In June 2026, KPMG pulled a flagship AI report after an outside investigation found that only 5 of its 45 citations accurately pointed to real sources. The rest were fabricated, distorted, misattributed, or too vague to verify, including detailed case studies about UBS, Swiss Federal Railways, and Transport for London that those organizations disputed as inaccurate. KPMG's own website, at the same time, was running an article on how strong AI guardrails help you "scale AI faster."

Weeks later, IBM lost roughly 25% of its market value in a single trading day, the steepest one-day decline in the company's history, close to $69 billion. CEO Arvind Krishna's explanation to investors was blunt: the company "faltered."

Neither company lacks money or talent. IBM invented more of modern computing than almost any company alive, and KPMG's entire business model is verification. But they didn't fail at the same thing, and the difference matters. IBM missed a strategic shift, clients reallocating budget toward AI infrastructure faster than IBM adapted its own positioning. KPMG missed a verification step: nobody checked the citations against their sources before the report went out under the firm's name.

Two different failures, same underlying cause. Not the technology itself, but the operating discipline required to run it well. That gap between what a company can afford to buy and what it actually knows how to run is the real story of enterprise AI in 2025 and 2026. MIT's Project NANDA examined more than 300 enterprise AI initiatives, representing an estimated $30 to $40 billion in spend, and found the overwhelming majority failed to produce measurable business impact. The detail that matters more than that headline number is who converted the rest of it into results. Large enterprises ran the most pilots and had the lowest rate of scaling one into production, while strong mid-market performers moved from pilot to full implementation in roughly 90 days, against nine months or longer at large enterprises.

Size didn't buy an advantage here. In several of these stories it bought a more expensive way to fail. What follows is organized by failure mode rather than industry, because the same patterns show up in industries that otherwise have nothing to do with each other.

Failure mode one: strategy that didn't survive the shift it was supposed to plan for

The clearest data point here is a pair of surveys. Writer, an enterprise generative AI platform vendor, published a 2025 study of 1,600 knowledge workers, 800 of them C-suite, finding companies with a formal AI strategy report success at more than double the rate of companies without one, 80% versus 37%. Info-Tech Research Group's 2026 survey of 551 senior leaders landed on almost exactly the same ratio through a different sample: 60% of enterprises with a governed AI strategy reported measurable impact, versus 20% without one.

IBM's July drop fits here more precisely than the "AI failure" headlines suggested at the time. Krishna's own account wasn't that IBM's AI underperformed. It was that the company didn't adapt quickly enough as customers shifted spend toward AI infrastructure and away from the software and services IBM sells, which is a strategic-adaptation failure, not a botched rollout. The distinction matters: IBM's stock didn't fall because a project failed, it fell because leadership didn't see a market shift coming and didn't have a plan ready when it hit. That's still a strategy failure, just a different one than the KPMG or Deloitte stories below, and treating all of these as interchangeable would be the same kind of sloppiness this piece is arguing against.

Failure mode two: verification that failed at firms whose entire job is verification

This is the category that should unsettle a professional-services client the most, because it hit the professionals hardest, and it happened four times at the same tier of firm within about a year.

KPMG's citation failure wasn't isolated. Deloitte delivered an independent review to the Australian government, contract value A$440,000 (roughly US$290,000), that cited academics who confirmed they never wrote the cited work and quoted a court ruling that didn't exist. An outside university researcher caught it before Deloitte's own process did, and the firm agreed to refund part of the contract, roughly A$97,000. EY Canada pulled a report after researchers found fabricated data and a citation to a McKinsey report that was never published. PwC Middle East had reports from 2024 through 2026 flagged for fabricated citations and broken references.

Saying none of these firms had any verification process at all would overstate the evidence. What the record actually supports is narrower and just as damning: whatever verification controls existed didn't catch material fabrications before publication, at firms that sell verification as their core product, one of which, Deloitte, employs more than 460,000 people worldwide. That decision, whether a human checks output against a source before it ships, is exactly as available to a twelve-person firm as it is to a firm with 460,000 employees, arguably more available, since a smaller firm has fewer places for the gap to hide.

