Why Do Most Enterprise AI Projects Fail? What Scale Does to a Good Idea

Most enterprise AI projects fail because of what a large company does to a project, not because of what the AI can do. Big organizations launch many pilots at once, choose them far from the people who do the work, steer the budget toward results that are easy to show the board, and then ask each pilot to connect to hundreds of systems and clear several layers of review before it can touch a real workflow. MIT's 2025 research found that enterprises (firms with more than $100 million in revenue) led in pilot count and put more staff on AI than smaller firms did, yet reported the lowest rate of turning pilots into scaled systems (Challapally, Pease, Raskar, and Chari, The GenAI Divide, MIT NANDA, July 2025, p. 7).

The general reasons AI projects fail, from choosing the wrong problem to leaving nobody in charge of adoption, are covered in our guide to why AI projects fail. What's specific to size is a set of mechanics that let a sensible idea at a 40,000-person company die in a steering committee, and those mechanics matter for an operation of 10 to 200 people too.

That matters more than it sounds. The vendors, frameworks, and AI programs built for large companies are now being sold to mid-size ones, and a lot of what fails at the enterprise level fails because of habits a smaller business would be copying on purpose.

Key Takeaways

• Most enterprise AI projects fail for reasons that come from size: too many pilots, chosen far from the work, judged by what's easy to report, and slowed by integration and review.
• MIT's 2025 research found enterprises over $100 million in revenue ran the most pilots and staffed AI most heavily, yet had the lowest pilot-to-scale conversion, taking nine months or longer where top mid-market firms took about 90 days.
• Asked to split a hypothetical $100 of generative AI spending, executives in MIT's research put roughly half into sales and marketing, because those results are easy to measure, while back-office automation often returned more.
• The average enterprise runs 957 applications and only 27% of them are integrated, according to a 2026 survey of 1,050 IT leaders, so every AI project starts with a connection problem.
• In Writer's 2025 survey of 800 US executives, 68% said generative AI had created tension between IT and the rest of the business, and 72% said their company builds AI applications in silos.
• A 10-to-200-person operation shouldn't copy the enterprise playbook wholesale. Copy the written strategy and the accountability, and skip the pilot portfolio, the platform license, and the committee.

Table of Contents

• Why do most enterprise AI projects fail? The short answer
• More pilots, more staff, fewer results
• Six reasons enterprise AI projects fail
• Why AI projects fail at Fortune 500 companies
• What a mid-size operation should copy, and what it should skip
• A 140-person manufacturer that ran AI like a Fortune 500 company
• How to avoid the enterprise failure pattern
• Where Avolis fits
• Frequently Asked Questions
• Continue Learning

Why Do Most Enterprise AI Projects Fail? The Short Answer

Most enterprise AI projects fail because a large organization adds distance between the decision to build something and the work it's meant to change. A central team picks the use case, a separate IT group has to connect it, a risk or legal function has to approve it, and the people who would use it every day hear about it last. Each handoff is reasonable on its own, but together they stretch a pilot's timeline into quarters, and a pilot that takes three quarters to reach a real workflow often loses its sponsor, its budget, or its relevance before it gets there.

MIT's researchers put the cause in plain terms. After studying more than 300 public AI initiatives and interviewing people at 52 organizations, they concluded that the main barrier to getting value wasn't integration or budget but organizational design, and that companies succeeded when they gave implementation authority to the people close to the work while keeping clear accountability for results (MIT NANDA, p. 19).

More Pilots, More Staff, Fewer Results

Enterprise AI failure happens even in the organizations with the most resources. MIT found that enterprises ran more pilots and assigned more people to AI than mid-market firms, and that this intensity didn't translate into success. Top-performing mid-market companies reported about 90 days from pilot to full implementation, while enterprises took nine months or longer (MIT NANDA, p. 7).

The same research found where enthusiasm runs out. MIT reported that budgets and excitement were usually enough to launch a pilot, but that converting pilots into systems built into the workflow, with value that lasted, was rare (MIT NANDA, p. 7). Money opens the door to a pilot and does very little to carry it the rest of the way.

That gap between launching and finishing is the shape of enterprise failure. A large company can afford to start almost anything, so it starts a lot, and its real constraint becomes the narrow pipe that every pilot has to pass through to reach production: the same integration team, the same security review, the same handful of managers whose sign-off each project needs. Our page on why AI consulting projects fail to reach production covers that last stretch from the buyer's side.

