Why Do AI Projects Fail? Common Reasons and What the Statistics Count

Avolis Research Group

·

·

24 min read

MIT's 95%, the 80% RAND cites, and Gartner's 85% count different things. What AI failure statistics actually measure, and five reasons projects fail.

AI projects mostly fail because of decisions made before anything is built, not because the technology breaks. Across the major studies, five causes come up again and again: picking the wrong problem, never building the tool into the way the work actually runs, finding out too late that the data for that task isn't there, leaving cost and value undefined, and assuming people will adopt a tool nobody owns. In RAND's 2024 interviews with experienced AI practitioners, 84% named a leadership-driven cause as the primary reason projects fail, ahead of data, infrastructure, or the limits of the technology itself.

The statistics are where most explanations go wrong. You've probably seen "95% of AI pilots fail," "80% of AI projects fail," and "85% of AI projects fail," often in the same article and treated as one finding. They aren't. Each number counts something different, one of them was never a failure statistic at all, and another is a citation of a citation. The clearest recent measurement comes from S&P Global. It found that the share of companies abandoning most of their AI initiatives before production jumped from 17% to 42% in a single year (S&P Global Market Intelligence, Voice of the Enterprise: AI & Machine Learning, Use Cases 2025, May 2025).

Once you know what each number counts, the reasons behind them get much clearer, and so does what a business of 10 to 200 people can do about them.

Key Takeaways

  • AI projects fail mainly because of decisions made before the build: the wrong problem, no fit with the workflow, missing data for that task, undefined value, and no owner for adoption. In RAND's 2024 study, 84% of interviewees named a leadership-driven cause as the primary one.
  • The famous failure rates measure different things. MIT's 95% counts organizations seeing no measurable return from generative AI, IDC's 88% counts pilots that never reached production, and S&P Global's 42% counts companies that abandoned most of their initiatives.
  • Gartner's often-quoted "85% of AI projects fail" comes from a 2018 prediction that 85% of projects would deliver erroneous outcomes through 2022, which is a claim about bias, not abandonment.
  • RAND's "more than 80%" is attributed in its own footnotes to a 2022 Fortune article, so it's an estimate RAND repeated rather than a rate RAND measured.
  • In MIT's funnel, 20% of organizations piloted a task-specific AI tool and 5% implemented one successfully, so roughly one pilot in four succeeded, not one in twenty.
  • Abandonment is rising. Companies scrapping most AI initiatives before production rose from 17% to 42% between S&P Global's 2024 and 2025 surveys, and the average firm scrapped 46% of its proofs of concept.

Table of Contents

Why Do AI Projects Fail? The Short Answer

AI projects fail when the people paying for them never pin down which piece of work should change, by how much, and who will own the change once the tool is live. That's the central finding of a 2024 RAND report built on interviews with 65 experienced data scientists and engineers (Ryseff, De Bruhl, and Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RAND, 2024). RAND found five root causes, and the first one dominated the rest.

In RAND's order, they are a misunderstood or miscommunicated problem, a lack of the data needed to train an effective model, and a focus on the newest technology over the user's actual problem. The last two are inadequate infrastructure to manage data and deploy models, and problems that are simply too hard for AI to solve. Of RAND's interviewees, 84% named one or more leadership-driven causes as the primary reason projects fail. Data problems came second, with 30 of the 50 industry practitioners describing persistent data-quality issues, and the remaining causes were each raised by a quarter to a third of participants.

The pattern matters because it tells you where to look. If most failures came from weak models, the fix would be waiting for better ones. Because most come from decisions made before anyone writes a line of code, the fix is in how the project is chosen and scoped, which is work a business owner can do and doesn't need an engineer to start.

How Many AI Projects Fail? What Each Statistic Counts

Published failure figures run from 30% to 95%, and the range is that wide because the studies count different things. Some measure organizations, some measure projects, and some measure pilots. Some are surveys of what happened, and others are analyst predictions of what will happen. Before you repeat any of these figures to a partner or a board, it's worth knowing which kind you're holding.

