The CFO role is being redefined in some pockets of the economy more rapidly than the inherited finance playbook is willing to acknowledge. The instruments of the function, the quarterly forecast, the variable cost model, the headcount-driven productivity plan, the capital allocation framework built around predictable returns, were all calibrated for a steady-state environment. That environment is eroding, though unevenly. Artificial intelligence is now embedded in the operating cost structure of an expanding share of businesses, in the products customers expect, and in the competitive dynamics of an increasing number of industries. The finance function is one of the last parts of the organization to be redesigned for this reality. In most companies, it has not been redesigned at all.
A recent extended interview with Krishna Rao, the CFO of Anthropic, offered a useful look inside a finance function that has been forced to redesign earlier than most. The interview is worth reading on its own terms. The more useful exercise, for finance leaders whose own companies are not building AI models but are nonetheless touched by them, is to treat it as an early observatory: a finance function operating today under conditions that may approach those of the broader corporate world in three to five years.
Anthropic is not a typical company. It carries an unusually high talent density, an unusually concentrated AI exposure in which the same input that drives the cost base also drives the product, and a growth profile that compresses cycles other firms experience over a decade into a quarter. Generalizing from one frontier lab to the rest of the finance profession would be premature. What is available today is something narrower and more useful: a set of concrete examples of how the function transforms under the pressures AI introduces, which can be read as hypotheses about what may carry over to a generalist large-cap CFO, and what may not. This article walks through five of those hypotheses. None of them require the reader to be an AI specialist. Several of them are uncomfortable, because they cut against instincts that have served finance leaders well for a generation.
The first thing to break is the forecast cycle
The quarterly forecast is the load-bearing instrument of most finance organizations. It anchors the board conversation, drives the operating plan, and serves as the baseline against which performance is measured. It assumes that the world three months from now is mostly the world today, with adjustments at the margin.
That assumption is starting to fail across a widening set of industries. AI capability is improving in step changes rather than incrementally, and customer adoption tends to follow those step changes rather than track a smooth diffusion curve. A reporting cycle that lands a forecast three months after the conditions that produced it are gone is no longer a planning tool. It is a historical document.
The replacement is not more frequent forecasting in the same format. It is a different posture toward uncertainty. The posture visible inside AI-native finance organizations, including Anthropic’s, is to plan against a range of scenarios rather than a point estimate, and to update the underlying assumptions continuously rather than on a fixed calendar. The “cone of uncertainty,” an established forecasting concept, captures the structure of the problem: the widening range of possible outcomes that opens up over twelve to twenty-four months when growth is compounding rather than linear, or when the underlying technology is changing on a cycle measured in months. Krishna Rao describes operating against such a cone, and notes that Anthropic maintains a deliberately low threshold for revising its priors. If a new data point breaks an assumption, the assumption is replaced, not defended.
The transferability question, for a generalist large-cap that is not facing Anthropic’s growth profile, is which parts of this posture carry over and which do not. The cone of uncertainty as a frontier lab experiences it, where a twelve-month horizon can contain order-of-magnitude differences in usage, is specific to that setting. But the underlying disciplines are not. Most large enterprises now operate inside at least one input that compounds on short cycles, whether that is AI capability, regulatory regime change, customer behavior in digitally mediated channels, or the cost curve of a critical supplier. The parallel running view, the explicit small set of scenarios, the willingness to retire a prior when the evidence breaks it: each of these is a practice that scales down. The hypothesis worth testing in a generalist context is that a finance function which runs both a quarterly board pack and a continuously updated operating picture will outperform one that does not, even when the underlying business is not growing at frontier-lab rates.
The practical implications for finance leaders willing to act on that hypothesis are concrete. Stop treating the quarterly view as the live operating picture. Build a parallel running view that is updated when material new information arrives, not on a fixed schedule. Replace the single base-case plan with an explicit small set of scenarios, including a high-end case that the standard process would normally discount. And accept that the forecast is no longer a contract with the future. It is a working hypothesis, and the discipline that matters is how quickly it can be revised.