The legal industry shows the same failure at higher stakes. Researchers have now identified more than 1,000 court filings containing fabricated AI citations, while broader international databases tracking AI hallucinations in legal proceedings report still higher totals. Two prominent cases involved firms nobody would call under-resourced. In Lacey v. State Farm, Ellis George LLP and K&L Gates were ordered to pay $31,100 after roughly 9 of 27 citations in a brief were found to be incorrect in some way, including at least two nonexistent authorities and several fabricated or inaccurate quotations. The underlying research involved CoCounsel, Westlaw Precision, and Google Gemini. In Wadsworth v. Walmart, lawyers representing the plaintiffs, including two attorneys from Morgan & Morgan, cited nine cases in motions filed with the court. Eight did not exist.

Failure mode three: speed replacing judgment at the exact moment judgment was the job

Some decisions were never supposed to be fast. Insurance claims and customer commitments are two of the clearest examples, and several companies used AI to make them faster instead of better.

Cigna's PxDx system, a computer-assisted claims-review system rather than a generative AI product, was used by Cigna physicians to deny payment for more than 300,000 claims over a two-month span in 2022, averaging 1.2 seconds per case, according to ProPublica. Cigna has disputed ProPublica's characterization of the system, saying PxDx uses software rather than AI or an algorithm to match procedure and diagnosis codes. UnitedHealth faces a separate class action alleging that more than 90% of claim denials associated with its nH Predict system were reversed on appeal, a figure alleged by the plaintiffs rather than independently established. On February 13, 2025, a federal judge dismissed five of the case's seven claims but allowed breach-of-contract and implied-covenant claims to proceed.

Klarna ran a comparable experiment in customer service and later rebalanced its approach. In early 2024, the company said its AI assistant was handling two-thirds of customer-service chats and performing work equivalent to roughly 700 full-time agents, a statistic that became one of the most widely cited examples of AI replacing knowledge work at scale. By 2025 Klarna was bringing human customer-service workers back into the operation. CEO Sebastian Siemiatkowski told Bloomberg that cost had become too dominant a factor and that the result was lower-quality service. Klarna did not abandon AI. It moved toward a model designed to preserve human support when customers wanted or needed it.

Air Canada's chatbot failure was smaller in dollar terms, about C$812 including damages, interest and tribunal fees, but important as an early AI-liability case. The bot gave a customer incorrect information about bereavement fares. After he relied on that information and Air Canada refused his refund request, the airline argued that it should not be responsible for information supplied by the chatbot. British Columbia's Civil Resolution Tribunal rejected that argument, finding that the chatbot was part of Air Canada's website and that Air Canada was responsible for the information it provided. The decision is not binding precedent for every company or jurisdiction, but it has become a frequently cited example of the legal risks surrounding customer-facing AI.

McDonald's ended a roughly three-year automated drive-through ordering trial with IBM in 2024 after the technology generated mixed results and numerous customer complaints, some captured in viral videos showing incorrect or bizarre orders. McDonald's did not abandon voice AI. The company said the experiment had reinforced its belief that voice ordering would eventually play a role in its restaurants, while ending that particular implementation.

In another widely shared incident, users manipulated a ChatGPT-powered chatbot on a Chevrolet dealership's website into agreeing to sell a new Tahoe for $1 and describing the exchange as a legally binding offer. No $1 sale took place. The incident exposed a more general control problem: a customer-facing model had been deployed without sufficient protection against users manipulating its instructions and eliciting statements far outside its intended sales-support role.

Failure mode four: scale assumed to be a moat, and mostly wasn't

JLL's 2025 Global Real Estate Technology Survey found that 88% of real estate investors, owners and landlords had begun piloting AI, while only 5% reported achieving all of their AI goals. Salesforce's Agentforce rollout offers a less definitive but useful caution. By 2026, reporting described the platform as still early in its adoption curve, with customers citing data-preparation burdens and limitations in more complex interactions, even as Salesforce reported rapidly growing Agentforce revenue. The evidence here is more mixed than in the other cases, so treat it as an early deployment warning rather than a confirmed failure.