Figure Source What it actually counts Measured or predicted
95% MIT NANDA, July 2025 Organizations getting no measurable P&L return from generative AI Measured (interviews, survey, public deployments)
88% IDC with Lenovo, 2025 AI proofs of concept that didn't reach wide deployment (4 of every 33 did) Measured (survey)
85% Gartner, February 2018 AI projects predicted to deliver erroneous outcomes through 2022 Predicted
More than 80% RAND, 2024 AI projects that fail, "by some estimates," citing a 2022 Fortune article Secondhand estimate
46% S&P Global, May 2025 Average share of a company's AI projects scrapped between proof of concept and broad adoption Measured (survey of 1,006)
42% S&P Global, May 2025 Companies abandoning the majority of their AI initiatives before production, up from 17% Measured (survey of 1,006)
Over 40% Gartner, June 2025 Agentic AI projects predicted to be canceled by the end of 2027 Predicted
At least 30% Gartner, July 2024 Generative AI projects predicted to be abandoned after proof of concept by the end of 2025 Predicted
Why Do AI Projects Fail? Common Reasons and What the Statistics Count - Avolis AI Eight "failure rates," eight different things counted Widely cited AI failure figures, labeled by what each measures Measured Predicted Secondhand estimate MIT NANDA 2025 Orgs with no measurable gen-AI return 95% IDC 2025 Pilots not reaching production 88% Gartner 2018 Projects with erroneous outcomes 85% RAND 2024 Projects failing "by some estimates" 80%+ S&P Global 2025 Average share of pilots scrapped 46% S&P Global 2025 Firms abandoning most initiatives 42% Gartner 2025 Agentic projects canceled by 2027 40%+ Gartner 2024 Gen-AI projects dropped after pilot 30%+
Sources: MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025; IDC research with Lenovo as reported by CIO, March 2025; Gartner press releases of 13 February 2018, 29 July 2024, and 25 June 2025; RAND RR-A2680-1, 2024; S&P Global Market Intelligence, Voice of the Enterprise: AI & Machine Learning, Use Cases 2025, May 2025. Bars are not comparable with one another; that's the point of the chart.

Read together, the measured figures describe a simple pattern: most organizations start more AI work than they finish, and many of the projects that do finish don't produce a return anyone can point to. That's a long way from "95% of AI is useless," which is how the numbers tend to get repeated.

Where the Famous Numbers Come From

The three most repeated AI failure figures each come with a caveat that rarely travels with them. Read in the original, each one says something narrower than its reputation.

Gartner's "85%" was a prediction about bias

The figure usually written as "Gartner says 85% of AI projects fail" comes from a February 2018 press release about CIO adoption plans. The actual sentence predicted that through 2022, 85% of AI projects would "deliver erroneous outcomes" because of bias in data, algorithms, or the teams managing them (Gartner, "Gartner Says Nearly Half of CIOs Are Planning to Deploy Artificial Intelligence", February 2018). That's a forecast about flawed outputs, made when only 4% of the CIOs Gartner surveyed had implemented AI at all, and its forecast window closed nearly four years ago.

RAND's "more than 80%" is RAND quoting someone else

RAND's report opens by noting that "by some estimates, more than 80 percent of AI projects fail," twice the failure rate of IT projects that don't involve AI. Its own footnote for the 80% figure points to a July 2022 Fortune article, and the "twice the rate" comparison points to a 2023 Harvard Business Review piece. RAND's real contribution is its interview study and the five root causes that came out of it. The 80% is borrowed context, and "RAND found that 80% of AI projects fail" is a claim RAND never made.

