Two surveys of enterprise AI, published a day apart after a year of rising spend, found the same number and read it in opposite ways. Dun & Bradstreet asked 10,000 businesses across 32 countries and found 28% report broad or strong returns on their AI work, unchanged from the previous quarter. Deloitte asked 3,500 leaders and found 28% reporting revenue gains from AI. That is up from 20% in January, a shift the firm reads as the money starting to pay. So your return on investment (ROI) is stalled at 28%, or rising to 28%, and which one your board hears first decides the meeting. Both counts are honest, which is the trap: two firms can agree on the arithmetic and still disagree about whether you are winning.
The 28% measures two different things
Deloitte’s 28% counts any company that reports a revenue gain somewhere in the business. Dun & Bradstreet asks for breadth. The return has to hold across the whole business, a much harder claim about one team’s workflow. Forty-four percent of its respondents sit one step below that bar, reporting pockets of ROI on some pilots. Just 9% say most pilots have scaled with clear gains.
That is how one week produced “solidifying” and “unchanged from last quarter” from two samples that agree on most other things. The firms asked different questions, and both sets were fixed before the first call. A board shown either figure is being shown a definition, and the definition is what survives the meeting.
Timing makes that definition costly. Sixty-five percent plan to raise AI spending over the next twelve months, so the returns question now lands next to the renewal. I read the two releases side by side and had to mark where Deloitte’s question ends and Dun & Bradstreet’s begins. A stalled return and a rising one are opposite orders to a budget committee, and both arrive labelled as the state of enterprise AI.
By the numbers
- 81% against 12% — Data readiness: Where enterprise data is fully ready for AI, 81% report broad or strong returns. Where it is only partly ready, 12% do [1].
- 55% against 2% — Strong returns: On strong ROI alone the split widens. That is 55% of the fully ready, against 2% of those with little or no readiness [1].
- 40% against 23% — Acting versus accounting: Four in ten let AI systems act on decisions without human approval, at least to a moderate degree [2]. Nearly one in four (23%) say accountability is unclear after a bad outcome [2].
Where the number stops being useful
The teams getting past this argument stopped reporting a percentage and started reporting a denominator. The tools are not exotic. Seventy-one percent of Deloitte’s respondents track total AI inference and application programming interface (API) spend. Only 44% track the cost of one inference, a single model answer, which is the number that turns a bill into a price. Reading the bill is not the same as pricing the unit of work, and the gap between them is where most ROI arguments live.
The second habit is slower, because it comes before a pilot rather than after one. Only 8% of the Dun & Bradstreet sample say their enterprise data is fully ready for AI at scale. A workflow that scales before that check often returns as a second project with the same name. What I am watching is whether that 8% moves before the next quarterly read, because the readiness split is the one line that would move the 81% figure.
Neither habit makes a better survey number this quarter, and that is the honest cost. A per-unit meter means a data model change, a new pipeline event and a definition several teams must sign. A readiness check stops work instead of starting it. Both are calls by engineers who would rather be shipping, and both pay off too slowly for a quarterly survey to show it.
The catch
Both releases rest on self-reports, and people tend to answer in ways that keep the budget room calm. Dun & Bradstreet sells commercial identity data. Its chief data officer ties readiness to the graph his firm sells, so the link the survey reports is also his pitch. Deloitte’s fieldwork closed in June, so its revenue trend is an early-summer reading released in late October. The January comparison also spans two editions of one study.
The match at the centre of this piece may be luck. Two separate samples landing on the same share is ordinary, and neither release shows the firms setting out to differ. What I keep circling is that both surveys count the top half of a ratio while the bottom half sits with a small group. Until the unit cost is on the page, the ROI percentage is a report about reporting.
At a glance
The shift: Two large AI surveys, a day apart, both reported 28% and read it in opposite directions. The number: 81% of businesses with fully ready data report broad returns, against 12% with partial data. The risk: The percentage is a definition, and the unit cost underneath it is tracked by fewer than half.
Related reading
- AI Coding Seats: The 75-Point Gap — more on Mental Models & Strategy
- Your AI Incident Clock — more on Mental Models & Strategy
- Your Gateway’s 104-Minute Trap — more on Mental Models & Strategy
Sources
[1] Dun & Bradstreet, “Dun & Bradstreet Survey of 10,000 Businesses Finds AI Returns Stable for Most, but Data-Ready Companies Pull Ahead on ROI” (AI Momentum survey, Q4 2026), 28 October 2026 — https://www.prnewswire.com/news-releases/dun–bradstreet-survey-of-10-000-businesses-finds-ai-returns-stable-for-most-but-data-ready-companies-pull-ahead-on-roi-302922158.html
[2] Deloitte, “Enterprise AI Enters an ROI Reckoning: More Use, Uneven Value, Reveals Deloitte Survey” (The State of AI in the Enterprise, 2026), 29 October 2026 — https://www.deloitte.com/us/en/about/press-room/enterprise-ai-enters-an-roi-reckoning.html