The sharpest data point in this category comes from Sinch, and it deserves a more careful read than the topline number suggests. Sinch found that 74% of surveyed enterprises had rolled back or shut down at least one live AI customer-communications agent after a governance failure. Among organizations with fully mature guardrails, that figure rose to 81%. Read quickly, that looks like evidence governance does not work. Sinch argues almost the opposite: organizations with stronger controls identify problems sooner and stop problematic deployments faster. Governance was never going to prevent every failure. Its value is in detecting problems early, limiting damage, and creating a process for correction or shutdown.

That also reframes the MIT finding from the opening. Large enterprises ran the most pilots but recorded the lowest pilot-to-scale conversion rates, while top-performing mid-market companies moved from pilot to full implementation in roughly 90 days. MIT's researchers found that successful implementations tended to focus tightly on specific business problems and execute against them deliberately. Bigger budgets and larger AI teams did not automatically translate into better conversion.

What this actually means if you're not IBM or KPMG

The cynical version of this takeaway, that big companies are bad at AI and smaller ones will win by default, doesn't hold up against the evidence here. Several of the failures above involve companies that moved fast and still lost. Klarna reversed course on its own. McDonald's ended its own pilot. Moving fast without discipline fails at any size, and a smaller company chasing the same shortcut would likely fail the same way, just with a smaller headline attached.

The narrower and more defensible version is this: none of the failure modes above are really about company size. IBM's was strategic adaptation. KPMG's, Deloitte's, EY's, and PwC's were verification. Klarna's was scope and judgment. The legal cases were accountability. Sinch's data is about monitoring, not prevention. None of that shows up as a line item on a budget, it's a set of decisions about how a company operates, and a company with 40 employees and a clearly defined problem can build the verification step and the governance guardrail that a company with tens of thousands of employees somehow shipped a $290,000 report without.

That's the actual confidence gap, and it isn't between large companies and small ones. It's the gap between AI ambition and operational readiness and controls, between what a company is confident enough to claim in public and what it actually verified before making the claim. Closing it comes down to defining the business problem, the success metric, the governance, and the verification step before the pilot starts rather than after the report ships. None of that requires IBM's balance sheet. It just requires someone deciding to do it before the deadline forces the shortcut.

A note on sourcing, given the subject: every claim above is drawn from public company statements, court and tribunal filings, regulatory disclosures, or reporting from named news organizations, cross-checked against primary sources where available. Figures described as allegations are labeled as such rather than presented as established fact.

Pick a side. That's what the discourse demands.

Half your feed says AI is the biggest thing since electricity, about to rewrite every industry and every job description. The other half says it's a bubble running on venture capital and CFO anxiety, headed for the same graveyard as every overhyped technology before it. Both sides cite real data. Both sides sound certain.

Both sides are asking the wrong question.

"Does this technology work" and "will this technology work for your company" are two different questions. The public debate is stuck on the first one. Your business only cares about the second.

The capability is real. That argument is over.

Strip out the hype and the backlash and look at what's left. Large language models draft documents, summarize mountains of text, write working code, and handle a real share of routine customer interactions. Five years ago that was science fiction, and even then, only a Fortune 500 with a nine-figure technology budget could have chased it.

Is the capability uneven? Yes. Performance varies by task, by model, by how well the system around it was built. Some industries need a level of reliability the technology can't promise yet. Anyone who tells you it all just works is selling something. But anyone who tells you there's nothing there is arguing with evidence sitting in plain sight.

Here's what actually changed, and why I keep saying AI is what I call "the great equalizer." The capability that used to require a Fortune 500 balance sheet, a data science team, and a multi-year integration project is now within reach of a company with a hundred employees. Or forty. Or fifteen.

