Why Are AI Projects Failing? What's Different About the Current Wave

AI projects are failing right now because starting one has become almost free while finishing one is as hard as it ever was. A generative AI tool can produce an impressive demo in an afternoon, so businesses launch projects faster and under more pressure, before anyone has decided which work should change or how they'd know it had. In IBM's 2025 survey of 2,000 CEOs, 64% admitted that the risk of falling behind pushes them to invest before they clearly understand the value, and the same CEOs said only 25% of their AI initiatives had delivered the return they expected (IBM Institute for Business Value, CEO Study 2025, May 2025).

The underlying reasons projects fail haven't changed much. Our guide to why AI projects fail covers the five that keep surfacing across the major studies, from picking the wrong problem to leaving nobody in charge of adoption. What has changed is the speed and the setting, and that's what turns an old set of mistakes into a rising failure rate. S&P Global found the share of companies abandoning most of their AI initiatives before production jumped from 17% to 42% in a single year, as organizations rushed to deliver generative AI at pace (S&P Global Market Intelligence, May 2025).

For an owner of a 10-to-200-person operation, the more useful question is which conditions of the current wave you're walking into, and which of them you can switch off before you spend anything.

Key Takeaways

• AI projects are failing more often because generative AI made starting cheap and fast, while the work of finishing (fitting the tool into the workflow, checking its output, measuring the result) stayed just as hard.
• Pressure drives much of it. In IBM's 2025 CEO Study, 64% of CEOs said the risk of falling behind pushes them to invest before they understand the value, and the same CEOs said only 25% of their AI initiatives had delivered the ROI they expected.
• A lot of AI use never becomes a project at all. MIT found workers at over 90% of companies used personal AI tools for work, while only 40% of companies had bought an official subscription.
• Output that looks finished often isn't. In a 2025 study of 1,150 US employees, 40% had received AI "workslop" in the past month, and KPMG's global survey found 66% of employees using AI at work rely on its output without checking its accuracy.
• People can't reliably feel whether AI is helping. In METR's 2025 trial, experienced developers took 19% longer with AI but believed it had made them 20% faster.
• The technology kept improving while abandonment rose, which points at how projects are chosen and run rather than at the models.

Table of Contents

• Why are AI projects failing? The short answer
• What changed: the demo got cheap, the finish didn't
• Five conditions behind the current wave
• How the current wave feeds the classic failures
• Isn't the technology getting better?
• What this looks like in a smaller operation
• A roofing contractor's first AI rollout
• How to avoid the current wave's mistakes
• Where Avolis fits
• Frequently Asked Questions
• Continue Learning

Why Are AI Projects Failing? The Short Answer

AI projects are failing because they're being started faster than they're being defined. When a tool can draft a proposal or answer a customer email on day one, the hard parts of a project (deciding which workflow should change, plugging the tool into the systems people already use, checking what it produces, and measuring whether anything improved) feel optional, so they get skipped. IBM found that half of the CEOs it surveyed said the pace of recent investment had left their organization with disconnected, piecemeal technology, and only 16% of AI initiatives had scaled across the enterprise.

Those numbers come from large companies, but the pattern travels well. The business buys access before it decides on a use, people try the tool at their own desks, and a few months later nobody can say what it changed. It's what happens when the cost of starting drops to nearly zero and nothing replaces the discipline that a high starting cost used to force.

What Changed: The Demo Got Cheap, the Finish Didn't

Before generative AI, an AI project had a natural filter built in. Building a model meant gathering data, hiring or contracting specialists, and spending months before anyone saw a result, so only projects with a clear business case tended to get approved. That filter was slow and expensive, and it also forced someone to write down what the project was for.

Generative AI removed the filter. Anyone in the business can now open a chat window, paste in a job description or a customer complaint, and get back something that looks like a finished piece of work, which makes every possible use case feel one small step from done. The trouble is that the distance between "it produced a good answer when I tried it" and "it reliably does this job inside our operation" barely shrank. The tool still has to get the right information from the right system, hand its output to the next step, handle the jobs that don't fit the pattern, and earn enough trust that people stop redoing its work by hand.