This is not an argument for abandoning quarterly cycles. They serve real governance purposes and will continue to anchor reporting. The argument is that the operating reality and the board pack should no longer be the same document.
Compute as a strategic category
Most CFOs cannot say, with any precision, how much their company will spend on compute over the next three years. The number is buried inside cloud bills, model licenses, infrastructure budgets, and engineering line items, scattered across multiple cost centers. The procurement process treats it as a variable input that scales with demand. The board reporting treats it, when it appears at all, as a sub-component of IT or technology cost. None of this is adequate for what it is becoming, at least at the frontier, and may not remain adequate elsewhere.
A brief translation is in order, because much of the conversation about compute is conducted in language finance leaders are expected to follow without acknowledging the gaps. Compute, in this context, means the specialized chips and the data centers that run AI workloads, together with the electricity required to operate them. It is the physical substrate on which any AI-enabled product or process runs. Costs include hardware procurement or cloud rental, the data-center capacity to house the hardware, and the power to keep it running. Multi-year contracts are now routine. Pricing is volatile. Capacity is constrained in ways that traditional CFOs have not seen since the semiconductor cycles of the 1990s.
At the frontier of the industry, compute is now the single most consequential procurement category in the business. Public disclosures put cumulative commitments at major AI labs in the range of one hundred billion dollars and rising. Krishna Rao reports that he personally spends thirty to forty percent of his time on compute decisions. That figure sounds extreme for a CFO, but it reflects how central compute has become to the business at the frontier. In a small but growing set of companies outside the AI industry, where AI is becoming a material part of the operating cost structure, an early form of the same migration is visible: compute is starting to move from a sub-line of IT spend toward a procurement category that warrants its own treatment. Whether this migration becomes universal, and on what timeline, is one of the open questions of the next several years. The frontier-lab example is the most concrete evidence that the trajectory is real.
Three properties of compute spending make it resistant to standard finance treatment.
The first is non-fungibility of timing. You cannot buy a meaningful quantity of high-end AI compute on short notice. The lead times for data-center build-out and chip allocation are measured in years. Procurement requires committing to capacity that will land twelve, eighteen, twenty-four months out, on the basis of a demand forecast that will almost certainly be wrong. Buy too little and you cannot serve the customer demand that materializes. Buy too much and you have built a stranded asset.
The second is fungibility of use. Once compute capacity exists, the same hardware can be redirected on short notice between revenue-generating customer workloads, internal research and development, and employee productivity tools. The chip that handles a customer request in the morning can be retrained on internal data in the afternoon. This breaks the standard cost-of-goods-sold framing, in which the cost of serving a customer is a distinct, variable line. The cost is not variable in the conventional sense. It is a pooled resource that supports current revenue, future capability, and internal productivity simultaneously, with the allocation among them shifting hour by hour.
The third is opacity to traditional unit economics. The temptation in finance is to assign a per-customer cost and reason about gross margin from there. In an AI-intensive business, that number is not stable. It depends on how much of the underlying compute is being used for training versus inference, how efficient the current model generation is, and how the same capacity is being amortized across other workloads. A clean unit-economic story is achievable, but only by ignoring most of what is happening.
One portable hypothesis for finance leaders outside the AI industry is that the classification problem arrives in miniature long before it arrives at the scale of a frontier lab. The same external AI spend in a typical enterprise will support production features, employee productivity, and experimental capability work. Forcing it into the legacy COGS or operating-expense categories produces numbers that look clean but no longer describe the business.
