The headline number that has C-suites rattled
Earlier this year, MIT's NANDA initiative published one of the most circulated AI numbers of 2025. Out of more than 300 enterprise generative AI deployments tracked, 95% delivered no measurable impact on the profit and loss statement. Only 5% extracted real, attributable value.
The same companies have collectively spent between $30 billion and $40 billion on enterprise generative AI in the past two years. Most of them have very little to show their boards.
That is uncomfortable, and it is also the most useful research published on AI ROI all year. Once you stop arguing with the headline, the rest of the report explains exactly why the 95% are stuck and what the 5% are doing instead.
The problem is not the model. It is the learning gap.
The MIT researchers expected to find the usual suspects: model quality, regulation, data privacy, change management. They found something else.
The biggest barrier, by a wide margin, is what they call "the learning gap." Generative AI tools demo brilliantly. They struggle in real workflows because they do not retain feedback, do not adapt to a specific business, and do not improve with use. Every interaction starts from zero.
Inside the 95%, this shows up as the same loop. A team buys a tool. The tool produces an impressive first draft. The tool then makes the same mistake on Monday that it made on Friday, because nothing has been done to teach it the company's edge cases, tone, terminology or constraints. Adoption stalls. The pilot quietly fades.
Inside the 5%, the tool is wrapped in a process that captures corrections, feeds them back, and incrementally fits the system to how the business actually works. The model is not the differentiator. The feedback loop is.
Most of the budget is going to the wrong half of the business
The MIT data also shows where the money is being spent versus where the money is being made.
Roughly half of all enterprise generative AI budgets sit in sales and marketing. That is where the executive sponsors live, and that is where the demos are flashiest. It is also where ROI is hardest to attribute, because pipeline outcomes are slow and noisy.
The 5% who report measurable returns tend to invest in back and middle-office work: finance operations, procurement, contract review, customer support triage, internal knowledge retrieval. The wins are less glamorous, but they are visible inside one quarter, not four. They show up as fewer outsourced hours, faster cycle times and reduced agency spend. Those are numbers a CFO can sign off on.
For leaders trying to find AI ROI fast, the lesson is direct. If you cannot point to a function where the cost line will visibly fall, you are unlikely to be in the 5%.
Buy beats build, by a factor of two
A second pattern in the data is sharper than most people expect.
External partnerships and off-the-shelf vendor solutions reach successful production deployment about twice as often as internally built solutions. According to the report, around two-thirds of vendor-led deployments succeed, against roughly one-third for internal builds.
There is a reason. Enterprise AI tools are deceptively expensive to build well. The model itself is a small share of the work. The hard parts are the integrations, the evaluation harness, the security review, the prompt and tool architecture, the human-in-the-loop process, and the ongoing tuning. Vendors who do this for a living amortise that cost across hundreds of customers. An internal team starts from zero.
This does not mean nothing should be built in-house. It means the bar for "we will build this ourselves" has moved. If a credible partner exists for the use case, the default should be to buy, deploy quickly, capture value, and only consider building once the workflow and data flywheel are proven.
The shadow AI economy is already inside your business
The MIT survey also documents something most leadership teams underweight. Employees are not waiting.
Around 90% of knowledge workers in the survey said they use personal AI tools for work tasks at least occasionally. Only about 40% of their organisations had purchased an official enterprise AI subscription.
So a significant share of measurable AI productivity inside the average company today is happening on consumer ChatGPT or Claude accounts, on personal devices, with no governance, no audit trail, and no contribution to enterprise ROI. The work is getting done. The value is leaking outside the company.
Closing that gap, by giving employees a sanctioned, integrated tool with the right data access and the right guardrails, is one of the fastest ways to convert shadow productivity into reportable ROI. It is also one of the lowest-cost moves a leadership team can make.
What the 5% measure, that the 95% do not
When the report compares the 5% to everyone else, the differences in measurement are striking. The 5% define ROI before the pilot starts, in numbers their finance team agrees with. They tie each deployment to a single named owner with profit-and-loss accountability. They run pilots in weeks, not quarters. They kill projects fast when the metric does not move. And they treat the feedback loop, the integrations, and the data quality as the real product, not the model itself.
The 95% mostly run pilots without a defined success metric, distribute responsibility across committees, measure activity such as logins, prompts and sessions rather than outcome, and let unsuccessful pilots quietly age out instead of being shut down.
Same models. Same vendors. Same year. Wildly different financial outcomes.
So how do you actually measure AI ROI in your business?
The honest answer is that AI ROI is not measured at the AI layer. It is measured at the workflow it touches. If a contract review takes six hours today, the question is whether it takes three hours next quarter. If support tickets cost £8 each to resolve today, the question is whether they cost £5 in six months.
Pick the workflow, baseline the cost, deploy the tool, capture the feedback loop, and re-measure. That is the loop the 5% run, repeatedly, on small surfaces, until the savings compound.
If you would like a structured view of which workflows in your business are most likely to deliver measurable AI ROI inside one or two quarters, book a discovery call with Samson Redline Technologies.