AI pilots: only 5% of tools reach production, MIT finds

Most AI projects that never leave the pilot stall for management reasons, and the language model weighs little in that count. The most quoted figure on the subject comes from MIT Project NANDA’s 2025 report “The GenAI Divide”: only 5% of custom enterprise generative AI tools built for specific tasks reached production. In the report’s funnel, 60% of organizations evaluated this kind of tool, 20% reached the pilot stage and 5% reached production with sustained impact. The number is in the original, but it comes from 52 interviews and 153 survey responses, the authors call it directional, and the yardstick for return is six months after the pilot. Read next to the research from Gartner, RAND Corporation, BCG and McKinsey, the report shows percentages that vary widely and causes that repeat: a poorly defined problem, data that is not ready, AI placed beside the workflow instead of inside it, and projects with no owner accountable for the result.

At Iasbeck&Co we implement CRM and AI agents in B2B operations and train the teams that work with them, and many of my conversations with executive teams start with a pilot that worked in the demo and never became routine. I wrote this for the people who approve the AI budget, for the sales, marketing and CS leaders who will run the agents, and for HR, which usually gets called once the problem has turned into adoption.

How many AI projects never leave the pilot?

Each study measures something different, and the percentages do not add up. MIT measured the move from evaluated tool to tool in production, Gartner made predictions about abandonment after proof of concept, RAND interviewed technical professionals about causes of failure, and BCG and McKinsey asked executives whether the company already captures value. The table summarizes the six sources with the method of each.

Source	Year	Sample and method	Main finding
MIT, Project NANDA	2025	Review of more than 300 public initiatives, interviews at 52 organizations and a survey of 153 leaders at four conferences, January to June 2025	5% of task-specific generative AI tools reached production with sustained impact
Gartner	2024	Analyst prediction	At least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025
Gartner	2025	Survey of 1,203 data management leaders in July 2024, plus prediction	63% lack or are unsure whether they have the right data practices for AI; through 2026, 60% of projects without AI-ready data will be abandoned
RAND Corporation	2024	65 interviews (50 industry practitioners with at least five years of experience and 15 academics), August to December 2023	Five root causes; 84% of industry interviewees named leadership causes as the main one
BCG	2024	Survey of 1,000 executives in more than 20 sectors across 59 countries, self-assessing 30 capabilities	26% of companies have the capabilities to move beyond proof of concept and generate value; 74% have yet to show tangible value
McKinsey	2026	Online survey of 1,719 participants in 97 countries, May to June 2026, weighted by GDP	6% are AI high performers; about three quarters of them redesigned workflows, against one quarter of the rest

What did the MIT report measure?

Project NANDA’s report came out in July 2025, signed by four authors connected to the MIT Media Lab, and presents itself as “preliminary findings.” The sentence that went viral is in the executive summary: 95% of organizations would be getting zero return from the US$30 to 40 billion that companies invested in generative AI. The data behind it is narrower. In section 3.2, the authors separate two kinds of tool: general-purpose ones, such as chat assistants, were evaluated by 80% of organizations and implemented by 40%, while embedded or task-specific tools, where the corporate project lives, followed the 60%, 20% and 5% funnel.

How MIT defined success

In the methodology appendix, success means deployment beyond the pilot with measurable KPIs, and return is measured six months after the pilot. In the chart note, a tool “successfully implemented” is one that users or executives described as causing a marked and sustained impact on productivity or the bottom line. The yardstick depends on what the interviewee reported and expects within six months an effect that is hard to measure in that window. Next to the chart, the authors write that the numbers are “directionally accurate,” come from individual interviews rather than official company reporting, and that the definition of success changes from one organization to another.

The criticism the report deserves

The limitations are in the document itself, and almost no headline repeated them. The appendix acknowledges selection bias, since those who agree to talk about AI may differ from those who decline, and admits that six months may understate success for complex systems. There is an internal inconsistency: the opening of section 3.4 says 50% of the generative AI budget goes to sales and marketing, and the next paragraph mentions about 70% for the same exercise, in which executives split a hypothetical 100 dollars across functions. Project NANDA builds infrastructure for AI agents, and the report concludes that the way out is agentic systems with memory, which is the project’s own thesis. Finally, the PDF stayed at the MIT address until August 18, 2025 and then went offline, according to Internet Archive records, which is why the link points to the archived copy.

The report is still useful for anyone buying AI. In the sample, tools bought from external partners reached deployment about 67% of the time, against 33% for those built in house, and the authors warn that the correlation does not prove causation. Mid-market companies took an average of 90 days from pilot to full deployment, against nine months or more at large ones, and the most cited barrier was resistance to adopting new tools.

What does Gartner predict for projects that stop at proof of concept?

In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value. The deadline has passed and I did not find an open Gartner publication checking the result, so treat the number as a prediction.

