We measure AI productivity as subtraction
Ask how AI improves productivity and you will almost always get an answer framed as subtraction. More automation means fewer hours, which means fewer people required. AI productivity, in this telling, is an efficiency equation, and the output of the equation is a smaller organization. The value shows up as something removed.
That framing is not wrong, exactly. Efficiency is real and it matters. But it may be an incomplete way to measure the opportunity, and measuring the wrong thing leads to the wrong conclusions about what AI is actually for.
Some recent research offers a useful reason to question the subtraction frame. Ramp's Economics Lab, working with the labor-analytics firm Revelio Labs, linked observed AI spending from corporate card and bill-pay data to workforce records for 21,559 U.S. firms, from January 2021 through February 2026. If the subtraction story were the whole story, you would expect the heaviest AI adopters to be shedding headcount. That is not what the data shows.
Which raises a question worth taking seriously. What if the most important result of AI productivity isn't reducing the size of the organization, but increasing what the organization is capable of doing?
What the Ramp research actually found
The headline finding cuts against the reflexive assumption. Firms making the largest AI investments grew employment by roughly 10% in the two years following adoption, and entry-level headcount rose about 12% among those high-intensity adopters. Low-intensity adopters, by contrast, showed no statistically significant workforce change. The gains emerged gradually rather than immediately, and appeared across engineering, sales, administration, and customer service rather than in one corner of the business.
The gap between the two groups is stark in spending terms. High-intensity adopters spent roughly $33 per employee per month on AI in their first three months. Low-intensity adopters spent around $3. So the firms that invested materially more in AI were also the firms adding people, while the firms that dabbled saw no measurable change either way.
Now the essential caveat, because it determines what you are allowed to conclude. This is a correlation, not a causal finding, and the study's own authors are careful to say so. AI did not necessarily cause these companies to hire. The high-intensity adopters differed from other firms in ways that independently predict growth: they were already larger, more technically oriented, faster-growing, and more likely to be venture-backed before they ramped up AI spending. The strong effects were also concentrated in information-sector businesses, not spread evenly across the economy. So the honest reading is not "AI creates jobs." It is narrower and more interesting than that.
What the study actually does is puncture a simplistic assumption. It shows that deeper AI adoption does not automatically produce a smaller workforce. The companies going furthest with AI are not, as a group, shrinking. That does not prove AI expands organizations. It proves the subtraction story is not the only thing that happens, and often is not what happens at all. That alone is worth sitting with, because so much AI strategy is built on the assumption that it is.
Efficiency and capacity are not the same thing
To make sense of this, it helps to separate two ideas that get collapsed together whenever people talk about productivity.
Efficiency means doing today's workload with fewer resources or less time. Same output, lower cost. It is a subtraction: you take the existing work and shrink the inputs required to produce it.
Capacity means increasing the amount, frequency, depth, or quality of work the organization can perform. Same organization, more output. It is an addition: you take the freed-up ability and point it at work you could not do before.
Efficiency is genuinely valuable, and nothing here argues against it. But the crucial point is that efficiency creates a choice. When AI makes an existing process take less effort, that freed-up capacity does not have a predetermined destination. An organization can remove it, taking the savings as cost reduction and a smaller team. Or it can reinvest it, keeping the capacity and directing it toward additional capabilities and growth. Both are legitimate. Neither is universally correct, and the right choice depends on the organization, its market, and its ambitions.
But here is what the two-choice framing reveals: measuring AI productivity purely by hours eliminated only captures organizations that chose subtraction. It is blind to the ones that chose to reinvest, because their gains show up as more and better work rather than as a smaller headcount. If the Ramp data is any indication, a lot of the heaviest adopters are making the reinvestment choice, which is exactly why a subtraction-only measure misses them.
Apply this to Finance
Finance is an especially clarifying place to think this through, because most finance teams operate under real capacity constraints. There is more the function could do than there is time to do it.
Consider where a finance team's hours actually go. Collecting data from across systems. Reconciling those systems against each other. Preparing management reporting. Updating forecasts. Building the same recurring analyses each cycle. Investigating variances. Assembling executive reporting. Much of this is necessary, and much of it is manual, repetitive, and time-consuming. It is the work that fills the month and leaves little room for anything else.
So when AI and finance automation reduce the effort those activities require, the interesting question is not "how many finance roles can we now remove." For most teams, that is the wrong question, because the team was never overstaffed relative to what the business actually needs from Finance. The interesting question is: what can Finance do with the capacity it gets back?
