The useful reading of the MIT finding is not that generative AI has failed. It is that enterprise AI has run into the same old enterprise problem: a tool that does not fit the work is not made valuable by being technically impressive.
The report in question has been widely identified as MIT NANDA’s The GenAI Divide: State of AI in Business 2025. As reported by Tom’s Hardware in August 2025, the MIT work examined 150 interviews, a survey of 350 employees, and 300 public AI deployments, and found that only about 5% of enterprise generative-AI pilots were producing rapid revenue acceleration or measurable profit-and-loss impact.
This is one report, not settled consensus about every company using AI. The finding is still worth taking seriously because it says something precise about the gap between model capability and operational value. The researchers’ argument, as covered by several outlets, was not that the models were useless. It was that most task-specific enterprise tools were too brittle, too disconnected from actual workflows, or too poorly defined to improve business outcomes.
The failure was not mostly a model story
That distinction matters. A company can buy access to a strong model and still build a weak product around it. It can add a chatbot to a process that needed a workflow change. It can automate a fragment of a job while leaving the costly handoff untouched. It can make a tool available to employees without giving it the data, context, permissions, feedback loops, or narrow objective needed to do useful work.
In Axios’s account of the MIT study, the research was framed as a warning to investors who were assuming that heavy AI spending would automatically translate into heavy AI returns. Axios reported that MIT researchers studied 300 public initiatives and found that companies buying external AI tools were often doing better than those building internal pilots.
The buy-versus-build point should not be turned into a rule for every company. Some regulated or technically complex organisations may have good reasons to build internally. But the pattern is instructive. A vendor product that already knows a workflow, captures feedback, and ships into a defined operational problem may be more useful than an internal tool that demonstrates model access without changing the process around it.
The failure mode is familiar in enterprise software. A tool gets evaluated as a capability instead of as part of a job. The demo works. The proof of concept works. The pilot works well enough to justify another meeting. Then the tool reaches the messy middle of the organisation, where data is incomplete, exceptions are common, incentives are misaligned, and nobody has redesigned the work.
The phrase to watch is learning gap
Several reports on the MIT finding highlighted what the researchers called a learning gap. Investor’s Business Daily reported that the obstacle was not simply infrastructure, regulation, or talent, but the inability of many systems and organisations to adapt to real use. Tools failed to retain feedback, improve through use, or fit into the context of the business.
That is a quieter diagnosis than the public AI debate usually allows. It is not the claim that models cannot reason, or that AI is over, or that every deployment is waste. It is the claim that many enterprise tools behave as if work were a prompt-response exchange, when work is usually a chain of decisions, approvals, exceptions, data checks, customer constraints, and accountability.
A support tool that drafts replies but does not learn which replies are corrected by experienced agents remains shallow. A sales tool that generates summaries but does not connect to pipeline reality remains decorative. A finance tool that produces analysis without traceability may increase review work rather than reduce it. A procurement tool that cannot handle exceptions may help with the easy cases and leave the expensive cases untouched.
The learning gap is therefore not just technical memory. It is organisational memory. The system has to know what happened after its output was used. Was it accepted, edited, rejected, escalated, ignored, or corrected? Which corrections mattered? Which cases repeated? Which users trusted it, and why? Without that loop, the tool may remain frozen at pilot quality.
Workflow fit beats generic capability
The MIT finding also pushes against a common enterprise habit: buying horizontal capability and hoping departments will discover value on their own. Generic tools can help individuals, and they may be useful for writing, summarising, searching, coding, or analysis. But measurable business returns usually require a connection to a defined operational metric.
The difference is not cosmetic. “Help employees use AI” is not the same as “reduce the average time to resolve a Tier 2 support ticket without lowering customer satisfaction.” “Give analysts a chatbot” is not the same as “cut the manual reconciliation cycle by two days while preserving auditability.” One is a technology adoption goal. The other is an operating problem.
A 2025 MIT Sloan article on generative AI in finance made a related point from a different angle: AI can help with tasks such as accounting and hiring, but governance and the shape of the process matter. That is the enterprise reality the 95% figure points toward. Tools have to land inside actual controls, workflows, and decision rights.
This is why some successful deployments look less exciting than the pitch-deck version of AI. Back-office automation, customer-service triage, document processing, coding assistance, compliance review, invoice handling, claims routing, demand forecasting and internal search may not sound like the end of work. They can still be where returns appear, because the task is specific and the measurement is closer to the process.
The measurement problem cuts both ways
It is also worth being careful with the phrase “measurable business returns.” A project may improve worker experience, reduce annoyance, speed up a task, or improve information access without immediately showing up in profit and loss. Conversely, a company may claim AI value by attributing ordinary process improvements to a tool that did less than advertised.
The MIT estimate appears to focus on measurable business impact, not every possible use of generative AI. That is an important boundary. Employees may be using public tools informally in ways that save small amounts of time, while official enterprise initiatives fail to show financial returns. Shadow use can be real, useful, and hard to measure. It can also create security and governance problems.
McKinsey’s State of AI research has repeatedly found that many organisations are adopting AI while still working through governance, risk and value-capture questions. That broader pattern is consistent with the MIT warning: adoption is easier to count than return.
For executives, this creates a less comfortable dashboard. Number of pilots is not enough. Number of users is not enough. Number of prompts is not enough. The harder question is whether a tool changes a business process in a way that can be observed, trusted, and repeated.
The lesson for vendors is harsher than the lesson for buyers
For vendors, the MIT report is a warning against selling model access as if it were operational transformation. Enterprise buyers are not short of demos. They are short of tools that survive contact with legacy systems, fragmented data, compliance requirements, security constraints, and frontline exceptions.
The winning enterprise AI product may be less glamorous than a general assistant. It may have narrow permissions, strong logging, boring integrations, clear escalation paths, and a visible record of what it learned from feedback. It may solve one expensive workflow instead of promising to change the company.
For buyers, the lesson is equally practical. A task-specific AI initiative needs a task specific enough to test. It needs a baseline, a business owner, a feedback loop, and a clear decision about what happens when the model is wrong. It also needs a reason to exist beyond the fact that generative AI is available.
This is where the 95% figure is most useful. It strips away the assumption that AI investment automatically becomes AI value. The underlying models may keep improving. That will not, by itself, fix a tool that cannot learn from users, fit the system of work, or solve a problem the business can actually name.
A sober reading is better than a backlash
The wrong conclusion is that companies should stop experimenting. The better conclusion is that experiments should be smaller, sharper, and more accountable. A useful pilot is not a theatre piece proving that a model can generate plausible output. It is a test of whether a particular workflow improves when the tool is inserted into it.
There will be enterprise AI returns. Some are already visible in specific functions and companies. But the MIT report suggests that returns are not evenly distributed across the wave of task-specific initiatives. They appear where the problem is narrow, the workflow is understood, the tool adapts to feedback, and the deployment changes how work actually happens.
That is less dramatic than the broad promise of generative AI. It is also more useful. The question is no longer whether the model can produce an answer. The question is whether the answer moves through the business in a way that lowers cost, raises revenue, reduces risk, or improves the quality of a decision. Most enterprise pilots, according to the MIT estimate, had not yet crossed that line.
