Why Do Enterprise AI Pilots Stall? The Missing Context Layer

Updated October 2026. This article replaces our earlier post on governance, trust and safety.

Most enterprise AI pilots stall because the tool doesn't know your organization, not because the model is weak. MIT NANDA's 2025 study found the dividing line was whether a system keeps context, learns from feedback, and fits the way work actually gets done. A context layer gives every AI tool the same governed record of what your organization knows and decided, so it doesn't start from zero each time.

What did MIT's 2025 study actually find?

MIT NANDA's report, The GenAI Divide: State of AI in Business 2025, opens with a hard line: "95% of organizations are getting zero return." It adds that "just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact." (MIT NANDA, 2025)

That number traveled fast. Read the fine print before you repeat it.

• What counts as success. The report calls a task-specific AI tool "successfully implemented" when users or executives said it caused "a marked and sustained productivity and/or P&L impact." Its methods section measures ROI six months after the pilot.
• Who was studied. The authors label the work "preliminary findings." It rests on interviews with 52 organizations, survey responses from 153 senior leaders at four industry conferences, and a review of more than 300 public AI initiatives, all from January to June 2025.
• What the authors caution. They call the pilot figures "directionally accurate based on individual interviews rather than official company reporting," and warn that six months "may be insufficient" to judge complex enterprise systems.
• What critics add. In Forbes, Martin Gutmann wrote that the 95% number "isn't a robust finding," pointing to the six-month window, the lack of a financial audit, and a convenience sample. (Forbes, 2025)

The report's own chart is more useful than the slogan. For custom or vendor-sold enterprise AI tools, MIT reports: "Sixty percent of organizations evaluated such tools, but only 20 percent reached pilot stage and just 5 percent reached production." Generic chat tools, by contrast, "appear to show high pilot-to-implementation rates (~83%)."

So "95% of AI fails" overstates it. A fairer reading: in this sample, few custom enterprise AI tools went from pilot to a measurable business result within six months, while general chat tools spread easily. The report's explanation of why is worth more than the headline number.

Why do pilots stall if the model isn't the problem?

MIT's answer is blunt: "The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time." The divide, the authors write, "does not seem to be driven by model quality or regulation."

Users said the same thing in their own words. Asked what kept AI out of their core work, they named barriers such as "Too much manual context required each time" and "It doesn't learn from our feedback." One corporate lawyer in the study liked a general chat tool for first drafts but drew a line at sensitive contracts: "it doesn't retain knowledge of client preferences or learn from previous edits. It repeats the same mistakes and requires extensive context input for each session."

That gap shows up in trust. For quick tasks such as emails and summaries, 70% of users in the study preferred AI. For complex, multi-week work, 90% preferred a human colleague. MIT's conclusion: "The dividing line isn't intelligence, it's memory, adaptability, and learning capability."

Picture a common case. A new procurement lead asks an AI assistant why the company dropped a supplier two years ago. The model can write a fluent answer. It can't know that the real reason lives in a contract amendment, a long email thread, and a meeting nobody wrote up. So it guesses, or the user pastes in background by hand. Either way, trust drops.

Is this a data problem too?

Yes. Gartner reaches a similar place from the data side.

• Gartner predicts that "through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data." In a July 2024 Gartner survey of 1,203 data management leaders, 63% of organizations either didn't have, or weren't sure they had, the right data management practices for AI. (Gartner, 2025)
• Gartner also found that by the end of 2025, at least 50% of generative AI projects were abandoned after proof of concept "due to poor data quality, inadequate risk controls, escalating costs or unclear business value." (Gartner, 2026)

Gartner describes the root of it plainly: in traditional data management, "uses of data are not well-documented, and data is often collected in siloes across various repositories, multiple systems and platforms." That's the context problem in different words. AI can't use what it can't find, can't trust, or isn't allowed to see.

What is a context layer?

A context layer is a governed layer between your existing systems and the AI tools your people and agents use. It does three jobs a model can't do alone:

1. Connect what's scattered. Files, records, and messages stay in the systems where they live, and the layer makes them findable together.
2. Remember the why. Decisions, the reasons behind them, and what happened next stay on record after people, tools, and projects change.
3. Govern access. It sets what information can be used, by whom, and where, and keeps the sources attached to every answer.

That's how Nvitis is built. Its four pillars map to those jobs:

• Unified Semantic Index connects data from systems such as Microsoft 365, Workday, Slack, AWS, and legacy applications into one context-aware foundation, without asking you to move files into a new drive.
• Institutional Memory keeps the reasons behind decisions as Living Storyboards: chronological timelines with short narratives that explain what each item is and why it matters.
• Dynamic Intelligence gives leaders live visibility into workflows, staffing, and initiatives so they can spot bottlenecks earlier.
• Governance controls what AI can use, protects sensitive data, grounds answers in cited sources to reduce hallucinations, and keeps an audit trail.