That's the point most coverage misses. The equalizer isn't a mid-market story. It's a true small-company story: the 30-person agency, the 60-person distributor, the 90-person law firm. Companies under 100 employees were never going to build a data science team, never going to fund a multi-year integration project, never going to get a straight answer from an enterprise vendor. They had no seat at this table. Now they do. That's my read of the market, not a settled fact, so let me be precise about it. Budget still matters: data preparation, integration, security, compliance, talent. None of that got free. But for a growing list of use cases, access to a capable model is no longer the barrier. For the first time in my thirty years around enterprise technology, the gap between a big company's capability and a small one's is less about budget and more about strategy.

And here's the part most small-business leaders miss: you may have the faster path, not the slower one. A company under 1,000 employees doesn't have six committees, a procurement gauntlet, or twenty years of legacy systems standing between a decision and a deployment. Under 100 employees, the advantage gets sharper still: the owner is the approval chain. A decision made Monday morning can be running by Friday. Big enterprises can't move like that at any budget. That speed advantage is real, and most of the companies holding it aren't using it, because they still assume transformative technology happens to bigger companies first.

One catch. The equalizer only works if you treat AI as a strategy question. Treat it as a shopping trip and you get the same result every technology wave has always delivered: ERP, CRM, cloud. Forty years of capable technology, and in every wave, some companies transformed while others spent millions to get a slightly more expensive version of what they already had. The technology being real has never guaranteed the outcome being good.

Now the uncomfortable numbers

This is where the pipe-dream crowd has a point. Not about the technology. About what companies are doing with it.

RAND Corporation studied AI project failure in 2024 by interviewing 65 experienced data scientists and machine-learning engineers. RAND notes that by some outside estimates, more than 80% of AI projects fail, roughly twice the failure rate of ordinary IT projects. RAND didn't produce that number; they cite it. What RAND produced is more interesting, and I'll get to it in a second.

Gartner projected in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. The actual number came in worse. By 2026, Gartner was reporting it had reached at least 50%, and their reasons read like a coroner's report: poor data quality, inadequate risk controls, escalating costs, unclear business value.

S&P Global Market Intelligence found the same pattern from a different angle. Companies abandoning the majority of their AI initiatives before production: 17% in 2024, 42% in 2025. On average, organizations scrapped 46% of their proofs of concept before anything reached production.

Those numbers don't say AI lacks business value. They say capability and deployment are two different things, and the gap between them is getting wider while the models keep getting better. Sit with that for a second, because it should bother you.

The cause isn't what you think

Back to RAND, because their real finding is the one that matters. 84% of the practitioners they interviewed pointed to at least one leadership-driven root cause as a primary reason AI projects fail. Not the model. Leadership. RAND calls leadership-driven factors the most frequently cited category, ahead of anything purely technical.

Their five leading root causes: stakeholders misunderstood or miscommunicated the problem. The company lacked usable data. The project chased fashionable technology instead of a business problem. The infrastructure to manage data and deploy models wasn't there. And fifth, in a smaller share of cases, the problem was genuinely too hard for today's AI.

Read that list again. Four of the five are things a leadership team controls: problem definition, data, technology selection, infrastructure. Only one is a hard technical ceiling. Yes, hallucinations are real. Yes, some problems are still out of reach, and RAND says so. But for most of the companies filling up those abandonment statistics, that's not what killed the project. What killed the project was a decision, or the absence of one.

That reframes everything. An expertise gap is a solvable problem. Companies solve them every day: build the capability, hire it, partner for it, or bring it in from outside. Pick deliberately and any of those paths can work. What doesn't work is the default most companies are running right now, which is hoping a better model eventually fixes a problem that was never about the model.

Both camps are making expensive mistakes

Which side of the holy-grail debate you land on isn't just a dinner-party opinion. It's actively shaping bad decisions.

Believers buy first and plan later. Urgency feels responsible when you think the technology alone will carry you, so the sequencing work and the governance work get skipped, and eighteen months later there's a pile of subscriptions, no numbers, and a board asking questions. Those are the companies feeding the failure statistics and wasting real money doing it.

