MIT report: 95% of AI pilots fail to deliver returns

A new report from MIT’s NANDA initiative finds that nearly 95% of the enterprise generative AI pilots it analyzed produce no measurable impact on the bottom line. Just 5% manage to accelerate revenue quickly—a gap that calls into question the distance between investment in AI and the results companies can demonstrate.

The study, titled The GenAI Divide: State of AI in Business 2025, draws on 150 interviews with executives, a survey of 350 employees and an analysis of 300 public deployments. Its central argument is that the main obstacle is not model capability, but how organizations try to incorporate these systems into their processes.

Heavy investment, little bottom-line impact

The report places its findings in a market that has already attracted between $30 billion and $40 billion in enterprise investment in generative AI. Despite that spending, most projects remain stuck before reaching production or fail to show clear improvements in revenue, costs or profits.

That 95% figure requires an important clarification: it does not mean that 95% of all AI tools are useless or that they produce no isolated improvements. The metric refers to the pilots studied and their ability to generate a measurable effect on the profit and loss statement, known in business as P&L.

An application may save minutes when drafting emails or summarizing documents without turning that time savings into lower costs, more sales or additional capacity. The leap from individual productivity to an organization-wide financial return is precisely where many projects get stuck.

The problem is integration, not just the model

The researchers describe a learning gap: the tools do not retain enough context, do not adapt well to internal procedures and do not improve through the company’s day-to-day use. At the same time, organizations are not redesigning their processes to take advantage of them.

General-purpose services such as ChatGPT can be useful to an individual because they handle many different tasks and allow for flexible workflows. In a company, however, a tool must connect to databases, permissions, billing systems and approval workflows. It also needs to operate according to repeatable criteria and leave an auditable trail.

This difference helps explain why a convincing demo is not the same as an enterprise product. A model may draft a correct response and still fail as a system if it lacks the customer’s history, uses outdated information or requires an employee to manually review every step.

Buying and partnering work better than building from scratch

The adoption path also makes a difference. According to the report, projects purchased from specialized vendors or developed through partnerships succeed approximately 67% of the time. In-house builds achieve success at a rate equivalent to only one-third of that figure.

The comparison does not, by itself, prove that buying is always better. Companies developing their own systems often face more complex requirements, particularly in finance and other regulated sectors. But it does reveal the cost of trying to build a complete platform without the necessary expertise, prepared data or a sufficiently focused business problem.

Successful cases tend to focus on a specific task and have operational owners, rather than relying solely on a central AI lab. Knowledge of where time or money is being lost is usually held by the departments that carry out the process every day.

Budgets are flowing to the wrong areas

More than half of enterprise generative AI budgets are being allocated to sales and marketing, according to the study. Yet the clearest returns are appearing in internal tasks: administrative automation, cutting spending on outside agencies, replacing outsourced processes and streamlining operations.

The difference makes economic sense. Generating campaigns or marketing messages is visible and easy to showcase, but attributing an additional sale to AI is difficult. By contrast, eliminating an outside invoice or reducing the hours required to process documents creates a direct saving that is easier to measure.

The report also identifies the spread of shadow AI, meaning employees’ use of tools such as ChatGPT without authorization or official integration. The phenomenon indicates genuine demand, but it raises risks involving confidential data, access controls and regulatory compliance.

Jobs are changing first through vacancies, not mass layoffs

The labor effects observed so far are concentrated in customer support and administrative roles. Rather than announcing large cuts directly tied to AI, some companies are choosing not to fill positions left vacant when employees leave the organization.

The impact is greater in roles that had already been outsourced and were treated as lower-value work. This allows companies to cut costs without equally visible internal reorganizations, although it shifts the impact to vendors and outsourced workers.

A useful figure, but not a verdict on all AI

The sample combines interviews, employees and public deployments, so the 95% figure should not be mechanically extrapolated to every company and sector. Public projects may differ from internal systems that companies do not disclose, and some returns take longer to materialize than others.

Even with those limitations, the report identifies a recognizable problem: many companies have started by buying access to a model and only afterward looked for ways to use it. The technology arrives before the process, data, owner and financial metric have been defined.

More advanced organizations are already experimenting with AI agents that can remember information and take actions within defined boundaries. But those capabilities also increase the need for controls and reliable integration. The next test will not be whether an agent can complete a demo, but whether it can do so consistently and safely, while generating savings or revenue that a CFO can record.