In February 2025, the firm published a piece on AI-ready data. In a survey of 1,203 data management leaders conducted in July 2024, 63% of organizations said they did not have, or were unsure whether they had, the right practices, and Gartner predicted that through 2026 companies will abandon 60% of AI projects without AI-ready data. Analyst Roxane Edjlali sums up the criterion: data with issues is not ready for AI. The firm’s prediction for agents, with more than 40% of projects canceled by 2027, is discussed in the article on AI agents.

Why do AI projects fail, according to RAND?

The RAND Corporation study from August 2024, by James Ryseff, Brandon De Bruhl and Sydne Newberry, was written for the US Department of Defense and is the most useful of the six for understanding cause. The researchers interviewed 50 industry practitioners with at least five years of experience building machine learning models and 15 academics, and counted as a failure any project the organization itself perceived as failed. Five root causes came out of the conversations: leaders misunderstand or miscommunicate which problem AI should solve; the necessary data is missing; the team chases the newest technology instead of the user’s problem; infrastructure to manage data and deploy models is lacking; and the problem is too hard for AI.

Leadership causes and data causes came up spontaneously in more than half of the interviews as the main reason, and 84% of industry interviewees named at least one leadership cause as the main one. The other three appeared in a quarter to a third of the conversations. One RAND recommendation fits any budget committee: before starting, leaders should be willing to keep the team on the same problem for at least a year.

Where the 80% figure comes from

The RAND study helped spread the claim that more than 80% of AI projects fail, twice the rate of IT projects without AI. The text introduces the number with “by some estimates,” and the footnote leads to a Fortune article from July 2022 quoting an executive. RAND did not measure the 80%, and the number has a journalistic origin that predates the generative AI wave.

What are the causes, and which study supports each one?

The table brings together the causes that appear across the six sources and translates each into what I see in a sales operation running on a CRM.

Cause	How it shows up in the operation	Source
Poorly defined problem	The pilot tests “what AI can do” and nobody wrote down which metric needs to change	RAND (first root cause); BCG (leaders pursue about half as many opportunities as the rest)
Data not ready	The agent reads duplicate records, empty dates and stages filled in for the weekly report	Gartner 2025 (63% without adequate practices); Gartner 2024; RAND (second root cause)
AI outside the workflow	The tool runs on a separate screen with no memory of the customer, and the rep goes back to the old way	MIT (brittle workflows, no learning from context); McKinsey (workflow redesign at high performers)
No owner	The pilot belongs to IT and no business manager answers for the result	RAND (84% cite leadership causes); MIT (implementation authority in the business, accountability kept)
Adoption	The team approves the demo and keeps using a personal assistant or the spreadsheet	MIT (resistance to new tools); BCG (about 70% of challenges come from people and processes)
Cost and unclear value	Consumption grows in the third month and nobody can show finance the gain	Gartner 2024; McKinsey 2026 (one in five say operating costs constrained use)

The causes chain together: a poorly defined problem produces a pilot with no metric, and a pilot with no metric cannot prove value when the cost shows up. Of the six, only data infrastructure depends mainly on IT, and the others depend on decisions that executives and managers make before signing the purchase.

What do companies that take AI into production do?

The 2026 McKinsey survey calls high performers the companies whose respondents attribute at least 5% of EBIT to AI and describe the value as significant. They make up 6% of the sample, the same share as in 2025. About three quarters of them redesigned workflows because of AI, up from 55% the year before, and only a quarter of other respondents did the same. They are also twice as likely to report visible commitment from senior leaders. The broader market reading is in the article on the state of AI in companies.

BCG reached a similar conclusion in 2024. Of the companies surveyed, 4% generate value consistently and another 22% are starting to capture substantial gains. Leaders follow the rule of 10% of resources on algorithms, 20% on technology and data and 70% on people and processes, and pursue on average half as many opportunities as the rest.

Putting McKinsey, BCG and MIT together, the companies that scale choose a few processes and change the whole workflow of each one, instead of handing a tool to everyone. They name a business manager as owner and hold the vendor to business results, as if hiring a service provider. And they measure impact with a defined process, the subject of the article on AI ROI.

What does this have to do with your company’s CRM?

MIT shows that sales and marketing receive the largest share of the generative AI budget, and these are the areas where the agent reads and writes in the CRM. A proposal follow-up agent depends on a reliable date for the last interaction, an objective stage criterion and a recorded next step. A ticket triage agent depends on closed categories, customer history and a written priority rule. With those fields empty, the pilot works on the database the vendor prepared for the demo and fails on the real one.

An illustrative example: a B2B services company tests, for 60 days, an agent that writes follow-ups for open proposals. In the pilot, with 40 proposals the manager reviewed by hand, the drafts come out well. In the rollout, the agent starts reading 900 proposals, some duplicated, some in the wrong stage and many without the decision maker’s name. Reps start rewriting every draft, give up in the third week, and the project joins the statistics with the same agent that worked in the pilot.