The frontier-lab pattern is not yet the typical large-cap pattern. At a one-hundred-billion-dollar compute envelope, the three properties listed above bind hard enough to force a redesign of procurement and accounting. At a five-million-dollar enterprise AI bill, they bind softly enough to ignore for another year or two. The more interesting question, and one no large-cap CFO yet has a clean answer to, is the threshold at which they start to bind: the scale of AI spend at which the legacy classification breaks down badly enough to distort decision-making. Frontier labs are an early-warning system on that threshold. They have reached the breakdown point already, and the patterns they are developing in response are the most concrete examples available of what the response looks like. The transferable instinct is to begin tracking the underlying composition of AI spend before the scale forces the issue, so that the chart of accounts catches up to the business rather than lagging it.
A new measuring stick: return on compute
If the old categories no longer work, what replaces them? The answer beginning to take shape at the frontier is a metric called the return on compute, sometimes phrased as the return on the “full envelope” of compute spend. The idea is to evaluate the entire pool of compute capacity, across customer workloads, internal development, and employee productivity, as a single capital base, and to ask what aggregate return it is generating across the different time horizons each of those activities supports.
The metric is unfamiliar to most CFOs because it does not fit neatly into any existing slot. It is not a gross margin number, because it includes R&D-equivalent spending. It is not an R&D efficiency number, because it includes customer-facing operating cost. It is not a capital allocation number in the conventional sense, because the asset is consumed and replaced on a cycle measured in months. It is closer in spirit to a portfolio view, in which several activities with different payback profiles are evaluated against a shared capacity constraint.
The mechanics, in practice, are less exotic than the framing suggests. The components are recognizable. Direct revenue contribution from customer-facing workloads. Indirect contribution from productivity gains in internal use. Forward-looking contribution from capability development that opens up new addressable demand. What is novel is the insistence on evaluating these together rather than separately, and the recognition that the same capacity supports all of them.
For CFOs whose companies are not at the frontier, the metric is still relevant in a simpler form. The basic question is whether the aggregate AI spend, across all the categories it appears in, is generating a return that justifies its scale. Most companies have not asked this question in a structured way. They have asked it for individual deployments, often with optimistic assumptions, but not at the portfolio level. The number of finance organizations that can produce a coherent answer to “what is our total AI spend across all categories, and what return is it generating” is small. Building the capacity to answer that question is now part of the role.
Krishna Rao described his own benchmark in plain terms: the test of his procurement and allocation decisions is not whether each customer is profitable, or whether each model is cheaper than the last, but whether the full compute envelope is producing a robust return across all the time horizons it supports. That framing, stripped of its frontier-lab specifics, may transfer to any company whose AI spend has grown enough to matter.
The harder question, for a generalist large-cap that does not yet operate at frontier compute scale, is whether building the apparatus to track an aggregated return on AI capacity is worth the effort at the current spend level. The frontier-lab answer is that the discipline becomes essential when the same pool of capacity supports customer-facing workloads, internal productivity, and capability experiments, regardless of absolute scale. The hypothesis worth testing is that this is a leading indicator rather than a frontier-only phenomenon: that any company whose AI spend has reached the point of supporting more than one of those three uses simultaneously has already entered the territory where the metric clarifies decision-making, even when the dollar figures are two orders of magnitude smaller. The cost of building the view is modest. The cost of building it after the spend has scaled past the point of clean attribution is higher.
Pricing dynamics: rising consumption, falling unit prices
A second area in which traditional CFO instincts mislead is pricing. The consensus expectation a year ago was that AI prices would collapse, that intelligence would commoditize, and that finance leaders would be able to ride a deflationary curve down. The actual pattern has been more complicated, and it produces a counterintuitive consequence that any CFO whose business is touched by AI needs to understand.
One observation Krishna Rao described from inside Anthropic, and which carries an old economic name, is that when the unit price of AI capability falls, total consumption rises by more than the price reduction alone would predict. As a generalized pattern this is a single-firm data point, not yet a documented industry behavior. The underlying mechanism, however, is well established. It is the Jevons paradox, originally observed in the nineteenth century in coal markets: when efficiency improvements made coal-fired engines cheaper to run, total coal consumption rose rather than fell, because the cheaper price made coal economical for applications that had previously been out of reach. The hypothesis carried over from Anthropic’s experience is that the same dynamic now operates on AI capability. Each price reduction, or each efficiency improvement, opens up use cases that were previously too expensive to attempt. Customers find more places to apply the technology. Total spend rises.
