• Report
  • 6 min read

Where AI pilots stall on the way to production

In MIT NANDA's review, a quarter of enterprise AI pilots reached production. We followed a pilot through the stages where the 2025 and 2026 studies say it loses ground.

ResearchBy OliviaPublished 3 September 20266 min read8 sources

Pilots usually get the best a company can offer. The team has volunteered, someone has picked the data, and the sponsor wants a success to report. None of that carries over to production, which is where the research published since early 2025 finds most projects losing ground.

MIT NANDA’s review is the most quoted source on this. It covered more than 300 publicly disclosed AI initiatives, 52 structured interviews and 153 survey responses gathered between January and June 2025.1 Of the organisations that evaluated enterprise-grade AI tools, a third reached a pilot, and a quarter of those pilots reached production. Below, the stages are taken in the order a project meets them, with figures from Gartner, IBM, BCG and McKinsey where MIT has none.

Enterprise-grade AI tools: from evaluation to production

Share of organisations reaching each stage, January to June 2025

Evaluated a tool60%
Reached a pilot20%
Reached production5%

Source: MIT NANDA, The GenAI Divide, 2025, as reported by Virtualization Review, 19 August 2025.1

Before the pilotStaff are already using AI, on accounts the company does not manage

Only 40% of companies in the MIT study had bought an official large language model subscription. At more than 90% of them, staff said they used personal AI tools for work.1

That cuts both ways. The demand is already there, because employees have found their own uses for these tools. But when an approved tool arrives, it competes with whatever people already do on their phones, and if it can’t handle the same tasks, people stay with their own.

It is also a security problem, because work material is moving through accounts that IT cannot see. Banning personal tools without offering a better approved one usually pushes that use further out of sight.

Inside the pilotThe sample was cleaned by hand

A pilot team can clean its data by hand, but nobody can do that across a whole business. Gartner surveyed 1,203 data management leaders and found that 63% of organisations either lack the data management practices AI needs or don’t know whether they have them. Gartner expects organisations to abandon 60% of AI projects through 2026 where the data is not AI-ready.2

Do organisations have the data practices AI needs?

Share of data management leaders, Gartner

Lack them, or unsure63%
Report having them37%

Source: Gartner, 26 February 2025.2

Other surveys agree. In Informatica’s 2026 study of 600 data leaders, 57% named data reliability as a key barrier to moving AI from pilot to production.3 Fivetran found that 42% of enterprises had seen more than half of their AI projects delayed, underperform or fail because of data readiness.4

Gartner recommends five steps: align data to specific use cases, identify the governance requirements AI adds, make metadata active, prepare pipelines for both training and production, and keep checking the data once it is live.2 A useful early test for any pilot is whether its last-mile data preparation can be automated at all.

Inside the pilotCorrections that vanish

MIT identifies learning capability as the main barrier. The tools in its sample often failed to retain feedback, adapt to a team’s context or improve with use. One mid-market manufacturing COO quoted in the study said LinkedIn makes it sound as though everything has changed, while day-to-day operations look much the same.1

Few pilots test for this, although it is easy to do. Have five users correct the system on the same task over two weeks, then check how many of those corrections still hold at the end without being repeated. If almost none do, the pilot is a demonstration, and in production the system would make the same mistakes at full volume.

The approval meetingRunning costs show up after the decision

Gartner gives three reasons for its forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027: escalating costs, unclear business value and inadequate risk controls.5 In McKinsey’s 2026 survey, one respondent in five said their organisation was limiting AI use because of operating costs.6

Costs show up late because pilots are small. Usage is low, vendors sometimes discount it, and the people checking outputs are enthusiasts who volunteered. The full bill only arrives once the system is open to everyone.

CostWhy the pilot hid itQuestion for the approval meeting
Cost per run at full useVolumes were small and often discountedWhat does a run cost at ten times pilot volume?
Data preparationThe team cleaned the sample by handWhich preparation steps are automated, and who owns them?
Review of wrong answersPilot users reported errors informallyWho reviews outputs once the system is live, and how quickly?
Audit and loggingA controlled test did not need themWhat would an auditor ask to see?

