95% of enterprise AI investments are delivering zero measurable returns. That's not a prediction. It's what MIT NANDA found when they studied GenAI adoption across organizations in 2025. And a new field diagnostic from Andus Labs, published this week, traces exactly where the value disappears — and it's not in the model you picked.
The enterprise AI market is at a turning point. Companies have committed billions to AI transformation. Boards are demanding ROI. And yet most organizations can't point to a single financial metric that moved because of AI. The gap isn't a technology problem. It's an operating model problem.
Here's what's really happening — and what separates the 5% who are getting real results from the 95% who are sitting on an expensive portfolio of proof-of-concept projects that never reached production.
The Pilot Graveyard Is Real
Andus Labs calls it the Pilot Graveyard, and it's more common than most executives want to admit. According to S&P Global Market Intelligence's 2025 report on generative AI, 42% of companies abandoned most of their AI initiatives — more than double the rate from the year before. The average organization scrapped 46% of its proof-of-concept projects before reaching production.
Think about that for a moment. Nearly half of all AI pilots die before they go live. And the ones that do reach production often stall once the controlled environment goes away.
The reason is predictable: AI pilots are built to succeed under conditions the broader organization can't reproduce. You staff them with your best people, give them access to clean data, remove friction from their workflow, and declare success. Then the pilot ends, the old operating system reasserts itself, and nothing changes at scale.
What kills AI at scale isn't a bad model. It's outdated workflows, misaligned decision rights, and incentive structures that still reward the old way of working.
The 25 Patterns Blocking Your Returns
Andus Labs spent time in the field studying where enterprise AI adoption actually breaks down. The result is their Ground Truth Index — a diagnostic ranking the 25 highest-priority patterns blocking enterprise AI returns, drawn from a corpus of more than 200 field signals.
The index organizes failure patterns across six dimensions where AI adoption consistently breaks down:
- Capability — Can your workforce actually use AI effectively?
- Behavior and Trust — Do your people trust the tools, or route around them?
- Tech-Workflow Fit — Does AI actually integrate with how work gets done?
- Institutional Coherence — Are your processes and systems aligned with AI adoption?
- Leadership and Decision — Are decisions being made at the right speed and level?
- External Forces — Regulatory, competitive, and market pressures affecting adoption?
A single weak layer stalls the entire program. You can have the best foundation model, the most sophisticated RAG pipeline, and a beautifully designed chat interface — and still see zero returns if your incentive structures reward the old behavior.
This quarter, Andus Labs identified two Critical-tier patterns. These aren't edge cases. They're showing up everywhere.
Pattern #1: Trust Deficit
The top-ranked pattern blocking enterprise AI returns is Trust Deficit. Here's how it manifests: leaders treat probabilistic AI as a deterministic search engine, then call it broken when it doesn't behave like one.
I've seen this play out repeatedly in peer conversations across industries. A CFO gets excited about AI for financial forecasting, runs a pilot, and the model returns a range with confidence intervals instead of a single number. The reaction: "This thing doesn't work." A CTO deploys a code-review assistant, and when it flags something as a "possible" issue rather than a definitive bug, developers stop trusting it entirely.
The problem isn't the AI. The problem is the mental model people bring to it. Every enterprise database query returns a deterministic answer. Every ERP system gives you a precise number. AI doesn't work that way — and the organizational trust required to use probabilistic outputs effectively is something you have to deliberately build.
"You can't train your way out of a trust problem," noted Chris Perry, Founder and CEO of Andus Labs. "When people believe the tool is a threat, they'll use it to look compliant and keep working the way they always have. Readiness work changes what an organization rewards."
The Gallup data backs this up. As of April 2026, only 13% of U.S. employees use AI daily at work. Just 28% use it a few times per week or more. Access has spread faster than the habits needed to use AI effectively.
What this means for your organization: If adoption is stalling, check whether your culture actually rewards acting on AI outputs. Do your teams trust the model enough to use its recommendations? Have you redesigned work to make AI use the expected path rather than the exceptional one?
Pattern #2: Tempo Shock
The second Critical-tier pattern is one I find particularly interesting: Tempo Shock. Companies already have the budgets and the board mandates. What they lack is a way for machine-speed analysis to reach a decision before it expires.
Here's the problem: AI can analyze your pipeline data in seconds and flag that a deal is at risk. But your deal review cycle runs weekly, the risk gets added to next Thursday's agenda, and by the time the sales VP sees it, the customer has already signed with a competitor.
Machine-speed insight requires machine-speed decision architecture. Most enterprises don't have it.
This is where the gap between technical capability and business value becomes most visible. You've deployed a real-time inventory optimization system — congratulations. But if your procurement team's authority level requires VP approval for any order outside the monthly cycle, the AI's recommendation arrives and waits. And waits.
The fix isn't a better model or a faster processor. It's redesigning who can decide what, and how fast.
