Insights

Billions invested. Ninety-five percent returning nothing. A landmark MIT study reveals why enterprise AI keeps stalling, and makes the case that the culprit is not the technology but the absence of the planning and guardrails that would allow it to succeed.

A recent MIT study has confirmed what many practitioners already suspected but few organisations have been willing to say aloud: the GenAI failure rate is not a technology problem. It is a management problem. Ninety-five percent of AI pilots deliver no measurable business impact, not because the models are inadequate, but because the organisations running them treated procurement as a substitute for planning, and a pilot as a substitute for a decision.

The report, The GenAI Divide: State of AI in Business 2025, draws on structured interviews with 52 organisations, survey responses from 153 senior leaders, and analysis of more than 300 publicly disclosed AI implementations. Its central finding is blunt. Adoption is nearly universal, transformation is rare, and the gap between the two is not closing. If anything, it is widening, because the organisations piling into AI in 2025 and 2026 are making the same foundational mistakes as those who went first.

That 95% figure is worth sitting with. The instinct is to absorb it into a broader narrative about AI being "early" or "evolving," a collective agreement to treat failure as a stage rather than a signal. But 95% is not a growing pain. It is a structural outcome. And structural outcomes have structural causes.

The cause is not, as vendors often think, that organisations simply need better tools or more sophisticated models. Tools like ChatGPT, Anthropic's Claude, and Microsoft Copilot are genuinely capable and genuinely popular, with nearly 40% of organisations reporting deployment. But they boost individual productivity without moving the metrics boards actually measure — revenue, margin, cost structure. The technology lands. The organisation around it does not move. And the reason it does not move is that no one planned for what movement would require.

That COO acquired a capability without a mandate. No clear answer to what was supposed to change, who was responsible for changing it, or what the organisation would look like when it had. Without those answers, AI does what all ungoverned tools do. It gets used for whatever is convenient, by whoever finds it first, with no connection to the outcomes that actually matter.

Across interviews, surveys, and public data, MIT's researchers identified four structural forces that separate organisations on either side of the divide. Looked at together, they form a consistent pattern. In every case, the technology was acquired before the conditions for its success were established. No clear ownership. No defined scope. No guardrails around data, decisions, or outcomes. No plan for what the organisation needed to look like on the other side.

Sector-level inertia gets blamed on regulation, on risk culture, on legacy systems. Those are real constraints. But they are not the primary explanation for why Healthcare and Financial Services, sectors with enormous AI budgets and serious technical talent, are producing documentation pilots and backend automation while leaving clinical and customer models entirely untouched. The more honest explanation is that no one in these organisations has been willing to define, in writing, which decisions AI is authorised to influence and which it is not. That definition requires committing to a position on accountability that most executive teams are not yet prepared to hold. So instead they run pilots. Pilots are safe. Pilots do not require anyone to be responsible for the outcome.

No rocket science here. Large organisations are slower with more process, more stakeholders, more inertia. But that reading stops short of what the 90-day mid-market figure actually tells us. Smaller organisations do not have less complexity; they have less room to avoid commitment. When a mid-market company launches an AI pilot, someone senior is personally exposed if it fails. That exposure forces the questions such as who owns this, what does success look like, what happens to the people whose work changes... all to be answered before the project starts rather than deferred indefinitely. Enterprise governance structures, paradoxically, make it easier to launch without deciding. There are enough layers, enough committees, enough shared accountability that no single person needs to actually commit to anything. The pilot becomes a way of appearing to move while the organisation stays still.

The investment bias toward sales and marketing is usually explained as a measurement problem. Front-office metrics are legible; back-office returns are not. That is true, but it is only half the story. The other half is that AI in sales and marketing is politically safe in a way that AI in operations is not. A tool that generates outbound emails does not threaten anyone's role in a visible way. A tool that automates procurement decisions, restructures how finance approvals flow, or replaces a document review process touches someone's territory, someone's team, someone's sense of what their job is. The investment bias is not just about measurement difficulty. It is about the fact that the highest-ROI applications of AI are also the ones that require organisations to have honest internal conversations they have been avoiding. Back-office automation does not stall for lack of evidence. It stalls for lack of will.

The implementation advantage is real, but it is worth being precise about why. It is tempting to attribute it to external partners having better technology or deeper AI expertise. Sometimes that is true. More often, the advantage is structural. An external partner has a commercial incentive to reach production, because that is typically when the real contract value is realised. Internal teams have no equivalent forcing function. They can extend the pilot indefinitely without anyone losing anything, until the budget cycle ends and the whole initiative gets quietly defunded. The external partner also has something else, that is standing to ask the uncomfortable questions. A vendor asking "who is accountable if this goes wrong?" is doing due diligence. An internal employee asking the same question is being difficult. That asymmetry matters more than most organisations acknowledge.