Six Reasons Enterprise AI Projects Fail

Enterprise AI projects fail for six reasons that grow with the size of the organization. Each one exists at a smaller company too, but at a smaller company it's usually a conversation, and at a large one it becomes a structure.

1. Too many pilots, chosen too far from the work

Large companies tend to run AI as a portfolio, with a central team gathering use cases from across the business, scoring them, and funding a batch of pilots at once. That approach feels disciplined, but it separates the people choosing projects from the people who know which tasks actually drain their week. MIT found that the most effective buyers did the opposite: they stopped waiting for central approval and let the managers closest to the work propose and drive AI projects instead of leaving that to a central lab, while holding the vendor and the team to business metrics (MIT NANDA, pp. 18 to 20).

A portfolio also spreads attention thin. Writer's 2025 survey of 1,600 US knowledge workers, run with the research firm Workplace Intelligence, found that 72% of the 800 executives surveyed said their company develops AI applications in silos (Writer, "Key findings from our 2025 enterprise AI adoption report", March 2025). Ten pilots in ten departments produce ten partial answers, and no one is in a position to finish any of them.

2. Budgets go where results are easy to show

MIT asked executives how they would divide a hypothetical $100 of generative AI spending across business functions, and roughly half went to sales and marketing. The researchers' conclusion was that back-office automation in finance, procurement, and operations often returned more, and that the money followed sales and marketing because its results were easier to attribute: email response rates and demo volumes line up neatly with numbers a board already watches (MIT NANDA, pp. 9 to 10).

A procurement executive at a Fortune 1000 pharmaceutical company described the problem from the other side. If a tool helps a team work faster, how does that person justify it to a CEO when it won't directly move revenue or a cost line (MIT NANDA, p. 10)? Faster month-end closes and fewer compliance misses are real gains, but they're hard to put in an investor update, so they get funded last.

MIT found one exception. Manufacturers and healthcare providers in its interviews put minimal investment into sales and marketing and over-indexed on operations (MIT NANDA, p. 10). The businesses closest to physical work were already pointing their AI at the work.

3. Every project has to connect to hundreds of systems

An AI tool is only useful if it can read the information a task needs and put its result where the next person will pick it up. In a large company, that information is spread across an enormous number of applications that don't talk to each other. Salesforce's 2026 Connectivity Benchmark Report, built on a survey of 1,050 IT leaders conducted with Vanson Bourne, found that the average enterprise ran 957 applications, up from 897 a year earlier, and that only 27% of them were integrated (Salesforce, "Salesforce Announces 2026 Connectivity Report", February 2026).

The same survey found that 96% of organizations hit barriers when using their data for AI, and that legacy infrastructure or incompatible systems (37%) and integrating siloed apps and data (35%) were among the main hurdles holding back AI agents. Many enterprise pilots get around this by running on a clean extract of data pulled out by hand, which produces a result before the integration work is funded and leaves the real connection for later.

4. IT and the business pull in different directions

In a large company, the people who own the technology and the people who own the work report to different executives, measure different things, and often disagree about what AI is for. Writer's survey put numbers on it. Sixty-eight percent of executives said generative AI had created tension or division between IT and other parts of the business, and 42% said the process of adopting it was "tearing their company apart" (Writer, March 2025).

Executives and employees didn't even agree on whether it was working. Seventy-five percent of the executives in Writer's survey believed their organization had successfully adopted generative AI over the past year, against 45% of employees. Executives are the ones who approve projects and employees the ones who use them, and when those two groups see different results, nobody acts on the gap, and the project carries on in a state that's neither working nor cancelled.

5. Review layers slow projects until momentum runs out

Large companies carry real risk from AI: customer data, regulators, shareholders, and a brand that a single bad output can embarrass in public. Their answer is review, and review takes time. Even for AI agents, risk management, compliance and security, or legal implications was the most-cited hurdle in Salesforce's 2026 survey, named by 42% of IT leaders. MIT's interviews with procurement and IT sourcing leaders, backed by public procurement records, found enterprise cycles from request for proposal to implementation running as long as 18 months, with review one of the steps inside that cycle (MIT NANDA, p. 18).