MIT's "95%" uses a strict definition of success

MIT's Project NANDA report says 95% of organizations are "getting zero return" from $30 billion to $40 billion in enterprise generative AI investment (Challapally, Pease, Raskar, and Chari, The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025). The report counts a task-specific tool as successfully implemented only when users or executives described it as causing "marked and sustained" productivity or P&L impact. It labels its funnel figures "directionally accurate" and based on interviews rather than company reporting, and its cover calls the work preliminary findings. A tool that saved an office manager a few hours a week, but that nobody would describe as a marked and sustained change, would count as a failure under that test.

Why Do AI Projects Fail? Common Reasons and What the Statistics Count - Avolis AI Of the pilots, about one in four succeeded Share of organizations at each stage, embedded or task-specific gen-AI tools 60% Investigated 20% Piloted 5% Implemented 5 of 20 pilots = 1 in 4 Figures are MIT's own estimates from interviews, labeled "directionally accurate." Success means users or executives reported marked, sustained productivity or P&L impact.
Source: Challapally, Pease, Raskar, and Chari, The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025, pp. 3 and 6–7 (300+ public AI initiatives reviewed, interviews with 52 organizations, 153 senior leaders surveyed, January to June 2025). The one-in-four ratio is our arithmetic from MIT's figures.

What MIT's funnel actually shows: the 5% is a share of all organizations, including the 40% that never looked at a task-specific tool and the 40% that looked and stopped there. Of the organizations that actually ran a pilot, 5 out of 20 implemented the tool successfully, which is one in four. That's still a poor rate, and IDC found something similar with an easier bar to clear: 4 of every 33 pilots reached production. The lesson for a business owner is less "AI almost never works" and more "most pilots are started without the conditions to finish," which is something you control.

What MIT Found About Why AI Projects Fail

MIT's answer to why AI projects fail is what it calls the learning gap. Most generative AI tools don't retain feedback, adapt to the business's context, or improve over time, so they work for quick one-off tasks and break down in the workflows that matter. The report found that the divide between success and failure wasn't driven by model quality or regulation. It was driven by approach, with tools failing because of "brittle workflows, lack of contextual learning, and misalignment with day-to-day operations."

Two other findings from the report are more useful to an operations business than the headline. First, budgets tend to go where results are easiest to show rather than where they're largest. Executives in MIT's survey put about half of their generative AI spending into sales and marketing, yet the report found that back-office automation often produced faster payback and clearer cost reductions. The biggest savings it documented came from cutting outsourced processing and agency spend, not from cutting staff. Second, the report noted that manufacturers tended to over-index on operations rather than marketing, and quoted a mid-market manufacturing COO who said that apart from processing some contracts faster, nothing fundamental in their operations had changed.

MIT also found that externally partnered projects reached deployment about twice as often as internal builds, though it warns that the gap may reflect the organizations rather than the approach. Our comparison of AI development companies and in-house teams digs into what that finding means for who should build your first project.

Why AI Projects Fail, According to Gartner

Gartner's answer comes down to four causes: poor data quality, inadequate risk controls, escalating costs, and unclear business value. Its July 2024 forecast named all four when it said at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025 (Gartner, "Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025", July 2024). The same release put enterprise costs for building generative AI into a business model at $5 million to $20 million, which is a useful reminder of whose projects these forecasts describe.

The June 2025 forecast on AI agents (software that takes actions on its own rather than just answering) is sharper. Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, repeating three of the same causes: escalating costs, unclear business value, and inadequate risk controls (Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", June 2025). It also estimated that only about 130 of the thousands of vendors selling "agentic AI" offer the real thing, calling the rest "agent washing," which means rebranding existing chatbots and automation tools. Its analyst added that many use cases sold as agentic don't need an agent at all. Our guide to why agentic AI projects fail goes further into that pattern.

On data, Gartner predicts that through 2026, organizations will abandon 60% of AI projects that aren't supported by AI-ready data. The forecast rests on a July 2024 survey of 1,203 data management leaders, 63% of whom didn't have, or weren't sure they had, the right data practices for AI (Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk", February 2025). What "AI-ready" means depends heavily on the task, which is the subject of our page on why AI projects fail due to data quality.

