[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-body-where-the-95-percent-goes":3},"\nThe MIT NANDA initiative's report, The GenAI Divide: State of AI in\nBusiness 2025, produced the statistic that now opens every enterprise AI\nkeynote: roughly 95% of generative AI pilots deliver no measurable profit\nand loss impact. It is the most quoted and most disputed number in\nenterprise AI; the methodology has been publicly challenged, and it is a\nresearch initiative's estimate, not an institutional MIT finding. S&P\nGlobal Market Intelligence supplies the corroborating trend from separate\ndata; 42% of companies abandoned most of their AI initiatives in 2025, up\nfrom 17% the year before.\n\nWhatever the precise figure, the direction is not in doubt, and the way\nthe number is being read is the real problem. In boardrooms it has become\nshorthand for \"AI does not work yet\", and that reading produces one of\ntwo responses: wait for the technology to mature, or buy a more capable\nmodel. Both responses miss what the research actually found.\n\n## What the report actually shows\n\nRead past the headline and the MIT researchers are specific about where\npilots died: overwhelmingly at the workflow boundary, not the model\nboundary. Generic tools produced fluent, plausible output that did not\nfit how decisions were actually made, and organisations quietly routed\naround them. The capability was present. The connection to a decision was\nnot.\n\nThat diagnosis matters because it tells you what will not fix the\nproblem. If the pilots were failing because the models were too weak,\nthe next model generation would rescue them. But they failed because the\nmodel's output never reached a decision. A better model in the same\ndashboard changes nothing, so the 95% will survive every model release.\nThe proof is in the timing. Abandonment climbed in the same year that\nmodel capability jumped.\n\n## What the two wrong responses cost\n\nThe misreading is not a semantic quibble, because each of its responses\ncarries a bill.\n\n**Waiting** treats the 95% as a maturity problem that time will solve.\nBut the constraint the research points at, wiring outputs into owned\ndecisions, does not improve with model releases; it improves with\norganisational work that takes quarters to do. An organisation that\nwaits is not holding its position. It is donating those quarters to\nthe minority already doing the wiring, whose systems are accumulating\nscored decision histories that a late starter cannot buy.\n\n**Upgrading** treats the 95% as a capability shortfall, and it is the\nmore expensive mistake because it looks like action. The new model is\nprocured, the benchmark is better, the demo is impressive, and the\noutput lands in the same dashboard, advisory, unowned and unscored,\nwhere its predecessor's output landed. Budget was spent making the\nunused thing more sophisticated.\n\n> The tell is in the timing: abandonment rates climbed in exactly the\n> year model capability jumped. If capability were the constraint, those\n> curves should have moved in opposite directions.\n\n## The reading that survives the evidence\n\nThe reading that fits the report's own evidence is simpler. The 5% did\nnot have better models. They wired predictions into real decisions: a\nnamed owner, a threshold for acting, and a record of whether acting\nworked. How to do that wiring is its own subject, covered in\n[the flagship essay](/insights/everyone-has-ai-few-get-value). The\npoint here comes first. None of that work can start until the statistic\nis read correctly, and in most boardrooms it is not.\n\n*Read the 95% as \"the models failed\" and you will wait or upgrade.\nRead it as \"the decisions were never wired\" and you have work you can\nstart on Monday.*\n\nWiring predictions to owned, scored decisions is the entire design brief\nof Prophesee. If you want to see what the 5% built, on your own data,\n[start here](/contact).\n",1786786820770]