Growth Theater vs Real Adoption: Why Enterprise AI Numbers Don't Match Reality (internal research)

The Numbers Tell a Story Most Leaders Don’t Want to Hear

Enterprise AI is one of the most widely announced technology investments of our era. Vendor keynotes, earnings calls, and boardroom slide decks are saturated with claims of transformation. Yet something strange is happening beneath the surface of those headlines: the outcomes don’t match the announcements.

A 2025 MIT NANDA report, The GenAI Divide: State of AI in Business 2025, examined more than 300 public AI initiatives, conducted 52 structured executive interviews, and surveyed 153 senior leaders across industries. Its finding was stark: despite $30-40 billion in enterprise generative AI spending, 95% of organizations are getting zero measurable return on that investment. Not modest returns. Not disappointing returns. Zero.

That number is worth pausing on. Because at the same time, McKinsey’s State of AI in 2025 reports that 88% of organizations say they use AI in at least one function, and 72% say they use generative AI specifically. How is it possible that nearly every company is “using AI,” yet almost none are capturing value from it?

The answer is what we at Datafi call Growth Theater: the widespread organizational behavior of adopting AI at the surface level, deploying tools, launching pilots, announcing initiatives, without building the underlying infrastructure that transforms AI from a question-answering novelty into a problem-solving engine that drives real outcomes.

What the Data Actually Shows

The MIT finding is not an outlier. Across every major research institution studying enterprise AI, the pattern is consistent.

McKinsey’s data shows that while 88% of organizations claim AI use, only 39% report any EBIT impact at the enterprise level, and for most of those, the impact is under 5% of earnings. Only approximately 6% of respondents, what McKinsey calls “AI high performers,” attribute 5% or more of EBIT to AI. McKinsey names the dominant failure pattern directly: “AI theater, organizations going through the motions of AI adoption without rewiring the operating model to capture real value.”

Forrester’s State of AI Survey 2025 surveyed more than 1,400 AI decision-makers and found that only 13% report positive EBITDA impact, and fewer than a third can connect AI contributions to P&L at all. Their analysts describe the phenomenon as an “endless pilot treadmill, producing activity without outcomes.”

S&P Global Market Intelligence’s Voice of the Enterprise research found that the proportion of companies abandoning most of their AI initiatives has increased from 17% to 42% in just one year. The average organization is now scrapping 46% of its proof-of-concept projects before they ever reach production. Gartner had forecast in mid-2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. The actual number has already blown past that estimate.

What’s causing this gap? After working directly with enterprise customers navigating these exact challenges, we believe the answer comes down to a fundamental misunderstanding of what AI actually needs to solve hard business problems, not just answer questions.

AI That Answers Questions vs AI That Solves Problems

There is a critical distinction that most enterprise AI strategies are failing to make. The AI tools most organizations are deploying today are excellent at answering questions. Ask a language model something general, something that requires knowledge of the world, something a well-read analyst might know, and it will perform admirably. But enterprise value is not created by answering general questions.

Enterprise value is created by solving specific business problems: optimizing a logistics route given today’s actual asset availability, predicting which pieces of manufacturing equipment will fail before next quarter’s maintenance window, identifying which insurance claims are likely to require escalation before a human ever reviews them. These problems require AI that knows your business, your data, your operations, your policies, your customers, with enough depth to act on that knowledge autonomously.

The overwhelming majority of enterprise AI deployments today are doing something far simpler. They are wrapping a general-purpose language model around a chat interface and asking employees to query it. The model has no access to the company’s actual data. It has no understanding of the business’s specific context. It cannot take action. It answers questions. And because it answers questions with impressive fluency, organizations count it as “using AI”, and wonder why the P&L hasn’t moved.

This is the structural heart of Growth Theater.

The Four Reasons Pilots Don’t Scale

Synthesizing findings from MIT, McKinsey, RAND, BCG, and Gartner, four structural barriers consistently explain why AI pilots succeed in demos and fail in production.

The first is data fragmentation. Most enterprise data is distributed across Salesforce, SAP, Workday, document repositories, data lakes, and dozens of SaaS tools, none of which were ever designed to be jointly queryable. Fewer than 20% of organizations have the data readiness required to support enterprise AI at scale. When a language model cannot access your actual data, it cannot solve your actual problems. It can only speculate.

The second is the absence of a learning loop. The MIT NANDA report identified this as the core technical barrier: “Most GenAI systems do not retain feedback, adapt to context, or improve over time.” This is why demos impress and production disappoints. A demo is stateless, it performs on a clean, curated set of inputs. A production system must remember what worked, adapt to what didn’t, and improve continuously. Most enterprise AI deployments have no mechanism for this.

The third barrier is workflow inertia. McKinsey found that workflow redesign is the single strongest predictor of EBIT impact across 25 organizational factors it studied, yet only 21% of firms have redesigned any workflows as part of their AI deployment. Layering AI onto broken or unchanged processes does not fix the process. It inherits the process’s limitations and adds new complexity.

The fourth is a governance vacuum. Without embedded policies and controls, AI in the enterprise is a liability as much as an asset. IBM’s 2025 Cost of a Data Breach report found that 63% of organizations experiencing AI-related security incidents lacked proper AI access controls. Gartner cites inadequate risk controls as one of the primary drivers of pilot abandonment.

The Shadow AI Economy: A Hidden Signal

One of the most telling findings in recent research is not about official AI programs at all. It is about what employees are doing on their own.

The MIT NANDA report found that while only 40% of companies have official enterprise LLM subscriptions, over 90% of employees use personal AI tools for work, tools they are hiding from IT. Microsoft and LinkedIn’s 2024 Work Trend Index found that 75% of knowledge workers use AI at work, and 78% of those bring their own tools, bypassing corporate systems entirely.

