Why most enterprise AI pilots never reach production, and what sets the successful ones apart

What the numbers really say

Several figures about failed AI pilots are in circulation, and presentations often lump them together. Yet each one measures something different – and that matters if you want to use them to make a case to leadership.

The most frequently cited is the report The GenAI Divide: State of AI in Business 2025 from the MIT NANDA initiative. According to it, 60% of organizations evaluated specialized enterprise AI tools, 20% reached the pilot stage and only 5% made it to production. The report defines success as deployment beyond the pilot with measurable KPIs, with the impact on returns measured six months after the pilot.

A March 2025 S&P Global Market Intelligence survey showed the same phenomenon from another angle: the share of companies that scrapped most of their AI initiatives rose year over year from 17% to 42%. In September 2025, BCG classified 60% of companies as laggards with minimal gains and only 5% as future-ready.

Why you shouldn’t take the 95% figure literally

The MIT NANDA report drew justified criticism, and it is only fair to know about it before you put the number into a board deck.

• A small, non-representative sample. In the methodology section, the authors themselves state that the figures are “directionally accurate”, that they are based on individual interviews rather than official company reporting, and that organizations willing to talk about their difficulties may differ systematically from the rest.
• A short window. Six months is not much time to evaluate a change to a business process. The report itself admits that this may understate the success rate.
• Unclear origin of the headline number. The analyst site Futuriom quotes Wharton professor Kevin Werbach, who says the document does not make it possible to trace where the 95% figure actually comes from.
• The authors’ interests. NANDA is an initiative promoting its own architecture for networked agents. The report’s conclusion naturally points toward it.

Second, “zero impact on the P&L within six months” is not the same as “technically non-functional”. Many pilots work; nobody has simply carried them through into a process where the savings or revenue could be measured.

The state of play in 2026: visible progress, but uneven

Newer data shows that things are moving. According to the McKinsey The State of AI in 2026 report (data collected from May 4 to June 8, 2026), 40% of respondents from large organizations are scaling AI agents, while among smaller ones the share remained at 22%. Still, 30% of large organizations with revenue above one billion dollars remain in the experimentation and pilot phase.

The group that attributes an EBIT impact of 5% or more to AI and also reports significant value makes up only 6% of respondents. What matters is how they differ: nearly three quarters of them say they fundamentally redesigned workflows because of AI, and they are twice as likely to say that senior leadership clearly demonstrates its commitment to AI.

Caution is warranted with agents. In June 2025, Gartner predicted that over 40% of agentic projects will be canceled by the end of 2027 – due to escalating costs, unclear business value and inadequate risk controls. It also warned of “agent washing”: of the thousands of vendors claiming agentic AI, by its estimate only about 130 actually offer agentic solutions.

The Czech context: fast adoption, weaker impact

The Czech Republic is not lagging behind in adoption. According to an SAP survey of 200 executives, published in September 2026 at the SAP NOW conference in Ostrava, 83% of companies have deployed AI in their software tools, up from 41% last year. However, AI is a key part of the long-term strategy for only 28% of businesses; another 54% have it in their strategy but are still mainly experimenting.

The same survey also names the obstacles: a shortage of skilled people, unclear rules and legislation, and insufficient or inconsistent data sources, which 31% of companies list among the biggest barriers.

On top of that comes a typically Czech problem, described in June 2026 by CzechCrunch. Jan Šmerhovský of Google Cloud states there that AI adoption in the Czech Republic reaches 48%, compared with an EU average of around 37%, but that companies tend to build everything in-house, even in areas outside their core business.

It is worth keeping in mind that this is the view of a vendor that sells off-the-shelf solutions. It does, however, match the data from the MIT NANDA report: external partnerships reached deployment in roughly 67% of cases, internally built tools in about 33%. The authors themselves add that the difference may reflect organizational capabilities more than the approach itself.

Five reasons a pilot stays a pilot

1. The pilot has no business owner and no metric

The project was born in an innovation team or in IT because leadership wanted “something with AI”. There is no one whose numbers the result will change, and no baseline against which the benefit could be measured. The MIT NANDA report states that successful organizations sourced ideas from line managers rather than a central lab, and evaluated tools by operational outcomes rather than model benchmarks.

2. AI gets bolted onto an old process

A chatbot next to an unchanged workflow just adds an extra step. McKinsey’s data shows that the difference comes from redesigning the process, not from the tool itself.

3. The demo runs on an export, production needs live data

A pilot built on a hand-cleaned CSV looks great. In production you need API access to systems, document-level permissions and up-to-date data. If key systems have no usable interface, the pilot will founder on integration – not on the model.

4. Nobody planned for operations

Who deals with the model answering differently after an update? Who monitors token costs, latency and error rates? Who is on call? Without answers, security and operations will rightly block the deployment.

5. Teams build what they could buy

A custom RAG platform, custom agent orchestration, a custom evaluation framework – and a year later the team is maintaining infrastructure instead of solving a business problem. Moreover, according to the MIT NANDA report, the main barrier is not model quality but the fact that most systems do not retain feedback or adapt to context. Solving that in-house is expensive.

Build vs. buy: decision criteria

Neither a simple “buy” nor “build” holds. It is more useful to split the solution into layers and decide for each one separately.

The counterargument is legitimate: buying means vendor dependency, data held by a third party and, in the age of “agent washing”, the risk that you are buying a rebranded chatbot. That is why, even with a purchased solution, you should keep the data, the integration layer, the evaluation set and the ability to swap the model in your own hands.

The infrastructure view: what must be ready before the pilot

A pilot that has a chance of reaching production is built from the start on the same foundations as any other service. If those foundations are missing, an AI project will only expose the gap – this is covered in more detail in the article on infrastructure chaos before adopting AI.

Evaluation in the pipeline doesn’t have to be complicated. It is enough that a prompt change cannot be merged without passing a set of test cases:

# example CI step – script names and thresholds are illustrative
ai-eval:
  stage: test
  script:
    - python eval/run.py --dataset eval/cases.jsonl --out report.json
    - python eval/gate.py --report report.json --min-accuracy 0.85 --max-cost-per-case 0.02
  rules:
    - changes: ["prompts/**", "config/model.yaml", "src/**"]

The same principle applies as with ordinary software: what is deployed manually and not measured won’t survive in production. Related practices are described in the articles on deployment automation and on monitoring and logging.

What to do on Monday

Also set termination criteria in advance. A canceled pilot is not a failure if it delivered an answer at a reasonable cost. A failure is a pilot that runs for a year, nobody uses and nobody has the courage to switch off.

Key takeaways

• The 95% figure is indicative. It rests on 52 interviews and a six-month measurement; the authors themselves acknowledge its limitations.
• Larger surveys confirm the direction. According to McKinsey 2026, 6% of respondents belong to companies with a significant AI impact on profit.
• Process redesign makes the difference. Nearly three quarters of the top performers fundamentally changed their workflows.
• Czech companies adopt AI fast but mostly experiment. 83% of companies have AI in their tools; for 28% it is a key part of their strategy.
• Buy the generic, build the integration. Nobody will do your data, permissions and connections to your own systems for you.
• Build the pilot as a future service. Identity, pipeline, evaluation, logging and an operations owner from day one.