Skeptics make the opposite mistake at the same price. They wait for certainty that isn't coming, because the uncertainty was never about the technology. Meanwhile the competitor who read the data correctly is compounding a lead, and the waiting room, as I've said before, has never had worse odds.

Both camps are staring at the wrong variable. The technology's maturity was never the question. Your organization's readiness is.

The questions that actually matter

Stop reading holy-grail-versus-pipe-dream content. Start asking the questions RAND's interviews point straight at:

Does every AI initiative on your list have a numeric definition of success, agreed before it starts? Is your data actually ready, or does everyone quietly know it isn't and nobody's scheduled the fix? Is each initiative attached to a real business problem, or to the technology of the moment? And one more that RAND doesn't ask directly but their findings beg for: if your executive sponsor changed jobs tomorrow, does the initiative survive?

Answer those well and AI will look a lot like the holy grail the optimists promise. Answer them badly and you'll join the abandonment statistics, and more capable models won't save you, because the model was never your bottleneck.

The debate will keep running for years. It's a good argument with no ending built into it. The only version worth your time is the one about your own company, answered with your own numbers. Focus on outcomes. Not the technology, not which side of the argument flatters your instincts. Neither one was ever going to decide your result.

I've carried a number for the better part of thirty years. Oracle, Pitney Bowes Software, ManpowerGroup, LexisNexis: different logos, different products, the same underlying job. Find out who has the problem, prove you can solve it, and guide them to a "yes" before the fiscal quarter/year closes.

In three decades of leadership, I watched a dozen technologies promise to fix sales. Marketing automation. Predictive lead scoring. Sales engagement platforms that promised to sequence the perfect cadence. Every one of them made the good teams better and made the bad teams worse, faster. It gave them "no place to hide." AI is doing the exact same thing right now, at a scale and speed none of the earlier tools came close to.

Here's the pattern I've watched repeat itself enough times to trust it: technology doesn't necessarily fix a process. It amplifies whatever process is already there.

What amplification actually looks like

Give a disciplined sales team a tool that automates repetitive tasks faster, and they'll use the time they saved to do more productive sales generating activities. The activities that matter. Give an undisciplined team the same tool, and they'll either not use it or use it to send more emails to more people who were never going to buy. Same technology. Undisciplined seller who doesn't follow a sales process. Opposite outcome. The difference was never the tool. Bad sellers often confuse "activities" with "results."

That's the uncomfortable truth sitting underneath most AI-for-sales pitches right now: the technology genuinely works. It writes competent emails. It personalizes outreach at a volume no human rep could match. It can draft a call summary and flag best next steps before you've finished your coffee. None of that is the problem. The problem is that most sales organizations are pointing an extremely capable amplifier at a process that was never disciplined to begin with, and mistaking the resulting noise for progress.

I've sat in enough pipeline reviews to know what an undisciplined process looks like even before AI enters the picture: an ICP (Ideal Customer Profile) that's really just "companies with a pulse and a budget," qualification that happens after the proposal instead of before it, and a forecast built on optimism rather than evidence. Add AI to that mix and here's what actually happens. Outreach volume goes up. Response rates go down, because prospects can tell generated personalization from earned personalization faster than most sales leaders want to believe. Reps spend more time reviewing AI-drafted material than they used to spend writing their own, because the AI doesn't know which of the fifteen personalization variables actually matters to this buyer. And the forecast gets worse, not better, because now it's built on a larger volume of lower-quality signal, run through a tool that will happily tell you a deal is 70% likely to close based on activity that never should have counted as progress in the first place. This will likely drive a wedge between Marketing, who can't measure campaign effectiveness and Sales, who think they are getting bad leads.