That is why I treat database preparation as part of the AI project, with a deadline and an owner. The roadmap for that cleanup is in the article on CRM data quality, and CRM projects that failed for the same reasons are covered in why CRM projects fail.

What these data do not tell you

None of the six sources measures Brazil on its own. MIT’s appendix admits the sample may not represent every region, BCG covers Asia, Europe and North America, and McKinsey weights by GDP, which gives more weight to larger economies. A mid-sized Brazilian sales operation may have different rates, higher or lower, and nobody has published that number with an open method.

Definitions of failure also change: MIT measures return six months after the pilot, RAND accepts the organization’s perception and Gartner works with predictions. BCG and McKinsey rely on executive self-assessment, and the person answering is often the one who sponsored the project. The high-performer data show correlation, and companies that already make money with AI may have more room to redesign workflows. BCG and McKinsey sell transformation projects and Project NANDA builds infrastructure for agents, and each has an interest in the conclusion it presents.

Two numbers circulate without a verifiable primary source. The 80% failure figure reaches RAND through a 2022 news article. And a 2025 survey by a market intelligence firm on abandoned AI initiatives appears in many news stories, but I did not find the original report in open access and left the number out.

What leaders do with these numbers

1. Swap the 95% for the funnel. The 60%, 20% and 5% data and the list of causes help the decision more than the headline, which applies to a small sample and a six-month yardstick.
2. Choose one process, with a metric and a baseline. RAND, BCG and McKinsey recommend focusing effort on a few problems. Record the metric’s value in the CRM before the pilot.
3. Name an owner in the business. The sales, marketing or CS manager answers for the agent’s result, and IT handles access and security.
4. Audit the fields the agent will read. Check duplicates, dates, stage criteria and required fields for the chosen process, and fix what is missing within the project schedule.
5. Put AI inside the workflow. The agent works on the customer record, with a defined approval rule, and the old process leaves the stage once the new one is validated.
6. Set the exit criterion and the date. The pilot ends with a decision: it goes into operations, changes scope or is shut down. The 90 days of MIT’s mid-market companies serve as a reference for the first review.
7. Add up the real cost. Include consumption, implementation, data cleanup, training and hours of human review, because rising cost shows up as a brake in both Gartner and McKinsey.

The role of HR and L&D in the move to production

BCG attributes about 70% of implementation challenges to people and processes, and MIT names resistance to new tools as the most cited barrier. For HR, training needs to happen in the process the pilot chose, with the agent running on the company’s data. MIT observed that many strong deployments started with people who were already testing AI on their own, and that group is a good starting point for building internal champions. The training metric is usage in the process, measured in the CRM before and after the class.

The projects that leave the pilot have a chosen process, prepared data, an owner in the business and a team that knows how to work with the agent. That is the work of the in-company applied AI training, built on your company’s processes and data, with groups for executives, managers and operations. It delivers the prompt library by area and the agent architecture, with what each agent reads, what it executes and who approves, the design most pilots lack when it is time to go into production.

Frequently asked questions

What percentage of AI projects never leave the pilot?

There is no single number. In MIT’s 2025 report, 5% of task-specific generative AI tools reached production with sustained impact, in a sample of 52 organizations. Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, and BCG found 26% of companies with the capability to move from proof of concept to value.

Is MIT’s 95% figure reliable?

The number is in the report, but it measures less than the headline suggests. It comes from interviews and surveys with a small sample, uses a six-month window to measure return, and the authors themselves call it directional. It works as a sign that moving into production is hard, not as the market’s failure rate.

Why do AI projects fail?

The causes that show up most in the research are a poorly defined problem, data that is not ready, AI outside the workflow, no owner in the business, low adoption and cost without proven value. At RAND, 84% of industry interviewees named leadership causes as the main one, with data close behind.

How long should an AI pilot last?

In MIT’s report, the best-performing mid-market companies took an average of 90 days from pilot to full deployment, and large ones took nine months or more. I use 90 days as the deadline for the first decision, with an exit criterion written before the start.

How do I prepare the team to take AI from pilot to production?

Train people in the process the pilot chose, with the agent running on the company’s real data. Start with the professionals who already use AI on their own, to build internal champions, and define the approval rules with the managers. Measure usage in the process before and after the training to know whether behavior changed.

Do I need to fix the CRM before adding an AI agent?

You need to fix the fields the agent will read in the chosen process, not the whole database. Gartner predicts that 60% of AI projects without AI-ready data will be abandoned through 2026, and in the CRM that means handled duplicates, reliable dates, stage criteria and filled-in required fields. That cleanup goes into the project schedule, with an owner and a deadline.