BCG's 2025 survey of 1,250 senior executives and AI decision makers put numbers on the split between starting and finishing. It classed only 5% of companies as "future-built" and getting value from AI at scale, 35% as scaling and beginning to see value, and 60% as laggards reporting minimal revenue and cost gains despite their investment (BCG, The Widening AI Value Gap: Build for the Future 2025, September 2025).

BCG doesn't call its 60% failed; they're companies reporting minimal gains. Read alongside IBM's numbers, the shape is the same: most organizations have started, few have finished, and the middle is crowded with tools that are in use without changing how the business runs.

Five Conditions Behind the Current Wave

The current wave of failure comes from five conditions that barely existed three years ago. None of them is a new root cause. Each one makes an old mistake easier to make and harder to notice.

1. Pressure to move before anyone knows the value

IBM's finding is the clearest statement of this. Nearly two-thirds of CEOs said the risk of falling behind drives them to invest in some technologies before they understand what the investment will do, and only 37% said it's better to be "fast and wrong" than "right and slow." In other words, most of them are investing faster than they think is wise, which is a recipe for projects that start without a definition and end without a verdict.

Smaller businesses feel the same pull in a different form. It's the peer group where another contractor mentions their AI estimator, the software rep who adds an AI tier to the renewal, or the trade magazine with an AI issue every quarter. None of that is a business case, and a project started to answer it tends to be judged by whether the business "has AI" rather than by whether any job got faster or cheaper.

2. AI used at desks instead of built into the work

A great deal of current AI use never becomes a project at all. MIT's 2025 study of enterprise generative AI found that while only 40% of companies said they had purchased an official AI subscription, workers from over 90% of the companies it surveyed reported regular use of personal AI tools for work tasks (Challapally, Pease, Raskar, and Chari, The GenAI Divide, MIT NANDA, July 2025, p. 8). MIT calls this the "shadow AI economy," and it notes that this personal use often works better than the official initiatives do.

That sounds like good news, and for the individual it often is. McKinsey's August 2026 survey of 1,719 participants found eight in ten respondents said AI had improved their own productivity, yet the share attributing any EBIT impact to AI was stuck at 37%, essentially unchanged from a year earlier (McKinsey, The State of AI in 2026: On the Road to ROI, August 2026). For the business, desk-level use means the gains stay locked at one desk. An office manager who drafts collection emails faster with a personal chatbot has saved some time, but the invoicing workflow hasn't changed, nobody else benefits, and nothing is measured. When the business later launches an official AI project, it often starts from the same desk-level habit and assumes a workflow-level result will follow, which is where many pilots stall. Our page on why AI consulting projects fail to reach production follows that stall in detail.

3. Output that looks finished but isn't

Generative AI is very good at producing work that has the shape of a finished product. That's useful when a person reviews it, and costly when nobody does. Researchers from BetterUp Labs and the Stanford Social Media Lab use the term "workslop" for AI-generated work that looks polished but lacks the substance to move a task forward. Of 1,150 US full-time employees they surveyed, 40% said they had received workslop in the past month, and those who had received it estimated that 15.4% of the content they receive fits the description (Niederhoffer and colleagues, "AI-Generated 'Workslop' Is Destroying Productivity," Harvard Business Review, September 2025).

KPMG and the University of Melbourne found the same habit on a global scale. In their 2025 survey of more than 48,000 people in 47 countries, 66% of employees who use AI at work said they rely on its output without evaluating its accuracy, and 56% said they had made mistakes in their work because of AI (KPMG and University of Melbourne, Trust, Attitudes and Use of Artificial Intelligence: A Global Study 2025, April 2025). In an operation, those mistakes don't stay on a screen. They turn into a proposal with the wrong quantity, a lease summary that misses a renewal clause, or a scope letter that promises work the crew never priced.

