The most useful number in enterprise AI right now is a failure rate. MIT's Project NANDA put it at 95%: the share of corporate generative AI pilots showing no measurable return. Companies had spent $30 to $40 billion getting there.
Here is the part of that finding most coverage skipped, and the actual point of this post: the researchers did not blame the models. The pilots failed at integration, meaning the AI never became part of how work actually flows. It sat beside the workflow as a demo, answered questions in a chat window, and changed nothing about how the invoice got approved or the ticket got resolved.
If you run operations at a small or mid-sized company and you have been waiting for a clear lesson from the enterprise AI spending wave, this is it. The technology purchase is the easy part. The workflow change is the work. And on that specific job, a 60-person firm moves faster than an enterprise.
Key takeaways
- MIT's Project NANDA reviewed over 300 enterprise AI initiatives in 2025 and found 95% delivered no measurable P&L return, despite $30 to $40 billion in investment. The failures were integration failures, not model failures.
- McKinsey's State of AI 2025 found 88% of organizations now use AI somewhere, but only around 6% see significant enterprise-wide impact. The strongest predictor of impact was workflow redesign, which only 21% had done.
- The pattern in the successful minority is consistent: one specific workflow, deep integration, and a tool that keeps context, instead of a generic assistant bolted onto everything.
- Smaller firms hold a real advantage here. Redesigning a workflow across 60 people is a month of work. Across 10,000 people it is a change program.
01What did the MIT research actually find?
The study, titled The GenAI Divide: State of AI in Business 2025, came from MIT's Project NANDA and drew on a review of more than 300 publicly disclosed initiatives, 52 organizational interviews, and 153 executive surveys. The headline: 95% of integrated AI pilots produced no measurable P&L impact. The 5% that did were extracting real value, in some cases millions.
The difference between the two groups was not model choice or budget. The failing majority deployed generic tools that demoed well and broke inside real workflows. The successful minority picked one high-value workflow, integrated deeply, and shipped tools that kept memory and context across uses instead of starting from zero every session.
One caveat belongs next to that 95%, because a number this quotable deserves scrutiny. The report circulated as a preliminary document and several analysts challenged how the failure rate was defined, arguing it measured formal pilot programs while ignoring the informal AI use spreading through the same companies. Fair challenge. But the report's core mechanism, that value follows deep workflow integration and not tool deployment, holds up independently, because a second large study found the same thing.
95%
Enterprise generative AI pilots with no measurable P&L return (MIT Project NANDA, The GenAI Divide, 2025)
02Is adoption the same thing as impact?
No, and the gap between the two is now the defining fact of business AI. McKinsey's State of AI 2025 survey found 88% of organizations using AI in at least one function. Only around 6% were seeing significant enterprise-wide financial impact. Adoption is nearly universal. Impact is rare.
The same survey tested 25 organizational attributes to find what separates the two groups. The strongest single predictor of bottom-line impact was workflow redesign: rethinking how the work flows before adding the tool. Only 21% of organizations had done it. Nearly 80% layered AI on top of existing processes unchanged, which is how a company ends up with universal AI adoption and no measurable result.
Read those two studies together and the lesson stops being about AI at all. Tools amplify the process they are dropped into. A well-designed workflow gets faster. A vague one produces vague output faster. We made the same argument about accountability in enterprise AI from the governance side: the organizations that struggle with AI outcomes are usually struggling with process ownership first.
Adoption is nearly universal. Impact is rare. The difference is workflow redesign.
03What does the successful 5% actually do?
The pattern across both studies is specific enough to act on, and it looks nothing like buying a company-wide assistant seat.
Successful pilots start narrow: one workflow with real money or real hours attached, not a general productivity mandate. They integrate where the work already happens, inside the ticketing system, the approval chain, or the document flow, rather than in a separate chat window the team has to remember to visit. They keep context, so the tool knows what happened last week without being retold. And they hold a feedback loop, meaning someone owns checking output quality and correcting the system, the way you would supervise a new hire.
In our own operations we automated one narrow workflow at a time: a weekly performance report that assembles itself and lands in the inbox every Monday at 5pm, and a content pipeline in which metadata that used to be typed by hand is generated, checked by a person, then published. Small targets, boring by design, each one measured against the hours it used to take. That last clause is the discipline most pilots skip: if you cannot name the hours or the error rate a pilot should reduce, you cannot know whether it worked, and MIT's 95% suggests most companies never could.
| Failing pilot pattern | What the successful minority does instead |
|---|---|
| Generic assistant deployed everywhere at once | One workflow with measurable hours or money attached |
| Tool sits in a separate window beside the work | Integrated where the work already happens |
| Every session starts from zero context | Keeps memory and context across uses |
| Nobody owns output quality | A named person reviews, corrects, and tunes |
| Success undefined at launch | Baseline measured before the pilot starts |
04What NOT to do
Do not run a pilot without a baseline. The single most common failure we see is a pilot nobody measured a before state for, which guarantees the after state is an argument instead of a number.
Do not automate a workflow you have not mapped. This is the same rule we apply to approval chains: automation multiplies the process it is given, including the broken parts.
And do not treat AI output as finished work in anything customer-facing or compliance-adjacent. The successful pattern keeps a person in the loop precisely where errors are expensive. A pilot that removes review from a high-stakes workflow is not efficient, it is unpriced risk.
05Getting started: a pilot that can actually prove itself
1. Pick one workflow where the hours are countable: report assembly, data re-entry between systems, first-draft responses to routine requests.
2. Measure the baseline for two weeks before touching anything. Hours spent, error rate, cycle time. This is the step that makes every later claim checkable.
3. Redesign the workflow on paper first. Decide where the tool acts, where a person reviews, and what happens when the tool is wrong.
4. Integrate into the system the team already uses. If using the pilot requires opening a new tab, adoption will quietly die there.
5. Review at 60 days against the baseline, and be willing to kill it. A killed pilot with clean numbers teaches more than a zombie pilot nobody will call failed.
Frequently asked questions
Integration, not model quality. MIT's 2025 research found failing pilots deployed generic tools beside the workflow instead of inside it, kept no context between uses, and had no owner measuring results. The models performed; the organizations never rewired the work around them.
Buy the model, build the workflow. Foundation models are commodities you rent through an API. The value sits in the integration layer: how the tool connects to your systems, where human review sits, and how context is kept. That layer is specific to your operation, which is why generic tools underdeliver on it.
Deciding how the work should flow before adding the tool: which steps disappear, which get automated, where a person reviews, and who owns the outcome. McKinsey's 2025 survey found it was the strongest single predictor of financial impact from AI, and that only 21% of organizations had done it.
Baseline first: two weeks of measured hours, error rates, and cycle time on the target workflow before the pilot starts. Then compare at a fixed review date against the same measures. Without the baseline there is no ROI calculation, only impressions.