The Five Common Reasons AI Projects Fail

Line the studies up side by side and the same five reasons keep surfacing. The names differ (RAND says "misunderstood problem," Gartner says "unclear business value," MIT says "misalignment with day-to-day operations"), but they describe the same handful of mistakes. RAND's other two causes, weak infrastructure for deploying models and problems too hard for AI, mostly trouble teams that build their own models, which a 10-to-200-person business rarely does.

Reason What the evidence says What it looks like in a 40-person operation Early warning sign
1. The wrong problem RAND: leadership-driven causes were the primary reason for 84% of interviewees An AI chatbot for the website when the real drain is quote write-ups Nobody can say which hours or errors the project removes
2. Not built into the workflow MIT: brittle workflows and tools that don't learn the business's context A tool the estimator has to copy results out of and paste into the job system Staff use it "on the side" and the old process keeps running
3. Data missing for this task RAND: 30 of 50 described persistent data-quality issues; Gartner: 60% abandonment forecast Job histories spread across three systems and a shared inbox The first real test uses sample data instead of your records
4. Cost and value never defined Gartner: escalating costs and unclear value; S&P: cost was the worst-performing project measure A monthly bill nobody compares against hours saved No before-and-after number exists for the work
5. No owner for adoption MIT: resistance to new tools was the top barrier; S&P: high-failure firms met more employee resistance The dispatcher quietly double-checks every AI schedule by hand No named person is accountable once the vendor leaves

1. The project solves the wrong problem

The most common failure starts before anything is built. Someone decides the business needs AI, a use case gets picked because it's visible or because a vendor demoed it well, and months later the tool works as designed and changes nothing that matters. RAND describes data science teams working hard for months to deliver a model that makes little impact on the organization, usually because of a communication breakdown between the people who know the business and the people building the tool.

In a trades or manufacturing business, this usually looks like starting at the front of the house. A customer-facing chatbot is easy to picture and easy to demo, while the hours actually draining away sit in estimating, scheduling, and the paperwork that follows every job. MIT's budget finding, with half of spending going to sales and marketing while the back office delivered the faster payback, is the same mistake measured at scale.

2. The tool never becomes part of how the work runs

A pilot can succeed as a demonstration and fail as an operation. The difference is whether the tool sits inside the systems and steps people already use, or next to them. When the estimator has to open a separate screen, copy an answer, and paste it into the job management system, the AI is extra work, and extra work gets dropped the first busy week. MIT's learning gap describes the deeper version of this: tools that forget context and have to be re-taught the business every session. Deloitte's 2026 enterprise survey of 3,235 leaders found the same gap from the other side. Only 30% of organizations were redesigning key processes around AI, while 37% were using it at a surface level with little or no change to underlying processes (Deloitte, State of AI in the Enterprise, January 2026). The failures our sibling page on AI projects failing to reach production examines are mostly this reason in disguise.

3. The data the task needs isn't there

Data comes second in RAND's ranking and is the reason vendors mention most. It's also narrower than "your data isn't clean." The question is whether the specific records this task depends on exist, are complete enough, and can be reached. A quoting tool needs past quotes with the actual costs attached. A scheduling tool needs reliable job durations, and nothing else in the database will stand in for them. Plenty of businesses with messy data overall have exactly the records one workflow needs, and plenty of tidy ones are missing the single field that matters.

4. Nobody defined the cost or the value

Gartner's "escalating costs" and "unclear business value" are two halves of one omission. If nobody measured how long the work took before the project, nobody can show what the project saved, and the ongoing bill starts to look like pure cost the moment budgets tighten. In S&P Global's 2025 survey, cost was both the most common factor in deciding which projects to pursue and the worst-performing measure once projects were running. And among organizations that had invested in generative AI, 46% said no single business objective had seen a strong positive impact. Our guide to measuring AI ROI covers how to set the baseline before you start.

