, want something else: an improvement to operations they already run, a product that fits the systems they already have, and one that works when it arrives. Above all, pragmatists want to know that people like them, in their own industry, are already using it successfully.
That last point creates the chasm. The natural reference for a new product is its first customers, and its first customers are visionaries. Moore argued that visionaries make poor references for pragmatists. To a pragmatist, the visionary undervalues the experience of colleagues, cares more about the technology than about the industry, overlooks the systems already in place, and leaves a mess behind for others to maintain. So the vendor arrives with its proudest stories, which are about early test sites, and the pragmatist wanted to hear about installations that are up and running in production. Each side hears the other as missing the point.
How far to trust the chasm
A fair reader should know what kind of book this is. Crossing the Chasm is a practitioner's framework, built from consulting experience and case stories rather than tested data. It has been enormously influential, which is not the same as being right.
The best test of it I know came from three marketing scholars, Jacob Goldenberg, Barak Libai and Eitan Muller. In 2002 they examined sales of a large number of new consumer electronics products and found that between a third and a half followed a particular shape: an early peak, then a trough too deep and long to be noise, then a recovery that rose above the first peak.5 They called it a saddle, and explained it the way Moore would: an early market and a main market, adopting at different speeds. They then modeled the conditions under which a saddle appears, and found that the decisive variable was how much the two markets talked to each other. When communication between them was low, more than half of the simulated products showed a saddle. When it approached the level of communication within each market, fewer than 5 percent did.
So the chasm is common but not universal, and its depth is not fixed. It depends on whether the second group of buyers hears about the product from people they trust. That is the lever this essay is about.
The chasm inside the company
Moore was writing about markets. But move the picture inside a single firm and it fits enterprise AI uncomfortably well. The internal visionaries are easy to name: the chief executive who has asked every division for an AI plan, the innovation team, the enthusiasts who were using these tools before anyone approved them. They are indispensable. Without them nothing starts. They also tend to sponsor pilots in the way Moore's visionaries buy products, for the leap rather than the fit, and they are quite willing to work around the rough edges themselves.
The internal pragmatists are the heads of operations, the controllers, the risk and compliance teams, and the people who run the systems and answer the phone at night. They are not opposed to AI. They are responsible for the work it would change, and they know they will be the ones explaining its errors. Their question is not whether the tool is impressive. It is whether the work will still get done, on time and correctly, in a bad week.
Both groups can say yes to a pilot, and the yeses mean different things. The sponsor's yes is a yes to the possibility. The owner's yes is a yes to Monday morning. The Futurum paper, which assigns each of the failures it describes to an owner, has a sentence that could stand as this essay's summary: "Ownership rarely survives the departure of the executive sponsor."3 A pilot that has only the first yes is held up by one person's enthusiasm, and enthusiasm moves on.
Most pilot reviews end here, with the sponsor's stamp.
The owner's list keeps printing until someone names who runs it.
Test sites and production installations
Moore's sharpest observation is about evidence. Vendors crossing the chasm keep presenting the evidence that won the visionaries, and it does not move the pragmatists, who want proof of a different kind from a different source.
The same mismatch runs through most internal pilot reviews. The deck reports accuracy on a test set, hours saved in a sample, delighted early users. That is evidence about the technology, and it answers the sponsor's question. The owner's question is about operations: has anyone like us run this through a real month, with real exceptions, on the systems we actually use? A test-set score cannot answer it. Neither can the sponsor's enthusiasm, however sincere, because the sponsor is not the one who will carry the risk.
Moore's point about references also explains why the obvious remedy, a bigger launch, rarely works. Pragmatists trust people who share their problem. In his telling they communicate within their own industries, so a reference from a bank moves another bank and not a retailer. Inside a company the same rule holds by function. The reference that moves the finance team is another finance team. A glowing report from the marketing department, or from the innovation lab, is evidence from the wrong market.
The whole product, inside the firm
Moore's second big idea, which he credited to the Harvard marketing professor Theodore Levitt, is the whole product. There is a gap between what a customer is promised and what the vendor actually ships. Visionaries fill the gap themselves, because they enjoy the work and want the advantage. Pragmatists will not. For them the product is what ships plus everything else they need to get the result that made them want it.4
Inside a company, what ships is the model, the agent or the vendor's platform. Everything else is the long blue list on the receipt: data that is ready to use, connections to the systems the work already runs on, controls and an audit trail, training, support, and a named person who is accountable when it goes wrong. The October surveys read like inventories of missing pieces. In the Futurum data the most cited obstacles were a lack of skilled people, at 52 percent, data quality and availability, at 51 percent, and integration with legacy systems, at 44 percent, and the same three had led every survey wave for eighteen months.3 In BearingPoint's study, about 40 percent named legal regulation as the main barrier to scaling and 34 percent named integration with existing systems.1 In Further's survey only 5 percent believed their data was largely ready for AI, and among the firms that had pulled back an initiative, data quality was the reason cited most often.2
None of those pieces is a model problem, and none of them shows up in a pilot designed to win a sponsor. A pilot can run on a clean extract, a handful of volunteers and an engineer who fixes things by hand, and still produce a convincing slide. That is fine for the sponsor's yes. It is the very thing that makes the slide useless to an owner, because the owner knows the extract, the volunteers and the engineer will not be there in production.