None of that review is wrong. The trouble is what happens to a pilot during it. The sponsor moves to a new role, the business case was written for last year's priorities, and the vendor has shipped two new versions of the tool. Companies in banking, insurance, and healthcare face the heaviest version of this, which our page on why AI projects fail in regulated industries covers in detail.

6. Big programs carry big-project odds

When an enterprise turns AI into a transformation program, it inherits the track record of every large technology program before it, and that record is poor. McKinsey and the University of Oxford studied more than 5,400 IT projects and found that large ones (initial budgets above $15 million) ran 45% over budget and 7% over schedule on average, while delivering 56% less value than predicted. Every additional year a project was scheduled to run increased its cost overrun by 15% (Bloch, Blumberg, and Laartz, "Delivering large-scale IT projects on time, on budget, and on value," McKinsey, October 2012).

The study predates generative AI, but its finding on time applies to any long technology program. Time is the enemy of a technology project, and the enterprise habits above (portfolios, integration queues, review cycles) are all ways of adding time. An AI program planned as a two-year transformation starts with the odds of a two-year IT project, whatever the model underneath can do.

Why AI Projects Fail at Fortune 500 Companies

AI projects at Fortune 500 companies fail for the same six reasons, with two pressures turned up further: public scrutiny and a flood of vendors. A Fortune 500 company answers to shareholders and regulators in writing, so AI shows up in its filings as a risk before it shows up in its operations as a result. Arize AI's review of annual reports, as reported by Fortune, found that 281 Fortune 500 companies cited AI as a risk factor in their latest annual reports, up from 49 in its earlier count (Fortune, "Over half of Fortune 500 companies cite AI as a business risk", August 2024). A company that has told investors AI is a risk will, sensibly, review every AI project as one.

The vendor flood works against them too. A head of procurement at a major consumer goods company told MIT's researchers that they received many emails a day from vendors claiming to have the best generative AI product, and that with so many impressive demos, the company leaned on peer recommendations to decide whom to trust (MIT NANDA, p. 10). When choosing a vendor becomes a trust exercise, the safest-looking option wins, and the safest-looking option is often a broad platform that does many things adequately and no single workflow well.

MIT also observed far more internal build efforts than purchased or partnered ones in its sample, and the internal builds succeeded about half as often (MIT NANDA, p. 19). Large companies have engineering teams, so building in-house feels like the controlled choice, but in MIT's sample it was the less reliable one.

What a Mid-Size Operation Should Copy, and What It Should Skip

A mid-size operation should copy the enterprise habits that create accountability and skip the ones that create distance. Most enterprise AI practices travel badly to a business of 10 to 200 people, but three are worth keeping in some form.

Writer's survey found that at companies without a formal AI strategy, only 37% of executives reported being very successful at adopting AI, against 80% at companies with one. The lesson is cheap to apply. For a 60-person company, a strategy is a single page that names the work, the person who owns it, and how you'll know it worked.

A 140-Person Manufacturer That Ran AI Like a Fortune 500 Company

Take a 140-person precision parts manufacturer (an illustrative composite) whose largest customer is an aerospace prime with a well-publicized AI program. The owner attends the customer's supplier day, hears about its AI center of excellence, and comes home determined to do the same. The company forms an AI committee of six managers, collects 22 use-case ideas from across the business, scores them, and funds five pilots at once: AI-drafted sales outreach, marketing content for the website, a customer-facing chatbot, a production scheduling optimizer, and a tool to help the accounts payable team. It also signs a one-year enterprise license for a broad AI platform, because the platform covers all five.

Seven months later, the committee has met monthly and every pilot has a status slide. The sales and marketing pilots are being used by two people at their own desks, and the drafts are fine, but nobody can connect them to an order. The chatbot answers questions on the website but can't see order status, because order data lives in the ERP and the integration was never funded. The scheduling optimizer ran well on a spreadsheet export and stalled when the plant manager pointed out it didn't know about the two machines that are always down for maintenance on Tuesdays. The accounts payable pilot never started, because it ranked fifth on the committee's scorecard and the AP lead couldn't spare time for the kickoff.

The pilots were chosen by a committee rather than by the people who knew where the hours went, the budget favored the work that was easy to present, the integration that would have made the chatbot useful sat in a queue, and the whole program was planned on a timeline that let the year run out.

