A new study from MIT’s Project NANDA has revealed a striking gap in the state of artificial intelligence in business, with the vast majority of corporate investments in generative AI (GenAI) producing little or no measurable return.
Despite enterprise spending estimated between $30 billion and $40 billion, 95 per cent of initiatives are failing to deliver profit-and-loss impact. Researchers have termed this phenomenon the “GenAI Divide”.
The report, based on analysis of over 300 public AI initiatives, 52 structured interviews, and survey responses from 153 senior executives, shows that adoption is high but transformation remains rare. While widely used tools such as ChatGPT and Microsoft Copilot are embedded across many organisations, their impact is largely confined to individual productivity. Enterprise-grade or custom systems, meanwhile, face high rejection rates due to integration challenges and limited adaptability.
High adoption, low disruption
The data shows that more than 80 per cent of organisations have experimented with or piloted general-purpose AI tools, with nearly 40 per cent reporting deployment. However, these pilots have rarely translated into large-scale transformation. Only two of nine industries examined—technology and media—show signs of structural change. Other sectors, including healthcare, financial services, and manufacturing, have seen little beyond efficiency pilots and limited automation.
MIT’s researchers developed an “AI Market Disruption Index” to measure structural change. While media and technology scored highest, industries such as energy and materials registered near zero adoption. A mid-market manufacturing executive summed up the broader picture: “The hype on LinkedIn says everything has changed, but in our operations, nothing fundamental has shifted.”
The pilot-to-production chasm
The divide is clearest in the failure of pilots to reach production. Around 60 per cent of companies evaluated enterprise-grade AI tools, but only 20 per cent piloted them, and just 5 per cent reached production. By contrast, general-purpose tools such as ChatGPT demonstrated higher adoption, but executives stressed that these systems lack the memory and adaptability needed for mission-critical work.
This inability to retain feedback and evolve was identified as the core obstacle. As one CIO commented: “We’ve seen dozens of demos this year. Maybe one or two are genuinely useful. The rest are wrappers or science projects.”
The shadow AI economy
An unexpected finding was the extent to which employees are driving unofficial adoption through personal tools. Although only 40 per cent of companies reported paying for large language model subscriptions, more than 90 per cent of employees surveyed said they used consumer AI tools regularly for work. This “shadow AI economy” often provides greater day-to-day value than official initiatives.
Forward-looking organisations are beginning to analyse shadow usage to guide procurement, identifying which tools genuinely deliver productivity before formalising investment.
Investment bias
Corporate AI budgets are heavily concentrated on sales and marketing, with executives allocating around 70 per cent of hypothetical spend to these areas. By contrast, back-office functions such as procurement, finance, and compliance—where automation often yields the clearest returns—receive far less funding. The report suggests this is due to the visibility of sales metrics, which are easier to attribute directly to board-level targets, while internal efficiencies are harder to quantify.
One pharmaceutical procurement officer observed: “If I buy a tool to help my team work faster, how do I justify that to my CEO when it won’t directly move revenue?”
Why most pilots stall
Executives and users cited consistent barriers: poor user experience, lack of workflow integration, and models that fail to learn. Workers who rely on ChatGPT privately often find internal tools rigid or less effective, leading to rapid abandonment. While AI excels at quick tasks such as drafting emails or summarising documents, 90 per cent of respondents still preferred human colleagues for complex, multi-week projects.
The report identifies this “learning gap” as the defining feature of the GenAI Divide. Current systems are static: they do not adapt, remember context, or improve with feedback. Emerging “agentic” AI, which embeds persistent memory and learning, is seen as the critical bridge.
What works: buyers and builders
The organisations crossing the divide are those treating AI as a business service rather than a software purchase. Successful buyers decentralise authority to line managers, demand deep workflow integration, and hold vendors accountable to business outcomes. They prioritise partnership over procurement.
On the supply side, startups that succeed focus on narrow, high-value use cases—such as call summarisation, document automation, or repetitive code generation—rather than broad, generic offerings. They grow by embedding deeply into workflows and adapting continuously to feedback.
External partnerships are proving twice as successful as internal builds, with deployments achieved faster and at lower cost. Vendors that demonstrate trust, integration capability, and the ability to improve over time are securing multi-million-dollar contracts within months.
The emerging Agentic Web
The report concludes that the next phase of AI adoption will be defined by the rise of interconnected “agentic” systems. Frameworks such as Model Context Protocol (MCP), Agent-to-Agent (A2A), and NANDA are enabling networks of specialised agents that can coordinate, learn, and transact autonomously. This “Agentic Web”, researchers argue, may decentralise workflows in the same way the original internet transformed publishing and commerce.
With procurement cycles closing rapidly, MIT’s researchers warn that the next 18 months are critical. Organisations that commit to adaptive, learning systems now will establish competitive moats that are difficult to dislodge. Those that remain trapped in pilots risk being left on the wrong side of the GenAI Divide.