After approvalLarge organisations take longer to get going

MIT reports that mid-market firms took about 90 days from pilot to full implementation, while large enterprises took nine months or more.1 IBM’s survey of 2,000 CEOs suggests one reason: half said rapid investment had left them with disconnected, piecemeal technology.7

Time from pilot to full rollout

MIT NANDA, 2025

~90 daysmid-market firms
9+ monthslarge enterprises

Source: MIT NANDA via Virtualization Review, 2025.1

McKinsey’s 2026 survey shows what happens next. Among respondents from large organisations, 40% now say they are scaling AI agents, compared with 22% at smaller ones.6 Large firms start slower but seem to scale further, perhaps because the platform, approval and ownership work they do early pays off once projects are ready to grow.

For sponsors, the main implication is the timeline. In this evidence, nine months is normal for a large enterprise, so a plan written for three months will show a delay in month four even when the project is on track.

What the survivors shareThey change the work before they widen access

McKinsey found that nearly three-quarters of its high performers had fundamentally redesigned workflows because of AI, up from 55% a year earlier. Among other respondents the figure was about a quarter.6 BCG reports that 33% of its "future-built" companies use AI agents and spend about 15% of their AI budget on them, against 12% of scalers and almost no laggards.8

Share of each BCG group using AI agents

BCG survey of 1,250 senior executives, 2025

Future-built33%
Scalers12%
LaggardsAlmost none

Source: Boston Consulting Group, 30 September 2025.8

The same sequence appears across these studies. Organisations that reach production pick a process, give it a business owner and agree what a good result looks like before a model goes into it, and they open access to more people only after that.

For your next reviewSix symptoms and where to look first

SymptomLikely causeFirst check
Good demo, no production dateNobody in the business owns the workflowName the owner and the process the system will change
People prefer their personal chat toolThe approved tool cannot do a task staff rely onCollect the ten tasks people take to personal tools
Accuracy drops as the sample widensThe pilot data was cleaned by handAutomate preparation and measure error on unseen data
Different answers to the same business questionTerms are defined differently across sourcesPublish one definition per term, with an owner
Sponsor reports slippage in month fourThe plan assumed mid-market speedReset against a nine-month reference for large firms
Budget approved, then frozenRunning costs were never modelled at full volumePrice the system at ten times pilot use before the next review

These are starting points drawn from the research above. Check each one against your own project data.

Sources and method

The MIT NANDA report is cited through the trade press article named below.

  1. MIT Report Finds Most AI Business Investments Fail, Reveals "GenAI Divide" (coverage of MIT NANDA, The GenAI Divide: State of AI in Business 2025)Virtualization Review, 19 August 2025
  2. Lack of AI-Ready Data Puts AI Projects at RiskGartner, 26 February 2025
  3. New Global CDO Report Reveals Data Governance and AI Literacy as Key Accelerators in AI Adoption (CDO Insights 2026)Informatica, 27 January 2026
  4. Fivetran Report Finds Nearly Half of Enterprise AI Projects Fail Due to Poor Data ReadinessFivetran, 13 May 2025
  5. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027Gartner, 25 June 2025
  6. The state of AI in 2026: On the road to ROIMcKinsey & Company, 25 August 2026
  7. IBM Study: CEOs Double Down on AI While Navigating Enterprise HurdlesIBM, 6 May 2025
  8. AI Leaders Outpace Laggards with Double the Revenue Growth and 40% More Cost SavingsBoston Consulting Group, 30 September 2025

This report summarises published research and is not professional advice.

BIAAS 2027, Amsterdam

Hear it first-hand from the people doing the work

Two days of keynotes, panels and 1-on-1 meetings with senior data, analytics and AI leaders, on 23-24 March 2027.