The forecast is where this shows up first

Nowhere does broken-process-plus-AI show its work faster than the forecast. A forecasting methodology that was already loose, stage definitions nobody actually enforces, "commit" that means "I feel good about it" rather than "the buyer has done something only a buyer who's going to sign would do, gets worse the moment you introduce AI-generated deal scoring on top of it. The AI isn't wrong. It's doing exactly what it was asked to do: pattern-match against the data it was given. If the data was never disciplined, or worse wrong or missing, the AI just launders that lack of discipline into a number that looks more precise than it is. A gut-feel forecast that's wrong, everyone in the room knows to discount. An AI-scored forecast that's wrong gets trusted right up until the quarter it blows up the board meeting.

I've watched CROs walk into a leadership meeting with an AI-generated pipeline dashboard and more confidence than the underlying process has ever earned. That's not an AI problem. That's a discipline problem wearing a very convincing AI costume. Too often a CRO's credibility is derailed by bad information. And just as bad, you quickly lose confidence in the AI platform you spent money to build.

What actually has to happen first

The fix isn't slower AI adoption. It's sequencing. Before AI touches your sales motion, three things have to be true, and if they aren't, no tool fixes them.

Your ICP has to be specific enough that a rep can disqualify a bad-fit account in the first call instead of the fifth. "Companies with 50 or more employees in our target industries" isn't specific. "Companies with this revenue range, this specific operational pain, and this buying trigger" is specific enough that AI-assisted prospecting can actually target it instead of just scaling noise. Qualifying "out" bad-fit accounts early is critical to a healthy pipeline.

Your stage definitions have to describe a customer's buying cycle, not seller activity. A deal doesn't move stages because a rep sent a proposal. It moves because the buyer did something that only a buyer serious about solving the problem would do. If your stages are seller-activity-based, AI will happily accelerate you through a pipeline that was never real.

And your forecast methodology has to have a mechanism for penalizing bad calls, not just rewarding optimistic ones. If a rep can call a deal "commit" for three quarters running with no consequence for being wrong, adding AI scoring on top of that just makes the wrong calls faster and better-formatted.

Get those three things right, and AI becomes what it should be: leverage on a process that already works, freeing your best reps to spend more of their time on the most important things that actually move a deal forward, the things no model can do for them, which is read a room, navigate internal politics on the buyer's side, and build the kind of trust that survives a bad quarter. Skip that groundwork, and AI just becomes a faster way to generate activity that looks like pipeline and isn't.

The question worth asking before the next tool purchase

Every sales leader evaluating an AI tool right now is asking "what can this do?" That's the wrong first question. The right first question is: what does our process look like today, honestly, without the story we tell ourselves in the QBR? Because whatever that process actually is, gaps included, is exactly what AI is about to amplify. Fix the gaps first, or watch them get faster.

Thirty years of carrying a number taught me that the sales organizations who win aren't the ones with the best tools. They're the ones with the clearest process, applied consistently, long before the tool arrived. AI doesn't change that rule. It just raises the price of ignoring it.

If you want an honest read on where your sales process actually stands before you add AI on top of it, that's the exact diagnostic we run inside the AI Strategy Assessment. Start at embarkonai.com.

There's a word software vendors never put in the pitch deck: "shelfware". It means exactly what it sounds like: software or SaaS licenses that were budgeted, approved, and bought, and then never actually got used. It figuratively sits on the shelf, quietly renewing every year, a line item nobody remembers the justification for anymore.

This isn't a fringe problem. It's the default outcome for a shocking share of enterprise software spend. The number that gets quoted constantly: 96% of organizations admit at least some of the software they've purchased is shelfware, and 39% report that 21% or more of their enterprise software spend is wasted on it. This issue goes back further than most people realize. It's not from a modern SaaS-management vendor. It's from a Flexera and IDC licensing survey conducted back in 2013-14, over a decade ago. And here's the part that should bother you more than the age of the stat: it's still basically true today.

Vertice's current SaaS wastage benchmark, drawn from more than $75 billion in processed software spend, found that as of the second quarter of 2026, 65% of all SaaS licenses are either sitting completely dead (14%) or running at less than half their paid-for capacity (51%). Zylo's numbers tell the same story from a different angle. The average company still leaves somewhere between $18 and $20 million a year on the table in licenses nobody opens, and that figure has climbed, not fallen, since 2023.

