FEI Dallas
The community for forward‑thinking finance leaders in North Texas.
Start with the problem underneath every AI business case that has crossed your desk this year. Nobody agrees on how many companies are actually using this.
A Federal Reserve note published in April put three public surveys side by side. Census Bureau data showed roughly 18 percent of firms had adopted AI as of the end of 2025. The Real Time Population Survey, which asks individuals rather than firms, put work related generative AI use at about 41 percent as of November. The Survey of Business Uncertainty, which targets senior leaders, estimated that 78 percent of the labor force works at firms that have adopted AI, with about 54 percent at firms using large language models.
None of those numbers is wrong. They measure different populations with different questions, and the spread between them is the most useful thing in the data. It also explains why the figure in the last vendor deck you saw was so much higher than the one in the deck before it.
Three things drive the spread, and each one matters when you are evaluating a claim.
The first is who gets counted. A survey of firms weights a two person shop the same as a ten thousand person company. Weight the same Census data by employment and adoption jumps from 18 percent of firms to 32 percent of the workforce. Neither figure is more honest than the other, but they answer different questions, and a vendor will always quote the higher one.
The second is what counts as use. Asking whether a business used AI in a business function in the past two weeks produces a much lower number than asking whether any employee has used a generative AI tool for work. The first is an organizational commitment. The second includes someone drafting an email.
The third is the one most likely to trip you up. In November 2025 the Census Bureau broadened its own question, moving from AI used in producing goods or services to AI used in any business function. That change breaks the time series. Any growth rate that spans late 2025 is measuring a definitional shift alongside real adoption. The Federal Reserve note flags this directly, observing that before the methodological change, adoption had grown 68 percent for the year ending in September.
Census data shows AI use in the Finance and Insurance sector running at 33.9 percent as of early May, well above the 19.8 percent national rate and second only to Information at 39.7 percent.
That figure describes companies whose business is finance. It does not describe the finance function inside companies, and the two get conflated constantly. If you run finance at a manufacturer, the 33.9 percent number tells you nothing about your peers. It tells you about banks and insurers.
The function level picture is different and more useful.
Adoption tracks company size more strongly than almost anything else. Around 37 percent of businesses with at least 250 employees reported using AI, and 32 percent of those with 100 to 249, against the 19.8 percent national figure.
The more telling detail is at the other end. Between December and May, AI use rose among firms with at least 20 employees and did not change significantly among firms with fewer than 20. Adoption is not diffusing evenly down the size curve. It is concentrating.
Among very large firms in the knowledge intensive sectors, including finance, use rates reach 50 to 60 percent, or 60 to 70 percent on an employment weighted basis. If your board is comparing you to a Fortune 500 disclosure, that is the number they have in mind, and it describes a different company than yours.
Roughly 20 to 23 percent of businesses expected to be using AI within six months. The direction is up. The pace is nothing like the one implied by most vendor material.
A Census working paper published in April, drawing on the 2026 AI supplement to the same survey, looked at deployment function by function. Among firms that have adopted AI at all, 57 percent use it in three or fewer business functions.
The functions where it lands most often are sales and marketing at 52 percent, strategy and business development at 45 percent, and IT at 41 percent. Finance does not appear in that leading group.
The task level picture is similarly narrow. Writing, document analysis, and information search lead generative AI use, and 65 percent of firms confine use to three or fewer tasks. Twenty three percent of firms report workers using AI in work related tasks, rising to 41 percent on an employment weighted basis.
So the honest description of AI in the finance function today is not that it has transformed. It is that a minority of companies have deployed AI narrowly, most often outside finance, for a small number of document and analysis tasks.
For a CFO evaluating a proposal, that is useful context. You are not late. The field is thinner than the marketing suggests.
The most useful finding for anyone trying to measure this internally is that adoption runs in both directions at once.
The Census researchers found evidence of both top down and bottom up diffusion. Worker task use sometimes occurs without formal firm level adoption, and firm level adoption sometimes occurs without worker task use. Those are two different failure modes of measurement, and most finance functions have one of them.
In the first, the company has adopted nothing formally and people are using these tools anyway. Your reported adoption rate is zero and your actual exposure is not. In the second, the company has bought licenses, announced a program, and counts itself an adopter while the people who were supposed to use it carry on as before. Your reported adoption is high and your actual usage is near nothing.
Both distort the denominator in any return calculation you run. Before benchmarking your function against a national figure, it is worth establishing which of these two situations you are in, because they call for opposite responses.
The same working paper found a robust positive correlation between commercial performance and the breadth of AI integration, measured across functional deployment, task level use, and operational investment.
Read that carefully, because the direction of causation is not established. Firms that are performing well may simply have more capacity to invest in deploying AI across more functions. The paper reports correlation. It does not claim that breadth of deployment produces the performance.
Still, the association points somewhere specific. The pattern linked to performance is breadth, meaning AI operating in more places across the business, not depth in any single showcase use case. A pilot that lives permanently in one team is not the pattern in the data.
If the business case in front of you rests on headcount reduction, the evidence is thin.
AI related employment decreases occurred in only 2 percent of firms. Two thirds of users, 66 percent, rely on AI solely to augment tasks rather than replace them.
There is a real finding underneath that, and it is more nuanced than either side of the usual argument. The paper reports that functional breadth and operational investment are positively associated with employment decreases, while worker task integration shows no significant link to headcount reduction once functional integration and investment are accounted for.
Translated into planning terms, employees using AI tools at their desks is not what reduces headcount. Deliberate deployment across functions, backed by real operational investment, is the version associated with workforce change. Those are different programs with different price tags, and only one of them is what most companies are actually doing.
This matters for how you underwrite the spend. A proposal justified by seat licenses and productivity anecdotes is describing the augmentation pattern, where the evidence says headcount does not move. If the savings line assumes otherwise, the savings line is doing work the evidence does not support.
Five questions that separate a defensible case from a hopeful one:
1. Which population does your benchmark describe? A vendor survey of self selected respondents at large enterprises is not the same as a nationally representative sample, and the gap between them runs to tens of percentage points.
2. Does the comparison span November 2025? If it relies on Census data across that date, the definitional change is inside the growth rate.
3. Is the benchmark the finance sector or the finance function? These are routinely confused and they are not close.
4. Does the case depend on headcount reduction? If so, ask what makes this deployment resemble the 2 percent rather than the 66 percent.
5. Is this a pilot in one team or deployment across functions? The performance association in the data attaches to breadth, and a single contained pilot is not that.
Nothing here argues for sitting it out. Adoption is rising, the expectation figures point higher, and large firms in knowledge intensive sectors including finance are running use rates of 50 to 60 percent, or 60 to 70 percent employment weighted.
The argument is narrower. The credible case for AI in the finance function right now rests on capacity rather than cost. More analysis from the same team, faster document work, shorter research cycles. The evidence supports that. It does not yet support a headcount story, and building one into a plan means committing to savings the data does not show.
For a function whose entire job is knowing which numbers to trust, that distinction is worth getting right before the budget conversation rather than after it.
Leave a Reply