This has two practical implications for finance leaders.
For companies selling AI-enabled services, the standard instinct that lower prices erode margin can be wrong. If the reduction unlocks a step change in usage, total revenue and total contribution can rise. The relevant pricing question is not what level maximizes margin per unit, but what level maximizes the conversion of capability into adoption.
For companies buying AI capability, the inverse holds. The natural finance response to a vendor price cut is to bank the savings. The more productive response is to ask which previously uneconomical use cases now make sense, and to redeploy the freed budget toward those rather than letting it flow to the bottom line. The companies that capture the most value from the next several years of AI deflation are likely to be the ones that translate every price reduction into expanded adoption, not the ones with the lowest AI bill.
This is uncomfortable territory for traditional finance discipline, which is calibrated to find savings and convert them to margin. The discipline still applies, but the optimization function is different. The CFO is no longer minimizing the cost of a stable input. The CFO is managing the conversion ratio between falling prices and expanding use.
Of the patterns in this article, the buyer-side discipline is probably the most directly transferable from frontier labs to a generalist large-cap. The dynamic does not depend on running a frontier-scale AI operation. It depends only on being a buyer of AI capability in a market where vendor prices are falling. Every CFO who has approved an AI procurement in the last twelve months has been a participant in this market, often without recognizing the dynamic. The seller-side mirror, in which falling prices unlock total revenue rather than eroding margin, is more specific to companies whose product carries AI as a substantial input. For the typical large-cap, the practical work is on the buying side: monitoring price reductions across the AI stack and treating each one as a budget-reallocation question rather than a savings line. A simple operational rule, observable inside several frontier labs, is to default to redeployment unless an explicit case is made for taking the savings.
The finance function as a user, not just a buyer
A separate question, often raised too late in the planning process, is what AI does inside the finance function itself. Most CFOs have engaged with AI primarily as a procurement and capability question for the rest of the company. The harder and more immediate work is inside their own organization.
Inside Anthropic, Krishna Rao describes routine financial-reporting workflows, monthly financial reviews, statutory financial statements across legal entities, and management reporting, as now in a state where ninety percent or more of the preparation is produced by AI systems before any human review. The reported time compression for those tasks is from hours to roughly thirty minutes. How far this state has spread beyond a small number of frontier labs and forward-leaning enterprises is unclear, and most large-cap finance organizations are some distance from it today. What the frontier-lab example establishes is the capability ceiling rather than the current floor: it shows what is achievable today by an organization willing to invest in the workflow redesign. Whether a given large-cap reaches that ceiling in three years or seven is a function of investment, willingness, and the nature of the work, not of whether the underlying technology is ready. The work that survives the transition is judgment work: deciding what the numbers mean, what to do about them, and how to communicate them. The work that disappears is the preparation, reconciliation, and formatting that historically absorbed the majority of finance staff time.
Two patterns from current practice cut against common assumptions and deserve attention.
The first is that the people consuming the most AI capacity inside the function, measured directly by token usage, are typically not junior staff. They are senior specialists whose work involves stable, repeatable domain logic that can be encoded into reusable workflows. Krishna Rao, who maintains a token-usage dashboard for his team, has reported his head of tax as the single largest token consumer, building policy engines and automating compliance workloads. This pattern, if it generalizes, runs counter to the assumption that AI helps mostly at the entry level. The functions with the most to gain from internal AI deployment may be the ones where senior expertise is repeatedly applied to similar problems. For CFOs deciding where to direct first investments, this is at minimum a corrective worth taking seriously.