Even the classic version of this story holds up: a Gartner survey of 631 companies, fielded in December 2002 and reported in 2003, found 42% of CRM licenses purchased that year went completely unused, at an estimated $1.27 billion in industry-wide waste. That's more than twenty years ago. In my 30 years in the software industry, vendors have sometimes created extra licensing capacity for user growth through longer-term agreements or pricing discounts. However, even with that in mind, these numbers are still staggering. This means that the problem never got fixed. It just kept getting rediscovered by a new vendor with a new benchmark report.

If that sounds like an enterprise problem, Zylo's own breakdown says otherwise. The same report behind that range splits it by headcount: companies with 500 employees or fewer average $3.8 million a year in unused licenses, and companies between 501 and 2,500 average $9.5 million. Smaller companies do not escape this. They just lose smaller absolute numbers out of much smaller budgets.

Here's the part worth sitting with: every leadership team that has ever nodded along to a shelfware statistic and said "yeah, we should be more disciplined about software spend" is currently running the exact same experiment again, with AI, and mostly not recognizing it.

The AI version of the same story

The data on AI project outcomes reads like a rerun with a bigger budget. RAND Corporation's 2024 report is the one everyone quotes for the "80% of AI projects fail" line, and it's worth being precise about what RAND actually found versus what gets repeated on LinkedIn. RAND interviewed 65 experienced data scientists and AI engineers, 50 from industry and 15 from academia, about why AI projects fail. The 80%-plus figure isn't something RAND's interviews produced. RAND's own report frames it as "by some estimates," a number borrowed from earlier industry reporting, not a finding from their study. What RAND's interviews actually produced is more useful anyway: five root causes of failure, and the finding that 84% of its industry interviewees named leadership-driven issues, not the technology itself, as the primary reason projects failed.

Gartner's own track record on this makes the point better than RAND does. In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, and unclear business value. In January 2026, Gartner published what actually happened: at least 50% were abandoned after proof of concept, well past its own forecast. Gartner's updated analysis of hundreds of GenAI implementations also names a failure point the original four didn't cover: poor change management, where a technically sound tool sees minimal adoption because nobody managed the human side of the rollout. The forecast wasn't wrong about the direction. It was too optimistic about the size of the problem.

S&P Global's research tells the same story with a bigger sample. Surveying 1,006 midlevel and senior IT and line-of-business professionals across North America and Europe, S&P Global found that the share of companies abandoning the majority of their AI initiatives before production more than doubled year over year, from 17% in 2024 to 42% in 2025, with organizations reporting on average that 46% of projects are scrapped somewhere between proof of concept and broad adoption.

Different technology, different decade, dramatically similar pattern. And it's worth being precise about what the pattern actually is, because "companies waste money on technology" undersells how specific the parallel really is.

Same root cause, twenty years apart

Shelfware doesn't happen because the software doesn't work. A CRM platform generally does what a CRM platform is supposed to do. Shelfware happens because different people make the decision to buy and the decision to use it, at different times and for different reasons. A budget holder buys the license to solve a problem that looked urgent in a board meeting. The people who'd actually have to change their daily habits to use it weren't in that room, weren't trained properly, and had no particular reason to disrupt a workflow that, however imperfect, already got them through the week. Leadership decisions were not linked to user benefit, and many times users saw this as a reporting exercise for leaders, or worse, micromanagement.

RAND's interviews show the identical mechanism operating inside AI projects, just measured with a sharper instrument. Of the five root causes RAND identified, only one was a genuine technical limitation: applying AI to problems beyond the current state of the art. The other four were organizational. Leaders failed to communicate which problem they wanted solved and which metric defined success, or changed priorities faster than a team could deliver. Organizations lacked the volume and quality of data needed to train a model that worked. Data scientists reached for the most advanced technology instead of the most effective solution to the business problem. And organizations underinvested in the infrastructure required to govern data and actually deploy a finished model.