A different essay here argued that new software needs the work around the machine to be funded as a program. Moore's version adds something specific. The whole product is not a general list of complements. It is defined by one buyer's reason to buy, so it cannot be specified until you know whose yes you are trying to earn. Vendors have noticed. Several now send engineers to work inside a customer's own teams; the Futurum paper describes these programs and, to its credit, advises buyers to ask what capability stays behind when the engineers leave.3 That is the whole-product question in contract form.
One beachhead
Moore's prescription for crossing the chasm is military. Like the Allies on D-Day, a company should pick one beachhead, a single segment of pragmatist customers, and commit everything to winning it completely before moving on. Completely means the whole product for that segment, so that its customers become references for the next. Neighboring segments then fall, in his image, like bowling pins, because each one hears from a peer.4
Inside a company the beachhead is a workflow with an owner who already has a compelling reason to change it: a backlog they measure, a close that runs late, an error rate they report. Pick one. Ask its owner what they would need to run it on a bad Monday, and build all of that, not just the model. Then make the owner, not the sponsor, the reference, and point the next team at them. That last step is the lever the saddle study found. Figure 2 shows a model of it.
The model is crude, but its lesson is the one Goldenberg and his colleagues drew. The trough is not caused by the technology getting worse. It is caused by the first group running out while the second has not yet heard from anyone it trusts. Shorten that silence and the trough fills in.
An earlier essay argued that a pilot needs a way to be wrong, a result written down in advance that could stop it. That is about whether a pilot can teach you anything. This is about who has to believe what it teaches. A well-designed pilot can still win the wrong buyer.
Who blinks first
Moore was frank that pragmatists would rather wait. Waiting is rational for them: someone else pays for the debugging, and the references accumulate. He argued that the vendor's job is to make waiting the riskier choice, and that the way to do it is not pressure but the whole product and the references that a finished beachhead produces.
Inside a company there is a temptation to skip that work, because the sponsor has a tool that vendors lack: authority. A mandate can make a team use a tool. It cannot make the team believe in it, and the difference shows up in workarounds, shadow spreadsheets, and the quiet reversion that follows the sponsor's next reassignment. The owner's yes has to be earned in the owner's terms.
Objections, considered
"Mandates do work. Plenty of systems were adopted because someone senior insisted." True, and some mandates are right. But a mandate settles whether a tool is used, not whether the work is done well with it. The systems that were successfully mandated usually arrived with their whole product: a finished integration, training, support and a team that owned them. The mandate was the last step, not the first.
"AI is a general-purpose technology, not a product for one segment." It is general in what it can do, but it is adopted one workflow at a time, by particular people with particular systems and risks.
"These numbers come from firms that sell the cure." They do. BearingPoint is a consultancy, and I have used its figures only as Reuters reported them. Further sells data and AI services. The Futurum paper carries Snowflake's company profile and describes Snowflake's own embedded engineers. The surveys measure different things in different populations, and they should not be added together or compared directly. The argument here does not rest on any one figure. It rests on the agreement in their direction, and on the obstacles they name, which are what Moore would predict.
"Moore wrote about markets. An internal buyer cannot walk away." That is the most serious objection, and it limits the analogy. An internal owner usually cannot choose a different vendor, and budgets are set from above. But the mechanism that makes the chasm, pragmatists trusting the experience of peers who share their risk, does not depend on there being a market. It depends on people being accountable for results. Those people exist inside every firm, and they can always comply without committing.
The sponsor's yes starts a pilot. Only the owner's yes finishes one.
Questions before the next round of pilots
- Name the owner before the pilot. Who will run this work every day if it succeeds? If nobody can say, the pilot can only win a sponsor.
- Ask the owner's questions at the first review, not the last. Which systems, which exceptions, which bad week? Put the answers on the slide next to the accuracy score.
- Pick one beachhead. Five half-finished pilots produce five sponsors and no references. One finished workflow produces a reference that the next team will believe.
- Count the whole product. List everything the owner needs besides the model, and fund it, or admit the pilot is a demonstration.
- Build the reference on purpose. Let the next team hear from the owner who runs it, in their own terms, rather than from the sponsor who championed it.
Back to the review
The accounts payable manager in the opening scene was not being obstructive. Those were the second buyer's questions, and they were the right ones. A pilot that cannot answer them has not failed. It has simply won the wrong buyer first, and stopped.
The way across is not a louder sponsor. It is one workflow, finished, and an owner willing to tell a colleague that it works.