5. Nobody owns adoption after launch

The final reason is the one most often blamed on "the team," and it's usually a management gap instead. In MIT's survey of executives and frontline users, resistance to adopting new tools was the top-rated barrier to scaling. S&P Global found that companies with higher project failure rates were notably more likely to run into resistance from customers and employees. In practice, resistance often means the people doing the work were never asked what would help, so they keep the old process running alongside the new one, just in case. A dispatcher who still checks every AI-built schedule by hand is telling you the tool hasn't earned trust yet, and somebody needs to own the job of earning it.

Why AI Projects Are Failing More Often Now

AI projects are being abandoned more often than they were two years ago, and the main reason appears to be speed. S&P Global's survey found the share of companies abandoning most initiatives before production rose from 17% to 42% in a year, reversing what had been a downward trend, as organizations tried to deliver generative AI at pace. The same survey found that positive impact had fallen across every objective it tracked, with revenue growth dropping from 81% to 76% and cost management from 79% to 74%.

IDC's group vice president Ashish Nadkarni described many generative AI pilots as born at the board level out of panic and launched underfunded or not funded at all (CIO, "88% of AI pilots fail to reach production — but that's not all on IT", March 2025). Put those findings together and the picture is of more projects started faster with less scoping, which is precisely the condition that produces the five failures above. Our page on why AI projects are failing looks at the current wave in more detail.

There's a more hopeful finding in the same S&P data. Organizations with lower failure rates took a more holistic approach to choosing projects and were more likely to weigh compliance, risk, and data availability before starting. The difference between the two groups showed up at selection, not at execution.

What the Statistics Miss About Smaller Operations

Nearly every failure statistic in circulation comes from large organizations. S&P Global surveyed IT and business professionals across North America and Europe, and Gartner's forecasts describe projects costing millions. MIT's only size split is between enterprises (over $100 million in revenue) and the mid-market. That split is encouraging for smaller firms: MIT found enterprises ran the most pilots but had the lowest pilot-to-scale conversion, while top-performing mid-market companies averaged 90 days from pilot to full implementation, against nine months or longer for enterprises. Still, none of the studies above report a failure rate for firms of 10 to 200 people, so any precise figure you see for small businesses deserves a hard look at where it came from.

The closest thing to small-business evidence comes from the Federal Reserve's Small Business Credit Survey, which asked 6,525 employer firms with 1 to 499 employees about AI in late 2025. Nearly half (46%) said the business or its employees used AI. Of those users, about half were still experimenting, 44% had partially integrated it into their processes, and just 7% had fully integrated it (Federal Reserve Banks, 2026 Report on Employer Firms, March 2026). The most common use was writing or marketing, at 83%, which is help at an individual desk rather than a change to how a workflow runs.

The challenges those firms reported read like the enterprise failure list in plain language. AI users named accuracy (46%) and adapting tools to meet their business needs (43%) as their top problems. Firms planning to start worried most about finding tools that fit their business (54%) and the time it takes to implement them or train staff (37%). Fit with the workflow, in other words, is the small-business version of MIT's learning gap. None of these surveys measures the thing the failure statistics claim to describe, which is firms that tried an AI project and stopped. The Census Bureau, the Federal Reserve, and Statistics Canada all ask about current use, plans, and reasons for not using AI, but not about abandonment.

What the enterprise data does tell a smaller operation is useful in a different way. S&P Global found that skills shortages hit hardest in manufacturing, construction, and real estate, and that those same industries reported notably lower returns on AI investment. Close to half of the organizations short on skills were turning to outside integrators and consultants to fill the gap. A 40-person contractor sits squarely in the industries where the evidence says AI projects struggle most, and it has no data science team to absorb a mistake.

The trade-off cuts both ways. A small business can't afford many failed pilots, because it usually only gets to run one or two, and a stalled project eats the owner's attention as much as the budget. But it also has fewer systems, shorter approval chains, and the person who knows the work is often a few desks away from the person paying for the project. Several of the five failures above are much easier to prevent in a 40-person company than in a 40,000-person one, if somebody does the work up front.