Put the shelfware research and the AI research side by side, and they're describing the same organizational failure. A purchase decision made at a level disconnected from actual usage. No enforced definition of what "working" would even look like. No mechanism forcing a decision to fully adopt or fully stop, so the initiative just persists in limbo indefinitely. Two different technologies, two different decades, the same gap between the boardroom decision and the desk-level reality.

Where the parallel breaks, and why the break makes AI worse

The comparison isn't perfect, and where it breaks matters more than where it holds. Shelfware is mostly a quiet failure. An unused CRM seat auto-renews for years without anyone at the leadership level particularly noticing. It's expensive, but it's a relatively invisible expense, the kind of waste that survives because nobody has to look at it directly.

AI pilot failure doesn't fail quietly. It fails in front of the board, in front of employees who were told this was the future, in front of a leadership team that staked some credibility on the initiative when it launched. Fortune's June 2025 reporting on what's now being called "AI fatigue" describes exactly this dynamic: leaders at companies like Tines and Netskope describing how repeated failed iterations, and a relentless queue of new AI tools to review, wore down engineering teams, sales teams, and governance staff alike, to the point where the next initiative gets harder to fund, harder to staff, and harder to get genuine buy-in for, because the organization now has a specific, recent memory of AI not working.

Tines cofounder and CEO Eoin Hinchy told Fortune his team went through 70 failures on one AI initiative over the course of a year before landing on something that worked. Most companies don't have the appetite, or patience, to eat 70 failures before finding out whether the 71st one pays off.

S&P Global's finding that abandonment more than doubled year over year, combined with Gartner's own admission that its 30% projection undershot the real number, tells you this isn't a static failure rate settling into a stable baseline. It's compounding, exactly as the shelfware pattern did the first time around, when the technology was CRM instead of AI.

Shelfware quietly wastes money forever. AI pilot failure wastes money and then makes the next attempt more expensive to launch, in political capital if not in dollars. That's a meaningfully worse failure mode than the one it's echoing.

The fix was never about the technology, either time

And here is the part that should encourage you rather than depress you, if you run a company under 1,000 employees. Nearly every failure in this article comes from distance: the gap between the person who approved the purchase and the person who had to change how they work on Monday. That distance is a function of the org chart, not the budget. In a company with a hundred employees, or forty, or fifteen, those two people are often in the same meeting. Sometimes they are the same person. None of the discipline below requires an enterprise balance sheet. It requires someone with the authority to say no, close enough to the work to know when to.

Here's the useful part of the parallel: the reason it's worth understanding instead of just finding depressing. The fix that eventually reduces shelfware, where it does get fixed, has never been a better piece of software. It's utilization management: tying purchases to a named, specific business outcome before the purchase happens, measuring actual usage instead of just spend, and having the organizational discipline to kill or renegotiate what isn't being used instead of letting it quietly renew.

RAND's prescription for reducing AI project failure describes the same discipline. Make sure the technical team understands the project's purpose and its domain context. Choose problems enduring enough to be worth committing a team to for at least a year. Stay focused on the problem, not the technology. Invest up front in the infrastructure that governs data and deploys models. And bring technical experts in early to assess what AI can and cannot actually do before you commit. Put bluntly, success is 100% tied to ongoing leadership team discipline. Projects usually fail, or fall short, without a leadership mandate and follow-through.

None of that is a technology fix, in either case. It's an accountability structure applied before the spending happens, not an audit performed after the waste is already sitting on the books. The companies that build that structure for their AI initiatives are the ones who won't be the subject of this same article in another twenty years, when whatever comes after AI shows up with its own version of the same experiment.

The pattern is old enough now, and documented well enough now, that repeating it is a choice, not an accident. Shelfware taught the exact lesson AI is currently re-teaching, just with faster feedback and a higher price tag for missing it: buy against a plan, or expect a very familiar outcome dressed up in new technology.

That plan (the specific business outcome, the sequencing, the kill criteria) is what we build before a single AI purchase happens. It's the entire difference between running this experiment for the third time and finally learning from it.