4. Nobody can tell whether it's working

One of the more counterintuitive findings of the past two years is that people are poor judges of whether AI is saving them time. METR, a research nonprofit, ran a randomized trial in early 2025 with 16 experienced open-source developers working on 246 real tasks in their own codebases. With AI tools allowed, they took 19% longer to finish. Before the study they expected AI to speed them up by 24%, and afterward they still believed it had sped them up by 20% (METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity", July 2025).

METR's February 2026 follow-up, with 57 developers and newer tools, pointed toward a speedup this time, but METR called its own data "very weak evidence" because many developers refused to work without AI at all (METR, "Uplift Update", February 2026). The lasting lesson is that the feeling of being faster is a bad measuring stick, and a business that judges an AI project by how it feels will keep paying for tools that may or may not be helping.

5. Costs that grow with use

Older software usually came with a fixed price per seat. Many AI tools charge by usage, measured in tokens (the chunks of text a model reads and writes), so the bill rises as people use the tool more and as prompts get longer. EY's July 2026 survey of 534 US senior leaders found 82% of those investing in AI said their organization was concerned about token usage and related costs, and 98% of those using token-based tools said those costs had made them reconsider their approach (EY, US AI Pulse Survey, fifth wave, July 2026).

McKinsey's 2026 survey found the same pressure, with about 20% of respondents saying AI operating costs, including token costs, had constrained their AI use. In EY's survey, more leaders said token costs had pushed them to expand their rollout (37%) than to shrink it (15%). The point for a smaller business is narrower: a usage-priced tool with no budget ceiling and no measure of value is the easiest AI project to cancel the first time a bill surprises someone.

How the Current Wave Feeds the Classic Failures

Four of the classic failure reasons get easier to make in the current wave, and missing data for the task is the one that hasn't changed. The table pairs each condition with the reason it feeds, and shows what it tends to look like in a 50-person trades or manufacturing business.

None of these questions needs technical skill, and each one is cheap to answer before a project starts and expensive to discover after it fails.

Isn't the Technology Getting Better?

Yes, and that's the most telling fact about the current wave. Models are more capable than they were two years ago by almost any measure, and METR's February 2026 follow-up pointed toward a speedup with newer tools, even though METR rates that data as weak evidence. Yet abandonment rose sharply over the same stretch, from 17% to 42% of companies scrapping most of their initiatives in S&P Global's surveys.

If failure tracked the quality of the technology, better models would mean fewer failed projects. The fact that it hasn't worked out that way points back at the decisions around the technology: which job it's pointed at, whether it's wired into the work, who checks it, and whether anyone measured the result. IBM's 2026 CEO Study found that organizations that redesigned five core areas (technology, finance, HR, operations, and cross-functional collaboration) were four times more likely to have met their business objectives (IBM Institute for Business Value, CEO Study 2026, May 2026). The difference showed up in how the business was run, not in which model it bought.

This is also why waiting for the technology to mature isn't much of a strategy. A business that adopts next year's model with this year's habits will get this year's result.

What This Looks Like in a Smaller Operation

Nearly every survey above describes large companies, so it's fair to ask how much of it applies to a 30-person HVAC firm or a 120-person fabricator. Smaller firms are earlier in the wave, not outside it. The Census Bureau's Business Trends and Outlook Survey found that between December 2025 and May 2026, between 17% and 20% of US businesses used AI in some business function. Use was 37% among firms with at least 250 employees and 32% among firms with 100 to 249, while firms with fewer than 20 employees showed no significant increase over the period (US Census Bureau, "Large Firms With at Least 20 Employees Biggest AI Users", May 2026).

That gap matters for two reasons. First, the vendors and tools that pushed large companies through the cycle of fast starts and quiet abandonment are now selling to smaller ones, with the same demos and the same urgency. Second, a smaller operation has less room for a wasted attempt. It doesn't have a data team to absorb a bad pilot, and the person who gets pulled in to clean up after a tool is usually the office lead or the owner, whose time is already the scarcest thing in the building.

There's an upside too. A smaller business can see its own workflows end to end, the person who runs dispatch is a short walk from the person who signs the checks, and a decision that takes a large company two quarters can happen in a week. Every condition in the table above is easier to switch off at 50 people than at 5,000, provided somebody actually switches it off.