For technical leaders (CIOs, CTOs): The AI infrastructure you're building needs to connect to decision workflows, not just data pipelines. An insight that sits in a dashboard for three days isn't faster than an analyst's weekly report — it just cost more to produce.
For business leaders (CFOs, COOs): Your governance structures were built for a world where analysis took days. AI compresses that to seconds. If your decision rights and approval hierarchies haven't changed, you're paying for speed you can't use.
What the 5% Do Differently
The organizations getting measurable returns from AI share a pattern that has nothing to do with which model they deployed. They redesigned their operating model first — or simultaneously — rather than bolting AI onto existing workflows and hoping for adoption.
According to Wharton Human-AI Research and GBK Collective's October 2025 report, "Accountable Acceleration," enterprise GenAI has moved from exploration into disciplined adoption. 72% of business leaders are now formally measuring ROI. The organizations leading that measurement share three characteristics:
They changed what work looks like, not just the tools available. Successful deployments redesign the actual task — who does what, when, with what authority — rather than giving people a new tool to complete the same task. A finance team that used to build month-end variance analyses manually doesn't just get an AI copilot. They get a redesigned workflow where AI handles the first draft and the human's job becomes reviewing, interpreting, and acting on exceptions.
They made AI use the expected path, not the optional one. Every successful enterprise AI deployment I've heard about from peers has one thing in common: using the AI became the default, not a choice. The old path got harder, or disappeared entirely. This sounds obvious but is almost never done. Most organizations deploy AI as a voluntary productivity tool and then wonder why adoption plateaus at 20%.
They invested in trust infrastructure before capability. The 5% didn't skip straight to autonomous agents. They built a track record — model outputs that the organization could see, evaluate, and gradually learn to act on. Trust at scale doesn't come from a training program. It comes from repeated positive outcomes that people can point to.
Microsoft Just Put $2.5 Billion on This Thesis
When the largest enterprise software company in the world makes a $2.5 billion bet on deployment help rather than better models, that tells you something about where the real problem is.
Microsoft's Frontier Company, launched earlier this month, combines 6,000 AI engineers, industry specialists, and change management experts specifically to help enterprises redesign workflows, establish governance, and integrate AI into existing business systems. Judson Althoff, CEO of Microsoft's commercial business, said the initiative stems from realizing that "customers are in very different places right now, and trying to really figure out AI."
That $2.5 billion isn't going toward a better model. It's going toward the organizational change problem that Andus Labs has been cataloging in their field research.
The message is clear: the AI itself is increasingly a commodity. The operating model transformation is the hard part. And the gap between organizations that figure that out and those that keep running pilots is widening fast.
What You Should Do This Month
If you're leading an enterprise AI initiative — whether you're the CISO managing security copilots, the CFO reviewing AI ROI, or the CTO building the infrastructure — here's what the data says to focus on:
Run an honest adoption audit. Don't count licenses sold or tools deployed. Count how many people are using AI outputs to make actual decisions, and how often. If that number is low, you have a trust and workflow problem, not a technology problem.
Identify your Tempo Shock points. Map three to five places where AI analysis currently produces an insight that then waits for a human decision cycle that wasn't built for it. Those are your highest-leverage redesign opportunities.
Pick one workflow to actually redesign. Not augment — redesign. Remove the old path or make it significantly harder. Give the team using AI a meaningful difference in how their work is structured. Measure the output change, not the adoption rate.
Connect AI outputs to incentives. If your sales team's commission structure still rewards the same behaviors as before AI, don't be surprised when they route around the AI-powered pipeline tool. Behavior follows incentives. Always has.
Start measuring organizational readiness, not just technical readiness. Evaluate whether people have the decision authority they need to act on machine-speed analysis. Evaluate whether your culture rewards acting on probabilistic outputs or waits for certainty that never arrives.
The Bottom Line
The enterprise AI conversation has been dominated by model selection, infrastructure choice, and vendor comparison. Those decisions matter at the margin. But the research is clear: returns follow operating model redesign, not model capability.
The 95% who are delivering zero measurable returns from their GenAI investments aren't running worse models than the 5%. They're running the same models inside operating systems that weren't built to act on what AI produces.
The organizations that figure this out — that make the organizational changes required to actually use AI at machine speed — are going to build advantages that compound over time. The ones that keep running pilots, hoping the next model will fix the adoption problem, are going to look up in two years and wonder why the gap keeps growing.
The technology is ready. The question is whether your organization is.
Sources: Andus Labs Ground Truth Index, July 2026; MIT NANDA "The GenAI Divide: State of AI in Business 2025"; S&P Global Market Intelligence "Generative AI Shows Rapid Growth but Yields Mixed Results," 2025; Wharton Human-AI Research/GBK Collective "Accountable Acceleration," October 2025; Gallup "Rising AI Adoption Spurs Workforce Changes," April 2026; FlexJobs AI survey, 2026.
Follow the conversation:
- LinkedIn: linkedin.com/in/rberi
- X/Twitter: @rajeshberi