The data on enterprise-grade, task-specific AI tools shows precisely where the value disappears.

MIT attributes this collapse to a learning gap where tools that cannot retain context, adapt to feedback, or evolve alongside the workflows they are meant to support. That is a real problem. But it does not explain the gap between 20% and 5%. What explains that is simpler and harder to admit. Most organisations reach pilot stage without having answered the questions that production actually demands. Who is liable if the AI output is wrong and a decision gets made on it? What data is the system permitted to access, and who decided that? If a frontline employee believes the tool is producing something dangerous or biased, what is the process? These are not questions for legal to handle at the end. They are the conditions of responsible deployment, and in the absence of answers, staying in pilot is not caution. It is an organisation using the language of progress while avoiding the obligations that progress requires.

The CIO dismissing most demos as "wrappers or science projects" is right about the tools. But there is a harder question underneath it. If the enterprise tools are so poor, why are their own people using personal AI accounts to do their jobs every day, without being asked to? That is not just a technology gap. It is a culture gap. It says something about whether people feel equipped, trusted, and supported enough to do their best work through official channels, or whether they have quietly concluded that the fastest path to getting something done is to route around the organisation entirely.

Ninety percent of employees in surveyed organisations use personal AI tools for work tasks regularly, without IT approval. The standard organisational response is anxiety to shadow IT, data risk, compliance exposure. Those concerns are legitimate. But they are the wrong first question. The right first question is what it means that your people solved the adoption problem on their own, without a rollout plan, a training programme, or executive sponsorship.

It means the technology is not the barrier. It never was. Your employees found these tools, learned them, integrated them into their daily work, and by the MIT data, are getting genuine value from them. They did this in the absence of any organisational support. The implication is uncomfortable. The 95% failure rate is not a consequence of AI being hard to use. It is a consequence of organisations making it hard to use well. The shadow economy exists because the official path either does not exist or is so burdened with process that individuals bypass it entirely.

Locking down shadow AI without replacing it with something better is not a solution. It is an own goal. The organisations creating sanctioned paths that are genuinely faster, safer, and more capable than personal accounts are the ones who will bring shadow usage inside, make it visible, and actually learn from it. Those that respond with restriction alone will drive it further underground, lose the intelligence it contains, and still carry all of the data risk.

The organisations that have crossed the GenAI Divide are finding their highest returns in back-office functions — the places receiving the least investment attention and requiring the most difficult internal conversations. The MIT report documents the outcomes from deployments that actually reached production.

The MIT report emphasises that these gains arrived without significant headcount reduction, that ROI came from shrinking third-party spend rather than cutting internal staff. That is true for the organisations in their sample, and it is an important counterpoint to the AI-as-job-destroyer narrative. But it should not be read as a universal guarantee. What it actually reflects is that the organisations achieving real returns found a path that generated economic value without triggering the internal resistance that comes with visible workforce impacts. That path is eliminating BPO spend, cutting agency fees, replacing external contractors and exists precisely because it is politically navigable. The harder applications, the ones that touch internal roles directly, are still mostly in the avoided pile. The returns documented here are real. They are also, in some sense, the easy ones.

The organisations that have moved to the right side of the divide share something more specific than good intentions or adequate budgets. They made concrete decisions, early, that most organisations defer until they are forced. Three of those decisions stand out consistently in the MIT data.

The report described the "Agentic Web" as an emerging phase where AI systems that maintain persistent memory, learn from interactions, and coordinate autonomously across workflows, vendors, and platforms, with frameworks like NANDA, Model Context Protocol (MCP), and Agent-to-Agent (A2A) as early infrastructure. That phase has now arrived. These protocols are active components in real enterprise deployments, and the organisations that treated agentic AI as a future concern are finding it is a present one.

The implications for governance are significant and largely underdiscussed. A standard GenAI tool produces an output that a human reviews and acts on. An agentic system acts on its own through accessing data, making decisions, triggering downstream processes all within parameters set at design time. The quality of those parameters, and the robustness of the guardrails around them, now determines not just whether the AI is useful but whether it is safe. Every governance gap that organisations have been deferring under the cover of "we're still in pilot" becomes acute when the system starts taking actions rather than producing suggestions.

This CIO is right about the switching costs. Organisations that have built real feedback loops between their AI systems and their workflows are accumulating institutional knowledge that compounds. The gap between them and organisations still cycling through pilots is growing every month, not shrinking.

The GenAI Divide is not closing on its own. Organisations that skipped the planning, diffused the accountability, and deferred the guardrails are not going to catch up by running more pilots. The gap compounds. Every month an organisation with real feedback loops embedded in its workflows pulls further ahead, and every month of continued experimentation without commitment narrows what catching up can actually mean. The organisations on the right side got there by making harder decisions earlier, before they were forced to. That option still exists.