A Roofing Contractor's First AI Rollout

Take a 50-person commercial roofing contractor (an illustrative composite) with three estimators, a four-person office, and crews spread across a metro area. After a trade association meeting where two other contractors talk about AI, the owner signs the office up for a general-purpose AI assistant on a plan that charges per seat plus usage, about $40 a seat in the first month, and tells the team to "start using it."

The first month looks like progress. The estimators use it to draft proposal language, the office manager uses it for collection emails, and everyone agrees it's saving time. By the fourth month, three problems have surfaced. A proposal goes out with a warranty term the manufacturer doesn't offer, copied from a draft nobody checked against the spec sheet, and the company has to honor it on a 60,000-square-foot job. The monthly bill has nearly tripled since the estimators started pasting in full specification sets, and nobody owns it. And when the owner asks how much time the tool is saving, nobody can answer, because nobody timed the work before the rollout. The estimators feel faster, but the proposal backlog is the same length it was in the spring.

Lay the story against the five conditions and they line up one for one. The purchase answered peer pressure rather than a named problem, the tool stayed at individual desks, its output went out unreviewed, the "time saved" was a feeling, and the bill that climbed with use was never set against a result. The technology did what it was asked. The rollout never asked it anything specific.

The company comes back to it in the fall with a narrower brief. It times the office's work for a month and finds that job closeout eats about 18 hours a week: assembling warranty submittals, photo logs, and as-built notes into the packet the manufacturer and the building owner need before final payment is released. Late packets hold up cash on every finished roof. So the tool gets one job, drafting the closeout packet from the crew's daily logs and photos and saving it to the job's record in the project management system, and the project manager on each job reviews the packet against a short checklist (warranty terms, roof areas, and missing photos) before it goes out. Usage is capped at a monthly figure the controller watches, days from substantial completion to final payment are tracked every week, and the owner puts a date on the calendar three months out to decide whether it stays.

How to Avoid the Current Wave's Mistakes

You avoid the current wave's failures by putting back the discipline that cheap tools removed, which comes down to five short decisions before you buy, each aimed at one of the conditions above.

1. Name the job, not the tool. Write down the one workflow the project should change and roughly how many hours a week it takes today. If the best answer is "everyone will use it for everything," you're buying seats, not starting a project.
2. Decide where the output lands. The result should appear in the system the next person already works in, whether that's the job management software, the accounting package, or the scheduling board, rather than in a chat window someone has to copy from.
3. Assign a checker. Name the person who reviews AI output before it reaches a customer, a crew, or a contract, and write down what they check. This is the cheapest protection against finished-looking mistakes.
4. Measure the work, not the feeling. Time the task before the project starts and again after it's been running for a month (our guide to measuring AI ROI covers setting the baseline). A felt speedup is a hypothesis, and METR showed how far off it can be.
5. Cap the bill and set a review date. Agree on a monthly ceiling for usage costs, put one person in charge of watching it, and pick the date you'll decide whether the project stays.

If you already have a tool in place and the answers are mostly blank, that's a recoverable position. Our guide to recovering a failing AI project covers how to tell a slow project from a failing one and decide whether to fix, narrow, or stop it. If the project in question involves AI agents (tools that take actions on their own rather than just drafting), the risks compound, and our page on why agentic AI projects fail covers that case.

Where Avolis Fits

The conditions behind the current wave all come down to skipping the step between "we should do something with AI" and "here's the tool." That step is where we start. Every engagement begins with the diagnostic, in which we work through your workflows with the people who run them, measure volumes, time per task, and handoffs, and rank each one by what it would return and what it would take to build. You get a ranked map with a go, wait, or no on each workflow, and it's yours to keep whatever you decide next.

From there, we build the first workflow inside the systems your team already uses, set up the review step and the measurement with you, and stay embedded while it settles in, so the five questions above get answered with numbers rather than impressions. To see what a good diagnostic covers, read our guide to AI readiness assessment services. If you're weighing several partners, our guide to which AI consulting company to choose lists what to ask each of them.