The list of answers is long, and it is all work most finance teams already wish they had time for. More frequent forecasting instead of a quarterly scramble. Faster management reporting that reaches leadership while it still matters. Deeper variance analysis that explains why, not just what. Better cash forecasting. More scenario planning. More timely SaaS metrics and ARR reporting. Stronger go-to-market analysis. Genuine business partnership with the teams Finance supports. More proactive executive insight. More time spent understanding why results changed and helping leadership decide what should happen next.
Notice that none of these is "the same reporting, with fewer people." Every one is Finance doing more, or doing it better, or doing it sooner. That is capacity, not efficiency. And it is almost certainly where the larger prize sits for finance organizations, because a finance function that forecasts more often and explains results more deeply is worth far more to a business than one that produces the identical monthly package slightly cheaper.
Adoption is not the variable that matters most
This connects to something we have written about before: the gap between how widely Finance has adopted AI and how little measurable value most teams report from it. A large majority of finance organizations are actively using AI. A small minority say it has delivered clear, measurable value. (We covered that gap in Finance Has Adopted AI. So Where Is the Value?.)
The Ramp research, read alongside that adoption gap, points to a broader possibility. Maybe AI adoption is not the variable that matters most. Depth of integration may matter more.
The Ramp data is at least consistent with this. It was the high-intensity adopters, the ones spending an order of magnitude more per employee, who showed real workforce effects. The dabblers showed nothing. Now, spending is not the same as integration, and the two studies measure different populations and different outcomes, so this is a suggestive parallel rather than a proof. But the shape of it rhymes. Buying an AI subscription and fundamentally changing how work gets done are different things, and they produce different results. Light adoption tends to produce light results.
That distinction, between adopting AI and integrating it, is the one finance leaders should hold onto, because it explains why so much AI spending has underdelivered.
Why AI beside a manual process changes so little
Here is the mechanism. Adding AI to an existing manual workflow can improve individual productivity without changing the organization's overall capacity at all.
Picture a finance team that still, by hand, extracts data from each system, reconciles those systems against one another, rebuilds its reporting each month, maintains a web of disconnected spreadsheets, repeats the same recurring analysis every cycle, and waits weeks for management reporting to come together. Now give that team an AI assistant. It helps write a formula faster, drafts a commentary paragraph, summarizes a document. Real help, at the level of individual tasks.
But the process is unchanged. The extraction still happens by hand. The reconciliation still happens by hand. The reporting still gets rebuilt. The spreadsheets still have to be maintained. The weeks-long wait is still weeks long. The AI made pieces of the work faster without changing the work itself, so the team's actual capacity, how much it can produce and how quickly, has barely moved. This is AI beside the process, and it is why light adoption yields light results.
Real capacity gains come from changing the process, not decorating it. That means AI working within an integrated approach to the whole flow: financial data integration across systems, management reporting and month-end reporting that do not have to be rebuilt from scratch, recurring analysis that does not consume the cycle. When the manual work shrinks because the process changed, capacity genuinely expands. When AI just sits alongside an unchanged manual process, it does not.
Capacity does not remove the need for governance
One caution, so this argument is not misread. Expanding capacity through automation does not loosen the requirements that make financial information trustworthy. If anything, it raises the stakes, because more work flowing through an automated process means more riding on that process being sound.
Finance still requires trusted data, validation, reconciliation, deterministic calculation where exactness matters, traceability, and human accountability for the results. These do not become optional because AI is doing more of the work. The core financial numbers still need to be exact and reproducible. Every figure still needs to be traceable to its source. A person still has to be able to stand behind the output. AI should operate within Finance's controls, not in place of them. The goal is a finance function that can do more and still defend everything it produces, which is only possible if governance scales along with capacity rather than being traded away for it.
The question finance leaders should actually ask
Return to where we started. The reflexive way to measure AI productivity is by subtraction: How many hours did we eliminate? How many people do we no longer need? Those questions have their place, and for some organizations in some situations, cost reduction is the right call.
But for most finance organizations, they are the wrong primary questions, because they measure only the efficiency half of the equation and miss the capacity half entirely. They cannot see the finance team that forecasts twice as often, closes faster, and finally has time to explain the business to leadership, because that team's gains do not show up as anything removed.
A better question, and the one the Ramp research nudges us toward, is this: What can our finance organization accomplish today that it could not accomplish before?
That reframes AI from a cost lever into a capability lever. It measures the technology by what it adds rather than only by what it subtracts. And it fits what Finance actually needs, which is rarely a cheaper version of the same output and almost always more capacity to analyze, anticipate, and help the business make better decisions.
That may turn out to be AI's biggest contribution to Finance. Not a smaller team producing the same reports, but a finance function capable of far more than it is today. The organizations that measure for that, rather than only for hours eliminated, are the ones most likely to actually get it.