What a Failure Looks Like Up Close

Picture a 45-person commercial pest control company (an illustrative composite) with 28 technicians on routes, a four-person office, and an owner who keeps hearing that competitors are "using AI." The owner signs up for an AI phone and booking assistant at about $1,200 a month, because inbound calls feel like the bottleneck and the demo was impressive.

Six months later the assistant answers calls politely and books almost nothing. It can't see the routing software, and the company has never recorded how long each type of job actually takes, so it offers appointment windows the technicians can't reach, and the office manager ends up calling customers back to reschedule. Nobody measured how many calls the office was missing before the project started, so nobody can say whether the assistant helped. The owner cancels it and tells a peer group that AI doesn't work for pest control.

Run the same business through the five reasons and every one of them shows up. The problem was chosen by gut rather than measured (reason one). The tool sat outside the routing system (two). It needed real job durations to offer workable windows, and the company had never recorded them (three). There was no baseline for missed calls and no cost ceiling (four). And nobody in the office owned making it work, so the office manager treated it as one more thing to clean up after (five).

When the company later timed its office work, the biggest drain turned out to be technicians' service reports and renewal follow-ups, which took the office about 22 hours a week to retype and chase. That's a back-office workflow, the kind MIT found tends to pay back faster, and the records it needs already sit in the company's own system. Run against the five checks, the second attempt passes all of them. The target is measured at 22 hours, and the output lands back in the routing software the office already works in. The last 20 service reports are pulled to test against, there's a monthly cost ceiling with a 90-day review date, and the office manager is named as owner, with two hours a week set aside for it.

How to Keep a Project Off the Failure List

Most AI project failures are decided before the project starts, so the most effective prevention is a short set of checks at the beginning. None of them requires technical skill. They require someone to measure the work, write down what success means, and name who owns it. Each check maps to one of the five reasons above.

Before you commit, check that Prevents What "passing" looks like
The target workflow is named and measured The wrong problem "Quote write-ups take 14 hours a week across two estimators"
The tool will work inside the system the work already uses No fit with the workflow Output lands in the job system, not a separate screen
The records the task needs exist and can be reached Missing data 20 real records pulled and checked, not sample data
There's a baseline, a target, and a cost ceiling Undefined value Before number, success number, monthly budget, and a stop date
One named person owns adoption after launch No owner That person's name, and time set aside for the job

The first check does most of the work. If you can't name the workflow and put a number on what it costs you today, every later step is guesswork, and that's the condition RAND found behind most failures. Our guide to AI readiness assessment services covers how a structured diagnostic answers these questions across several workflows at once, and the readiness cluster includes a DIY version if you'd rather run it yourself.

If you're already partway into a project that's stalling, the same five checks work as a diagnosis. Find the first one that fails, because the later ones usually fail as a consequence of it. Our page on recovering a failing AI project walks through that triage, and AI pilot purgatory covers the specific case of a pilot that never ends.

When Stopping a Project Is the Right Result

Not every abandoned AI project is a failure. A pilot that was scoped with a clear target, tested against real work, and stopped because it didn't clear the bar has done its job, since it cost a few weeks instead of a year and told you something true about your operation. Some of the abandonment in the statistics above is exactly that, and the problem isn't that companies stop projects. It's that so many of them stop after spending money they never meant to risk, on projects they never defined well enough to judge.

The practical difference is whether the stop rule existed before the project started. Write down the number that would make you keep going and the date you'll check it, and a stopped pilot becomes a decision rather than a failure. Leave those out, and the project tends to drift until someone cancels it out of frustration, which is how a single bad experience turns into a belief that AI doesn't work for your kind of business.

Where Avolis Fits

Every failure pattern above is set in motion before the build, so that's where we start. Every engagement begins with the diagnostic, in which we work through the candidate workflows with the people who run them, measure volumes, time per task, and handoffs, and rank what's worth building by impact and effort. The output is a ranked map with a clear go, wait, or no on each workflow, and you keep it whatever you decide next. From there we build the first workflow inside the systems your team already uses and stay embedded while it settles in, so the five checks above don't end at launch.