Frequently Asked Questions

Why are AI projects failing?

AI projects are failing because generative AI made them cheap and fast to start, while integrating, checking, and measuring them stayed hard. Pressure plays a big part: in IBM's 2025 survey, 64% of CEOs said the risk of falling behind drives investment before the value is clear, and the same CEOs said only 25% of their AI initiatives had delivered the ROI they expected.

Why do most AI projects fail to deliver value right now?

Because much of it stays at individual desks instead of changing a workflow, so personal gains don't show up in the business's numbers. McKinsey's 2026 survey found eight in ten respondents said AI improved their own productivity, yet only 37% attributed any EBIT impact to AI, essentially unchanged from a year earlier.

Why are more AI projects being abandoned now?

Generative AI made projects cheap to start, so more of them begin without a named workflow, a reviewer, or a baseline. S&P Global found the share of companies abandoning most AI initiatives before production rose from 17% to 42% in a year, as organizations rushed to deliver generative AI at pace.

Is the AI technology the reason projects fail?

Rarely. Models have improved quickly, yet abandonment rose over the same period, which points at how projects are chosen and run. IBM's 2026 CEO Study found organizations that redesigned five core areas, including operations, were four times more likely to have delivered on business objectives.

Why are AI projects failing at small businesses?

Small businesses face the same pressure and tools as large ones, with less room for a wasted attempt. Census data show no significant increase in AI use among firms with fewer than 20 employees between December 2025 and May 2026. The best protection is naming one measured workflow, a reviewer for output, and a cost ceiling before buying.

Continue Learning

The models aren't what's failing. What's failing is the habit of buying access first and deciding on the purpose later, which generative AI made easy and the pressure of the moment made common. The businesses getting value out of this wave are the ones that put back the step a high price used to force: naming the job, deciding who checks the work, and measuring the result before calling it a success.

Before you buy another AI subscription, write down the one job it's for and how long that job takes today.

Why projects fail, in detail:

• Why do AI projects fail? Common reasons and statistics
• Why agentic AI projects fail
• Why your AI consulting project isn't reaching production
• Why AI projects fail due to data quality

When a project is already in trouble:

• How to recover a failing AI project
• AI pilot purgatory
• How to measure AI ROI

Starting the right way:

• AI readiness assessment services
• Which AI consulting company should I choose?

---

Sources

All sources retrieved 2026-10-08.

