Skip to main content
Market research

← Back to Topics

Why AI projects stall, and what works

What studies and surveys say about why AI projects fail to reach production or show a return, which factors recur, and where the evidence is weak.

In short

Nineteen studies from consultancies, research institutes and software vendors were read for the same question: why do AI projects stall, and what separates those that deliver? The most repeated answer is about workflow: projects succeed when AI is built into how the work is actually done, and fail when it is bolted on. Headline failure rates differ widely between studies, and the best-known one is contested.

Factors that recur

Tally made for this research: each of the 19 studies read was checked for whether it names a factor, and a study counts if its text names it. The counts are ours, not the studies'.

Factor namedStudies naming it (of 19)
Fit with existing workflows, or redesigning the process (success when present, failure when absent)13
Risk, governance, security and compliance9
Clear choice of use case or roadmap7
Gap between pilot and production7
Skills and training gaps7
Change management and adoption6
Data quality6
Executive sponsorship6
Value not measured, or unclear return5 to 6
Output quality or accuracy4
Involvement of outside partners (direction differs between studies)4

The main studies

StudySampleWhat it reports
MIT NANDA, "The GenAI Divide" (July 2025, preliminary)300+ public initiatives, 52 interviews, 153 leaders"95% of organizations are getting zero return." Only 5% of enterprise-grade custom tools reached production. External partnerships reached deployment about 67% of the time, against about 33% for internally built tools
RAND (August 2024)65 data scientists and engineersCites estimates that more than 80% of AI projects fail. The most common cause is misunderstanding or miscommunicating what problem the project is for
McKinsey, State of AI 20251,993 respondents, 105 countriesNearly two-thirds have not begun scaling AI across the enterprise. Workflow redesign is among the factors most strongly linked to business impact; high performers were 2.8 times as likely to have redesigned workflows (55% against 20%)
Gartner (June 2025, a prediction)Poll of 3,412 webinar attendees"Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls"
BCG, "Where's the Value in AI?" (October 2024)Not stated on the pageOnly 22% of companies have moved beyond proof of concept and only 4% create substantial value
BCG, "The Widening AI Value Gap" (September 2025)1,250 firms5% achieve AI value at scale; 60% report no material value. BCG recommends a 10-20-70 split: 70% of effort on people and processes, 20% on technology, 10% on algorithms
S&P Global / 451 Research (late 2024 fieldwork)1,006 IT and line-of-business professionals42% of companies abandoned most of their AI initiatives before production, up from 17%; on average 46% of proofs of concept were scrapped
Deloitte, State of AI in the Enterprise (January 2026)3,235 leaders, 24 countriesClose to three-quarters of companies plan to deploy agentic AI within two years; only 21% of them report a mature model for agent governance. Revenue growth from AI "largely remains an aspiration"
Bain generative AI survey (May 2025)199 US companiesMore than 80% of use cases meet or exceed expectations. Companies further along cite low-quality vendors, which are hard to spot early
Thomson Reuters, Future of Professionals 20252,275 professionalsOrganizations with a visible AI strategy were almost four times as likely to report AI benefits
Salesforce, SMB Trends (6th edition)3,350 leaders at firms with up to 200 employeesSecurity is the biggest concern (41% of AI-using SMBs). Growing SMBs were 1.8 times as likely to invest in AI
Clio, Legal Trends 2025Clio user dataFirms with wide AI adoption were nearly three times as likely to report revenue growth

Also read: Deloitte's January 2025 report, Gartner's 2024 prediction that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, Bain's 2025 Technology Report, Stanford's 2026 AI Index, Intuit QuickBooks' April 2025 small-business survey (68% of firms with up to 100 employees use AI regularly), a 2026 survey by the AI agency Creative Genius (312 businesses), and a desk analysis of 123 deployments by TechnBrains. The last two are published by AI service vendors and are treated as weak evidence.

What the studies say works

  • Fit the workflow. McKinsey, BCG, Deloitte, Bain and MIT NANDA all point to process redesign over adding a tool. The Creative Genius survey says the businesses with no measurable impact "skew heavily toward businesses that bought AI tools (ChatGPT Team, Microsoft Copilot) without building workflows around them".
  • Pick a problem and measure it. RAND finds that successful projects "are laser-focused on the problem to be solved, not the technology used to solve it." BCG and Deloitte link continued funding to demonstrated value.
  • Control the risk. Governance, security and compliance are a recurring factor. Deloitte found that close to three-quarters of companies plan to deploy agentic AI within two years, but only 21% of them report a mature governance model for agents.
  • Use outside help, with care. MIT NANDA and BCG find better outcomes with partners (BCG: companies that use the ecosystem are three times as likely to experiment with agents). Bain notes that production-stage companies cite low-quality vendors.
  • Keep a person in the loop. The TechnBrains analysis found that 48% of documented implementations still require human approval before AI output triggers action, and only 15% run fully autonomously.

How strong is the evidence

  • The 95% figure is contested. The MIT NANDA report calls its figures "directionally accurate based on individual interviews rather than official company reporting". It is preliminary and not peer-reviewed. Critics note that it relies on publicly announced initiatives, which over-represent large programs, and that its success test was profit-and-loss impact within about six months. The report itself says the success-rate differences "may reflect organizational capabilities rather than implementation approach alone." Its own text also reports about 83% pilot-to-implementation for general chatbots. The critiques come from the Marketing AI Institute and from an AI-automation agency blog.
  • Predictions are not measurements. Gartner's 40% is a forecast.
  • Samples skew large. Most of these surveys draw on large enterprises. Salesforce and Intuit are the main small-business samples, and both are vendor surveys.
  • Self-reported value. Return figures in Deloitte, Bain, Intuit and Thomson Reuters are reported by respondents.
  • No study of agency-built automations was found. Searches for outcome or churn studies of automations built by AI agencies returned only agency blog posts with no disclosed method. The closest evidence is the MIT NANDA partner comparison and Bain's note on vendor quality.

Sources

← Back to TopicsBack to top ↑