MIT: Why Most Enterprise Gen AI Pilots Still Fail
Global enterprise spending on Gen AI has surged to between US$30bn and US$40bn, yet 95% of organisations admit they have seen no measurable return on these investments.
A new MIT study, The Gen AI Divide: State of AI in Business 2025, analysed 300 public use cases and highlights what researchers term the “Gen AI Divide”: a gap between the small minority extracting significant value and the majority stuck in unsuccessful pilot schemes.
The report shows just 5% of embedded AI pilots deliver demonstrable P&L impact, with the divide driven largely by execution strategy rather than model sophistication or regulatory hurdles.
In parallel, off-the-shelf platforms such as ChatGPT and Microsoft Copilot are widely adopted, with more than 80% of enterprises are trialling them and nearly 40% already deploying.
Yet their impact remains confined to boosting individual productivity rather than transforming organisational outcomes.
Custom enterprise solutions tell a different story. While 60% of companies are assessing bespoke or vendor-led systems, only 20% progress to pilots and a mere 5% succeed in achieving production rollout.
The majority stumble due to fragile workflows and poor alignment with daily operations.
Why the “shadow AI economy” bypasses enterprise failures
The study reveals the rise of a vibrant “shadow AI economy,” where employees rely on personal subscriptions for professional tasks without corporate sanction.
According to the research, staff in more than 90% of surveyed companies report frequent use of personal AI tools at work, even as only 40% of organisations provide official enterprise licences.
One case in point is a corporate lawyer whose firm invested US$50,000 in a specialist contract analysis platform. Despite this, she routinely turns to ChatGPT for drafting.
“Our purchased AI tool provided rigid summaries with limited customisation options,” she says. “With ChatGPT, I can guide the conversation and iterate until I get exactly what I need.”
She also outlines clear guardrails:
“It’s excellent for brainstorming and first drafts, but it doesn’t retain knowledge of client preferences or learn from previous edits,” she explains.
“For high-stakes work, I need a system that accumulates knowledge and improves over time.”
These insights underscore the knowledge gap undermining enterprise AI efforts and widening the divide.
User sentiment further reflects this split: while 70% prefer AI for fast, routine tasks, 90% turn to humans for complex projects demanding sustained expertise.
How external partnerships double success rates
MIT finds that strategic partnerships with external vendors reach deployment 67% of the time compared to 33% for internal builds.
- AI will replace most jobs soon: Layoffs are limited and mostly industry-specific and executives remain divided on future hiring
- Gen AI is transforming business: Adoption is high, but only 5% of firms scale AI into workflows – most sectors show little change
- Enterprises are slow adopters: In reality, 90% have seriously explored AI purchases, showing eagerness, not hesitation
- Model quality and regulation are the barriers: The real issue is poor workflow integration and tools that don’t learn or adapt
- The best AI tools succeed on their own: Success comes when tools are customised, integrated, and tied to measurable outcomes
Leading adopters report implementation cycles as short as 90 days, compared with the nine months or more typically needed by larger enterprises.
Investment priorities are also misaligned: nearly 50% of AI budgets flow into sales and marketing, even though back-office automation consistently delivers higher returns.
Organisations with effective deployments cite annual savings of between US$2m and US$10m from reduced business process outsourcing, a 30% cut in external creative costs and around US$1m saved on outsourced risk management.
As one VP of Procurement at a Fortune 1000 pharmaceutical firm told MIT researchers:
“If I buy a tool to help my team work faster, how do I quantify that impact?
“How do I justify it to my CEO when it won’t directly move revenue or decrease measurable costs?”
Top performers report 90-day implementation cycles while enterprises typically require nine months or longer.
Meanwhile, investment patterns reveal misaligned priorities – sales and marketing capture 50% of AI budgets despite back-office automation often yielding higher returns.
Successful deployments report US$2m-US$10m annual savings through business process outsourcing elimination, 30% reduction in external creative costs and USD$1m saved on outsourced risk management.
A VP of Procurement at a Fortune 1000 pharmaceutical company told MIT researchers: “If I buy a tool to help my team work faster, how do I quantify that impact?
“How do I justify it to my CEO when it won’t directly move revenue or decrease measurable costs?”
MIT’s findings show that the most successful enterprises approach AI vendors as business service partners rather than traditional software suppliers, insisting on deep tailoring and measuring success against operational results.
Top-performing Gen AI startups achieve annualised revenues of US$1.2m within six to 12 months by first focusing on narrowly defined workflows before scaling outwards.
A CIO at a US$5bn financial services company illustrates the selection process: “We’re evaluating five different Gen AI solutions, but whichever system best learns and adapts to our specific processes will ultimately win our business,” he says.
“Once we’ve invested time in training a system to understand our workflows, the switching costs become prohibitive.”
The procurement challenge is echoed by a Head of Procurement at a major consumer goods firm:
I receive numerous emails daily claiming to offer the best Gen AI solution,” she says.
“Some have impressive demos, but establishing trust is the real challenge. With so many options flooding our inbox, we rely heavily on peer recommendations and referrals from our network.”
Another CIO in the study is frank about the noise: “We’ve seen dozens of demos this year. Maybe one or two are genuinely useful. The rest are wrappers or science projects.”