The restart is deliberately smaller. The owner asks where the office actually loses time, and the answer turns out to be the AP pilot that never began. Three clerks spend most of their week matching about 1,800 supplier invoices a month against purchase orders and receiving records, chasing mismatches by email, and the company misses early-payment discounts because invoices sit in the queue. So the project gets one job: match each invoice to its purchase order and receiving record inside the ERP, approve clean matches for the AP lead's daily review, and send mismatches to a short exception list the AP lead owns. Before anything is built, the team records invoices processed per clerk-hour and days from receipt to payment. The owner and the AP lead review the numbers together every two weeks, the platform license isn't renewed, and the next workflow waits until this one has a result.

How to Avoid the Enterprise Failure Pattern

You avoid the enterprise failure pattern by keeping the decision, the work, and the result close together. For a mid-size operation, that comes down to six choices you make before you spend anything.

1. Run one workflow, not a portfolio. Pick the single workflow that drains the most hours and finish it before starting a second. A finished project teaches the business more than five pilots in progress.
2. Rank by hours and dollars, not by visibility. Back-office work such as invoicing, purchasing, scheduling, and document handling often returns more than customer-facing projects, even though it's harder to show off.
3. Name the system of record first. Decide which system the AI reads from and writes back to, whether that's the ERP, the job management software, or the accounting package. If that connection isn't in the plan, the project is a demo.
4. Give the decision to the person who runs the work. The AP lead, the dispatcher, or the production planner should have the authority to say what the tool must do and when it's good enough. The owner keeps accountability for the result.
5. Write a one-page risk check instead of a committee review. Note what data the tool touches, who reviews its output before it reaches a customer or a supplier, and what happens when it's wrong.
6. Put a 90-day clock on it. Set a baseline before you start and a date to decide whether the project stays. MIT's top mid-market performers reached full implementation in about 90 days, and a narrow workflow in a smaller business should be held to a similar pace.

If the project in front of you involves AI agents, meaning tools that take actions across systems on their own rather than drafting work for a person to check, the integration and review problems above get sharper. Our page on why agentic AI projects fail covers that case. If a project is already stalled, our guide to recovering a failing AI project covers how to decide whether to fix, narrow, or stop it.

Where Avolis Fits

The enterprise failure pattern is a gap between the people who decide on AI and the people who do the work, and our method is built to close it. Every engagement starts with the diagnostic. Working with the people who actually do the purchasing, scheduling, or billing, we measure how each workflow runs (how much comes through, how long each step takes, where work changes hands) and rank the workflows by what each would return against what it would cost to build. The ranking belongs to you, and it's useful whether you build the first workflow with us, with someone else, or not yet.

From there, we build the first workflow inside the systems your team already uses, with the person who runs that work deciding what good looks like, and we stay embedded while it settles in and while the next one gets built. That's the enterprise's discipline (a written plan, a baseline, accountability for the number) without the committee, the portfolio, or the platform license. To see what a good diagnostic should cover, read our guide to AI readiness assessment services. If you're comparing partners, our guide to which AI consulting company to choose lists the questions to ask each one.

Frequently Asked Questions

Why do most enterprise AI projects fail?

Most enterprise AI projects fail because of how large organizations run them: many pilots at once, chosen far from the work, slowed by integration and review. MIT's 2025 research found enterprises over $100 million in revenue ran the most pilots but had the lowest pilot-to-scale conversion, taking nine months or longer to reach full implementation.

Why do AI projects fail at Fortune 500 companies?

Fortune 500 companies face the usual enterprise obstacles plus heavier scrutiny and a flood of vendors. Arize AI found 281 Fortune 500 companies cited AI as a risk factor in their annual reports, up from 49 in its earlier count, which tends to put every AI project through risk review, and long review cycles can outlast a pilot's sponsor and budget.

What percentage of enterprise AI projects fail?

There's no single reliable figure, because studies define failure differently. MIT's widely quoted 95% treats success as marked and sustained productivity or profit impact, which is a stricter test than whether a project was abandoned. Our guide to why AI projects fail shows what each common statistic actually counts.

Are smaller companies better at AI projects than enterprises?