This is not a security failure. It is a signal. Employees are finding real value in AI tools when those tools are accessible, flexible, and designed for the way they actually work. The problem is that individual productivity gains captured through shadow AI never translate into enterprise outcomes. The data that flows through these tools is not governed, the insights are not shared, the actions are not connected to enterprise workflows, and the learning loop that would make the organization smarter never gets built.

The shadow AI economy tells us that the demand for capable AI across every employee is real and urgent. What is missing is not enthusiasm. What is missing is a sanctioned platform capable of delivering that capability at enterprise scale, with governance embedded from the start.

What LLMs Actually Need to Function Across the Enterprise

This brings us to the technical reality that most enterprise AI strategies are avoiding. Language models are extraordinary reasoning engines, but they reason over what they know. And a general-purpose model, trained on the public internet, knows nothing specific about your business.

To function across an enterprise in genuinely capable roles, analytical, operational, strategic, an AI system needs the full context of the business. That means access to the complete data ecosystem: structured operational data, unstructured documents, real-time event streams, historical performance records, customer interaction histories, and the policies and controls that define how the business operates.

It also means that the AI must be able to operate within those policies autonomously. Not just retrieve information, but act on it. A customer service workflow that queries a language model and displays the response is not an AI system. A system that receives a customer inquiry, retrieves the relevant policy and account context, determines the appropriate resolution, takes the action, logs the outcome, and learns from the feedback, that is AI functioning in a real operational role.

Building that capability requires a vertically integrated approach to data and AI: a stack in which the data ecosystem, the governance layer, the agent framework, and the interface for non-technical users are all designed to work together from the ground up. This is not something you can assemble by integrating six point solutions. The integration gaps between those solutions are exactly where the learning loop breaks, where governance fails, and where pilots stall.

The Contextual Layer: Where Complex Agents and Workflows Are Built

The organizations that are capturing real value from AI, the roughly 5% that McKinsey identifies as high performers, share a common architectural commitment. They have built what we think of as the contextual layer: a governed, continuously updated representation of the business that AI agents can reason over, act within, and learn from.

This layer is not a database. It is not a vector store. It is the synthesis of everything the AI needs to understand the business in the way a deeply experienced operator would: what the processes are, what the policies say, what the data means in context, what happened the last time this situation arose, and what the constraints are on any action it might take.

When this layer exists, complex workflows become possible. Predictive maintenance systems that monitor asset condition across an entire fleet, surface anomalies before failures occur, trigger work orders, track resolution, and feed the outcomes back into the model’s understanding. Operations optimization systems that continuously balance cost, capacity, and customer commitments across a logistics network in real time. Strategic planning workflows that synthesize market signals, internal performance data, and competitive intelligence into decision-ready analysis, not as a one-time report, but as an ongoing process.

These are not speculative capabilities. They are what AI can do when it has access to the full context of the business and is permitted to function in fully autonomous roles. But they require the contextual layer to exist first. And building that layer requires treating data access, governance, and integration as first-class engineering problems, not afterthoughts.

Why Broad Enterprise Adoption Requires a Different Kind of Interface

There is one more dimension of this challenge that often goes unaddressed: the interface.

Most enterprise AI tools today are built for technical users. They require prompt engineering, query syntax, or at minimum a comfort with ambiguity that most non-technical employees do not have. The result is that AI capability in most organizations is concentrated in a small cohort of data scientists and AI engineers, while the majority of employees, operations managers, field supervisors, finance analysts, sales leads, customer service representatives, remain outside the system.

This is a profound constraint on enterprise value. The operations manager who oversees a manufacturing facility knows where the bottlenecks are. The logistics coordinator who manages carrier relationships knows which lanes are underperforming. The claims adjuster who has handled thousands of policies knows which patterns predict fraud. These are the people whose daily decisions, multiplied across an organization, determine business outcomes. Excluding them from AI is not a minor inefficiency. It is the single largest reason enterprise AI programs fail to move the P&L.

A Chat UI designed specifically for non-technical users, one that translates natural language into structured queries, surfaces the right context without requiring the user to know where to look, and delivers actionable insights rather than raw data, is not a convenience feature. It is the mechanism by which AI capability becomes available to every employee, in every role, across every workflow.

From Growth Theater to Real Transformation

The path from Growth Theater to real adoption is not mysterious. The research is clear about what separates the organizations that are capturing value from those that are not.

They buy proven platforms rather than building from scratch. MIT NANDA found that purchasing AI tools from specialized vendors succeeds roughly twice as often as internal builds. They focus first on back-office automation, the BPO replacement, the document review, the procurement workflow, where the economics are clearest and the governance requirements are most manageable. They redesign workflows before deploying AI, rather than layering AI onto unchanged processes. They transfer ownership from central IT to the line managers who understand the business problems. And they set transformation objectives, not efficiency objectives.

Most importantly, they treat the data foundation as the mission. Not the model. Not the interface. The data: governed, accessible, complete, and continuously maintained. Because AI can only be as intelligent as the information it has access to, and an AI system that cannot see the full context of the business can only ever answer questions about it.

At Datafi, everything we build is grounded in a single conviction developed over years of working directly with data and AI in production environments: transformative outcomes from AI come not from deploying smarter models, but from giving those models access to the complete, governed reality of a business and enabling them to act on it. That is the difference between AI that answers questions and AI that solves problems. And that difference is what separates the 5% from the 95%.

Datafi is building the Business AI Operating System, a vertically integrated data and AI platform that gives enterprises the full-context data ecosystem, embedded governance, and non-technical Chat UI required to deploy autonomous AI agents across every role and workflow. Learn more at datafi.co