If you're comparing partners, our guide to which AI consulting company to choose sets out what to ask any firm before you sign. For what a good assessment should cover, start with AI readiness assessment services.

Frequently Asked Questions

What percentage of AI projects fail?

Published figures range from about 30% to 95% because studies count different things. S&P Global found 42% of companies abandoned most AI initiatives before production in 2025, up from 17%, and the average company scrapped 46% of its proofs of concept. MIT's 95% counts organizations seeing no measurable return from generative AI, not individual projects.

Why do most AI projects fail?

Most AI projects fail because of decisions made before the build rather than technical faults: solving the wrong problem, not fitting the tool into the existing workflow, missing data for that task, undefined cost and value, and no owner for adoption. In RAND's 2024 study, 84% of interviewed practitioners named a leadership-driven cause as the primary reason.

What did MIT find about why AI projects fail?

MIT NANDA's 2025 report found that 95% of organizations saw no measurable return from generative AI, and blamed a "learning gap": tools that don't retain feedback, adapt to context, or fit daily workflows. It also found budgets favor sales and marketing even though back-office automation often paid back faster.

What does Gartner say about AI project failure?

Gartner predicted at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, weak risk controls, rising costs, and unclear value, and over 40% of agentic AI projects canceled by 2027. Its widely quoted "85%" was a 2018 prediction about erroneous outcomes.

Are AI projects riskier for small businesses?

No study reports a reliable failure rate for firms of 10 to 200 people, since most data comes from large enterprises. Small firms can afford fewer failed pilots and lack technical staff, but they also have simpler systems and shorter decision chains, which makes the main causes of failure easier to prevent up front.

Continue Learning

AI projects don't usually fail because the technology let them down. They fail because nobody decided, before spending money, which work should change, how they'd know it had, and who would own it afterward. The statistics make the problem look like a coin flip you can't influence, but most of the outcome is set by choices you make in the first few weeks.

Before you start your next AI project, write down the one number it's supposed to move.

Why projects fail, in detail:

When a project is already in trouble:

Starting the right way:


Sources

All sources retrieved 2026-10-08.

On comparing failure rates. The figures in the table and the first chart measure different units (organizations, projects, pilots), use different definitions of failure, and come from different years and samples. Some are surveys of what happened and some are analyst predictions. We list them together to show the differences, not to average or rank them.

On the MIT funnel. MIT reports 60% investigated, 20% piloted, and 5% successfully implemented as shares of organizations, for embedded or task-specific generative AI tools. The "one in four" figure is our own arithmetic (5 divided by 20) and assumes, as the report's funnel implies, that the organizations reaching implementation are a subset of those that piloted.

On MIT's report. The copy linked above is the widely circulated July 2025 PDF of the report. We read it in full and cite page numbers from it. The report gives two different figures for the share of generative AI budgets going to sales and marketing (about 50% in its takeaways, about 70% in one survey passage), so we use the lower, more often repeated one.

On the IDC figure. We could not retrieve IDC's underlying research, which is not public. The 88% and "4 of every 33" figures are as reported by CIO, which attributes them to IDC research conducted in partnership with Lenovo.

On small-business abandonment. We looked for a government or central-bank survey that measures firms that tried AI and stopped, covering the Census Bureau's Business Trends and Outlook Survey, the Federal Reserve's Small Business Credit Survey, and Statistics Canada's Canadian Survey on Business Conditions, and found none. Vendor surveys that report year-over-year drops in small-business AI use compare different samples rather than tracking the same firms, so we haven't used them.

On the pest control company. It's an illustrative composite, not a client. Its figures (45 people, $1,200 a month, 22 hours a week) 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

Ready to make AI work for you?

Book an AI readiness evaluation. If there’s nothing worth automating, we’ll tell you.