Back to MIT's barriers. "Too much manual context required each time" is the problem a context layer is designed for: the organization's record is already there, so people and agents don't rebuild it in every prompt. MIT also found that executives want this. Among the traits of winning tools, the report lists systems that "retain context (63% demand this)," and buyers named "clear data boundaries" as a selection test. One executive put it bluntly: "I can't risk client data mixing with someone else's model, even if the vendor says it's fine."

For more on the index and the memory pillars, see Breaking Silos and The Perpetual Enterprise Brain.

What about hallucinations, leaks and runaway costs?

Hallucinations, data leaks, compliance violations, and runaway token costs can quickly turn an exciting AI initiative into a liability. If an AI system is not grounded in trusted context and controlled by policy, it can produce output that sounds right while being materially wrong. If the system also has broad access to enterprise data, the risk grows because sensitive information may be exposed in the wrong context or to the wrong user. And many organizations do not anticipate how expensive AI becomes when every query pulls too much context, every workflow is poorly scoped, and every interaction is allowed to run without controls. For how a governance layer handles these risks, see Governance at Scale.

What won't a context layer fix?

Be honest about the limits. Context is necessary, but it isn't enough.

Gartner calls a lack of business value "the most fundamental failure mode," and lists poor change management as its own failure point. MIT's survey put resistance to adopting new tools at the top of the barrier list. A context layer won't choose your use case, set your success metric, or win over a skeptical team.

MIT's description of buyers who got results is a good guide. They "benchmarked tools on operational outcomes, not model benchmarks," and "sourced AI initiatives from frontline managers, not central labs." In short: pick a real problem, measure the outcome, and let the people who own the work lead.

A checklist before you approve the next AI pilot

1. Name the result. What business outcome will change, and how will you measure it at six months?
2. Ask where the context comes from. Which systems does the tool read? Does each user have to paste in background every session?
3. Test it on history. Ask, "Why did we decide this?" If it can't show its sources, it's guessing.
4. Set data boundaries first. Decide who can see what, and what may leave your environment, before the pilot starts.
5. Check what survives turnover. If the pilot's champion leaves, does the knowledge stay?
6. Start narrow. Pick one workflow where a team already feels the pain, and let that team's manager lead.
7. Plan adoption like a project. Training, feedback, and follow-up need an owner, not a launch email.

Frequently asked questions

What did the MIT 95% AI study find? MIT NANDA's 2025 report said "95% of organizations are getting zero return" from generative AI and that "just 5% of integrated AI pilots are extracting millions in value." It traced the gap to tools that don't retain feedback, adapt to context, or improve over time, rather than to model quality.

Is it true that 95% of AI pilots fail? Not as a general rule. The figure comes from preliminary research based on 52 interviews, 153 survey responses, and 300-plus public initiatives. Success meant a marked and sustained productivity or P&L impact, judged about six months after the pilot. The authors call their pilot figures "directionally accurate," and critics have questioned the sample and the short window.

What is a context layer for enterprise AI? It's a governed layer between your systems and your AI tools. It connects scattered information, keeps the reasons behind decisions, and controls who and what can use each piece, so every answer can show its sources.

How is a context layer different from enterprise search? Search finds documents. A context layer also keeps the story that connects them: what was decided, why, and what happened next. It applies access rules before anything reaches a person or an AI agent.

What is AI-ready data? Gartner says AI-ready data must be "representative of the use case," and that proving readiness depends on metadata "to align, qualify and govern the data." Its short version: "if the data has issues, then the data is not ready for AI."

Does a context layer replace our existing systems? No. Nvitis sits on top of the systems you already use and doesn't ask you to move files into a new drive.

Next step

If your pilots keep stalling at "it doesn't know us," start with the context, not a new model. Book a demo at nvitis.com/contact and we'll show you how a context layer works with the systems you already have.

The model is the easy part to buy. The context is what you have to keep.

Sources

1. Challapally, A., Pease, C., Raskar, R., Chari, P. The GenAI Divide: State of AI in Business 2025. MIT NANDA, July 2025. Originally hosted at nanda.media.mit.edu/ai_report_2025.pdf (that URL now redirects to the NANDA project page). Copy used: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
2. Gartner. "Lack of AI-Ready Data Puts AI Projects at Risk." Feb 26, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk
3. Chandrasekaran, A. "Why 50% of GenAI Projects Fail — And How to Beat the Odds." Gartner, Jan 26, 2026. https://www.gartner.com/en/articles/genai-project-failure
4. Gutmann, M. "Why That Viral AI Statistic Might Be More Hype Than Science." Forbes, Oct 8, 2025. https://www.forbes.com/sites/martingutmann/2025/10/08/why-that-viral-ai-statistic-might-be-more-hype-than-science/