Why Most AI Projects Stall

In August 2025 a report from MIT’s NANDA initiative, “The GenAI Divide: State of AI in Business 2025”, put a number on something many owners had already sensed. Fortune summarised its finding this way: about 5% of AI pilot programmes achieve rapid revenue acceleration, while the vast majority stall, delivering little to no measurable impact on profit and loss. The research rests on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public AI deployments. That is a modest sample, so it is best read as a signal rather than a census.

What matters more is the diagnosis. The report places the problem in a “learning gap” for both tools and organisations, and Fortune notes that it points to flawed enterprise integration instead of weak models. The software, in other words, is generally capable of the task. The trouble starts when it meets a real business with real habits.

MIT is not alone. In March 2025 S&P Global Market Intelligence reported results from a survey of more than 1,000 respondents in North America and Europe. The share of companies abandoning most of their AI initiatives had risen to 42%, up from 17% the year before, and the average organisation had scrapped 46% of its AI proofs of concept before they reached production. Respondents named cost, data privacy and security risk as the main obstacles.

Gartner, in a June 2025 press release, predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. The same release describes what it calls “agent washing”, where vendors relabel chatbots and older automation as agents, and estimates that only about 130 of the thousands of agentic AI vendors are real.

A 2026 follow-up points in the same direction. In April 2026 Gartner published a survey of 782 infrastructure and operations leaders, fielded in November and December 2025. Only 28% of AI use cases fully succeeded and met their ROI expectations, and 20% failed outright. Among leaders who had seen a failure, 57% said their initiatives failed because they expected too much, too fast.

The Gartner 2026 survey offers a useful hint about where projects fail. Its authors say failures were most often seen in ambitious areas such as auto-remediation and self-healing infrastructure, where leaders expected AI to outperform what the tools can reliably deliver. Success clustered in mature, well-understood uses: 53% of leaders reported their AI wins in IT service management. The pattern is plain enough. Narrow, established tasks work, and sweeping ambitions tend to stall.

For an owner this translates into a simple discipline. Pick one process that already costs you time or money, such as answering the same ten questions from new enquiries, or chasing quotes that have gone quiet. Write down today’s number: how many hours a week, how many enquiries go unanswered overnight, how many days a quote waits for a reply. A project with a baseline can be judged. A project described as “becoming an AI-first business” cannot.

McKinsey’s March 2025 survey tested 25 attributes of organisations against reported EBIT impact from generative AI. It found that redesigning workflows had the biggest effect. Only 21% of respondents using generative AI said their organisations had fundamentally redesigned at least some workflows.

The 2026 edition, published in August, shows the gap persisting. Nearly three-quarters of the firms McKinsey classes as high performers report fundamentally redesigning workflows because of their AI use, up from 55% a year earlier, against just one quarter of other respondents. Gartner’s 2026 survey found the same thing from the operations side. Among leaders with at least one success, the main reasons given were integrating AI into existing workflows and systems and securing full support from business executives.

In practice this means the assistant should live where your work already happens: your calendar, your inbox, your booking system, your customer records. An impressive tool that sits in a separate tab, and needs someone to remember to open it, is a pilot waiting to be forgotten.

Fortune’s account of the MIT report draws a sharp line here. Purchasing AI tools from specialised vendors and building partnerships succeed about 67% of the time, it reports, while internal builds succeed only one-third as often. The same report found the largest returns in back-office automation, such as cutting external agency costs and streamlining operations, even though more than half of generative AI budgets went to sales and marketing tools.

There is a fair counterpoint. McKinsey’s 2026 survey finds that larger organisations are beginning to take more ownership of their technology, and that 32% of respondents have decided against buying some software because agentic coding tools let them build it internally. McKinsey is careful to say this does not mean building everything. The firms moving fastest are deliberate about where to buy, where to build and where to develop enough capability to integrate and scale what works. A business with a small team and no engineers is well served by starting with what already exists.

A system that answers customers or sends money-related messages will sometimes be wrong, and the cost of that depends on who sees it first. McKinsey’s 2026 report lists human-in-the-loop design among the practices that high performers use to sustain their deployments, alongside senior-leadership role modeling, impact measurement and risk management.

The practical form is modest. Let the assistant draft the reply, the quote or the post, and let a named person approve anything that commits the business to a price, a date or a promise. Routine, low-risk steps can run on their own once the approvals have shown a clean record. Gartner’s June 2025 warning about inadequate risk controls is partly about this: a project that cannot show who checks its output is a project that gets cancelled after the first bad incident.

McKinsey’s 2026 data show that high performers are twice as likely as others to say their organisations have defined processes to measure the impact of AI initiatives. The same report finds a wide gap between personal and organisational results: 80% of respondents say AI has improved their own productivity, yet only 37% of organisations report any positive EBIT contribution, essentially flat on the year before. Feeling faster does not show up on its own in the accounts.

For a small business the measurement can be a single page. Record the baseline before launch, check the same figure each week, and set a date, perhaps eight to twelve weeks out, when you will decide whether to extend, change or stop. That date protects a project from the slow drift that turns many pilots into abandoned ones.

The takeaway for an owner choosing where to begin is short. Choose one process, write down its current number, use a proven tool that connects to the systems you already run, keep a person on the approvals, and review the result on a fixed date. Projects that do this are unglamorous, and the evidence suggests they are the ones that survive.