• IBM Institute for Business Value, "IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles," press release, 6 May 2025 (2,000 CEOs from 33 countries and 24 industries, surveyed with Oxford Economics, February to April 2025), retrieved 2026-10-08: https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles
• IBM Institute for Business Value, "IBM Study: CEOs Are Reshaping C-Suite Roles for the AI Era," press release, 4 May 2026 (2,000 CEOs and equivalent senior leaders across 33 geographies and 21 industries, February to April 2026), retrieved 2026-10-08: https://newsroom.ibm.com/2026-05-04-ibm-study-ceos-are-reshaping-c-suite-roles-for-the-ai-era
• Boston Consulting Group, "AI Leaders Outpace Laggards with Double the Revenue Growth and 40% More Cost Savings," press release, 30 September 2025, for the report The Widening AI Value Gap: Build for the Future 2025 (global survey of 1,250 senior executives and AI decision makers across nine industries), retrieved 2026-10-08: https://www.bcg.com/press/30september2025-ai-leaders-outpace-laggards-revenue-growth-cost-savings
• Dan Tinkoff, Lieven Van der Veken, and Michael Chui, with Tara Balakrishnan, The State of AI in 2026: On the Road to ROI, McKinsey & Company, 25 August 2026 (online survey of 1,719 participants in 97 nations, fielded 4 May to 8 June 2026; 36% from organizations with more than $1 billion in revenue), retrieved 2026-10-08: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
• Aditya Challapally, Chris Pease, Ramesh Raskar, and Pradyumna Chari, The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025, p. 8 (preliminary findings; 300+ public AI initiatives reviewed, interviews with 52 organizations, survey of 153 senior leaders), retrieved 2026-10-08: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
• Kate Niederhoffer, Gabriella Rosen Kellerman, Angela Lee, Alex Liebscher, Kristina Rapuano, and Jeffrey T. Hancock, "AI-Generated 'Workslop' Is Destroying Productivity," Harvard Business Review, 22 September 2025 (ongoing survey of 1,150 US full-time employees, BetterUp Labs and Stanford Social Media Lab), retrieved 2026-10-08: https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
• KPMG and the University of Melbourne, Trust, Attitudes and Use of Artificial Intelligence: A Global Study 2025, April 2025 (more than 48,000 people in 47 countries, November 2024 to January 2025), retrieved 2026-10-08: https://kpmg.com/sg/en/home/insights/2025/04/trust-attitudes-and-use-of-ai.html
• METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," 10 July 2025 (randomized controlled trial, 16 developers, 246 tasks), retrieved 2026-10-08: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
• METR, "Uplift Update," 24 February 2026 (57 developers, 143 repositories, 800+ tasks), retrieved 2026-10-08: https://metr.org/blog/2026-02-24-uplift-update/
• EY, "EY Survey: C-Suites Pivot from AI Adoption to Unlocking Value as Escalating Token Costs Trigger Fiscal Scrutiny," press release, 28 July 2026 (fifth wave of the EY US AI Pulse Survey, 534 US senior leaders at SVP level and above, fielded 24 April to 17 May 2026), retrieved 2026-10-08: https://www.ey.com/en_us/newsroom/2026/07/ey-survey-c-suites-pivot-from-ai-adoption-to-unlocking-value-as-escalating-token-costs-trigger-fiscal-scrutiny
• US Census Bureau, "Large Firms With at Least 20 Employees Biggest AI Users," America Counts story, May 2026 (Business Trends and Outlook Survey, December 2025 to May 2026), retrieved 2026-10-08: https://www.census.gov/library/stories/2026/05/ai-use-businesses.html
• S&P Global Market Intelligence, "AI experiences rapid adoption, but with mixed outcomes: Highlights from VotE: AI & Machine Learning," 30 May 2025 (survey of 1,006 IT and line-of-business professionals, North America and Europe), retrieved 2026-10-08: https://www.spglobal.com/market-intelligence/en/news-insights/research/ai-experiences-rapid-adoption-but-with-mixed-outcomes-highlights-from-vote-ai-machine-learning
• James Ryseff, Brandon F. De Bruhl, and Sydne J. Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RAND Corporation, RR-A2680-1, 2024, retrieved 2026-10-08: https://www.rand.org/pubs/research_reports/RRA2680-1.html

On what the surveys count. IBM's 25% is CEOs' own estimate of the share of their AI initiatives that met ROI expectations, not a measured failure rate. BCG's 60% counts companies reporting minimal revenue and cost gains, not abandoned projects. McKinsey's 37% counts respondents attributing any EBIT impact to AI. All three surveys lean toward large organizations.

On the METR figures. The 2025 trial covers experienced software developers working in their own repositories, so it shows how unreliable self-assessment of AI's effect can be, not how AI affects office or field work. METR's February 2026 update reports estimated speedups with wide confidence intervals and describes its new data as very weak evidence because of selection effects.

On the Census figures. The Census Bureau changed its core AI question in November 2025, from AI used "in producing goods or services" to AI used "in any business function," so rates from before and after that change aren't directly comparable.

On the workslop survey. The 40% and 15.4% figures come from an ongoing, self-reported online survey of US desk workers, as reported by the authors in Harvard Business Review.

On the roofing contractor. It's an illustrative composite, not a client. Its figures (50 people, about $40 a seat, a 60,000-square-foot job, 18 hours a week of closeout work) are representative examples, not measurements or benchmarks.

---

About Avolis Research Group

Avolis Research Group is Avolis's in-house research practice, focused on how operations-heavy small and mid-sized businesses actually adopt AI. It synthesizes primary economic research, government survey data, and results from real implementations into practical, vendor-neutral guidance.

More about Avolis and how we work · Get in touch