MIT NANDA's 2025 Report Says 95% of Organizations Saw No GenAI Return

The document describes a multi-method design: a review of more than 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected at four industry conferences. Those are the figures stated in the report itself; some early news coverage used different interview and survey counts.

The report's headline number is therefore a finding from this preliminary research program, not a universal failure rate for every enterprise AI project. Company-level data and quotations were anonymized, and the authors said the views were their own rather than those of affiliated employers. That context matters when applying the 95% figure outside the studied sample.

The authors attributed the divide less to model quality or regulation than to implementation. They pointed to brittle workflows, limited ability to retain feedback or context, and poor alignment with day-to-day work. More than 80% of organizations had explored or piloted general-purpose tools such as ChatGPT and Copilot, the report said, but only 5% of evaluated task-specific enterprise systems reached production.

The report also found a correlation between delivery approach and deployment. In its interview sample, customized tools built through external partnerships reached deployment about 67% of the time, compared with about 33% for internally built tools. The authors explicitly cautioned that these were self-reported outcomes, definitions of success varied, and the relationship did not prove that buying caused better results.

The useful lesson is narrower than the viral claim that AI projects simply fail. The report's evidence favors projects tied to a specific workflow, measurable operating outcomes, persistent context and accountable business owners. It also suggests separating individual productivity gains from effects that reach a profit-and-loss statement. For data and ML teams, the practical test is whether a pilot has an agreed baseline, a production owner and a credible measurement window before its results are generalized.