They can be faster. MIT found top-performing mid-market companies took about 90 days from pilot to full implementation, against nine months or longer for enterprises. A smaller company has fewer systems and shorter approval chains, but it also has less room for a wasted attempt, so the advantage only holds if it chooses one workflow carefully.

Should a mid-size company copy an enterprise AI strategy?

Copy the parts that create accountability and skip the parts that create distance. A one-page written strategy, a baseline measure, and a named owner travel well. A portfolio of pilots, a steering committee, and a broad platform license usually don't, because they recreate the slow handoffs that make enterprise projects fail.

Continue Learning

Enterprise AI fails because size puts distance between the decision and the work, and every layer of that distance adds time a project can't afford. A mid-size operation has most of that distance removed by default, which is an advantage worth protecting rather than giving away by copying a big company's structure.

Before you form an AI committee or sign a platform license, name the one workflow you'd fix first and the person who runs it.

Why projects fail, in detail:

• Why do AI projects fail? Common reasons and statistics
• Why agentic AI projects fail
• Why AI projects fail in regulated industries
• Why are AI projects failing? What's different about the current wave

When a project is already in trouble:

• Why your AI consulting project isn't reaching production
• How to recover a failing AI project

Starting the right way:

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

---

Sources

All sources retrieved 2026-10-08.

• Aditya Challapally, Chris Pease, Ramesh Raskar, and Pradyumna Chari, The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025 (preliminary findings; 300+ public AI initiatives reviewed, interviews with 52 organizations, survey of 153 senior leaders; pages 7, 9, 10, 18, 19, and 20 cited), retrieved 2026-10-08: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
• Salesforce, "Salesforce Announces 2026 Connectivity Report," 5 February 2026, for MuleSoft's 11th annual Connectivity Benchmark Report (1,050 IT leaders in nine countries, double-anonymous online survey with Vanson Bourne, October to November 2025), retrieved 2026-10-08: https://www.salesforce.com/news/stories/connectivity-report-announcement-2026/
• Writer, "Key findings from our 2025 enterprise AI adoption report," 18 March 2025 (1,600 US knowledge workers using AI at work, including 800 C-suite executives and 800 employees, surveyed with Workplace Intelligence), retrieved 2026-10-08: https://writer.com/blog/enterprise-ai-adoption-survey/
• Michael Bloch, Sven Blumberg, and Jürgen Laartz, "Delivering large-scale IT projects on time, on budget, and on value," McKinsey & Company, 1 October 2012 (more than 5,400 IT projects, with the BT Centre for Major Programme Management at the University of Oxford), retrieved 2026-10-08: https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/delivering-large-scale-it-projects-on-time-on-budget-and-on-value
• Fortune, "Over half of Fortune 500 companies cite AI as a business risk," 19 August 2024, reporting Arize AI's analysis of Fortune 500 annual reports, retrieved 2026-10-08: https://fortune.com/2024/08/19/over-half-fortune-500-companies-cite-ai-business-risk-chatgpt/

On MIT's figures. MIT defines enterprises as firms with more than $100 million in annual revenue. Its 90-day and nine-month timelines are averages reported by interviewees. The budget question asked executives to allocate a hypothetical $100 across functions; MIT's summary puts sales and marketing at about 50%, the body text gives approximately 70% for the same survey, and MIT describes the sub-category breakdown as directional. The build-versus-partner comparison comes from its interview sample of 52 organizations, and MIT notes the difference may reflect the organizations rather than the approach.

On the vendor surveys. Salesforce (which owns MuleSoft) and Writer both sell enterprise AI and integration products, and their surveys cover enterprise IT leaders and knowledge workers. The figures are used here as descriptions of enterprise conditions, not as measures of project outcomes. Writer's success comparison (37% versus 80%) is self-reported by executives.

On the McKinsey and Oxford study. It covers IT projects in general and was published in 2012, before generative AI. It's cited for how project size and duration affect results, not as a measure of AI project failure.

On the Arize figures. Arize AI sells AI observability software. The 281 and 49 counts are as reported by Fortune; Arize's headline growth figure (473.5%) is measured against 2022 filings.

On the manufacturer. It's an illustrative composite, not a client. Its figures (140 people, six managers, 22 ideas, five pilots, seven months, three clerks, about 1,800 invoices a month) 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