The second pattern needs to be stated directly, because it cuts against the assumption many CFOs are quietly making in their AI business cases. Introducing AI into the finance function does not necessarily lead to headcount reduction. The standard finance instinct is to take a productivity improvement and convert it into a smaller team. The dominant pattern observed at organizations adopting these tools at scale has been the opposite: finance teams have grown rather than shrunk. The binding constraint on output shifts from staff capacity to the ambition of the work, and the team grows to handle the expanded scope. Analyses that were previously uneconomical become viable. Reporting cycles get compressed and the freed time is redirected into deeper interpretation rather than into a headcount line on the savings ledger.
That said, this is the place where the talent-density caveat applies most directly. The pattern of finance teams growing rather than shrinking as AI is adopted has been observed primarily inside high-growth, high-talent-density organizations where the binding constraint on output was always ambition rather than cost. Whether the same pattern carries over to a mature large-cap finance organization, where headcount discipline is a longstanding management orientation rather than a growth-stage afterthought, is genuinely uncertain. The transferable observation is narrower: that converting AI productivity gains into immediate headcount reduction may be the wrong optimization, because it forecloses the option to redeploy the freed time into deeper analytical work that the function has historically been unable to staff. Whether the team ultimately grows, holds, or shrinks is a separate decision, and one that depends on the company’s growth profile and management philosophy. The transferable discipline is not to treat the first two questions as the same.
The practical implication, for finance leaders unwilling to wait for that question to settle, is straightforward. The capability is adequate today for the bulk of routine reporting and compliance work. The blockers are procedural and cultural, not technical. The firms whose finance organizations will look most different in three years are likely to be the ones treating this as current operating work rather than as a future project.
Talent density may now beat talent mass
A final theme worth surfacing, partly because it crosses into territory CFOs share with their heads of HR, is the implicit trade-off between team size and team quality, and the way that trade-off may be shifting in environments where current-generation AI tools are widely available.
The historical finance discipline has tilted toward team size. Scale of operation justifies scale of staffing. Productivity is measured at the team level, and headcount growth is treated as the principal lever for output expansion. This assumption is being challenged by a different model, sometimes phrased as talent density over talent mass: a small number of very high-caliber people, equipped with current-generation tools, produces results that a larger team of average performers cannot match.
The argument is most visible inside frontier research organizations, where the productivity gap between elite and average performers has always been large. It may be becoming more general. As AI tooling raises the output of the most capable users by multiples, while raising the output of average users by smaller increments, the gap widens. In an environment where that pattern holds, the marginal dollar is often better spent on tools that amplify existing high-performers than on incremental hires that add coordination overhead.
This cuts against most of the inherited finance and HR norms. It is harder to justify in budget conversations than headcount growth is, because the connection between tool spend and output is less direct than the connection between hires and output. It also requires a level of confidence in identifying high-caliber talent that most organizations have not built. None of which makes it wrong. The CFOs who navigate this transition will need to make capital allocation decisions in which the trade-off between people and tools is explicit, rather than buried inside the headcount plan.
The talent-density argument is the one in this article that travels least cleanly from the frontier to the typical large-cap. Inside a frontier research organization, the productivity gap between elite and average performers has always been wide, and the cultural muscle to act on that gap is already built. Inside a typical large-cap finance organization, both conditions are weaker. The compensation structure is flatter, the cultural premium on identifying and concentrating capital around the highest performers is lower, and the incremental productivity from current-generation tools may be more uniform across performance levels in roles where domain logic is more transferable. The frontier-lab observation does not refute the case for talent density at a typical large-cap. It does suggest the case is narrower, and that the most defensible version is at the margin: a budget choice between an incremental hire on an average team and an incremental tool investment on the highest-leverage individuals, where the latter is now harder to dismiss than it was three years ago. That decision, made repeatedly over a multi-year capital plan, compounds.
When exploration becomes cheap
The five hypotheses so far have each been about a specific instrument or trade-off: the forecast cycle, the compute line, return on capacity, pricing posture, internal AI use, talent density. A more general effect threads through them, and is worth surfacing on its own terms because it bears on the competitive dynamics of any sector exposed to differential AI adoption.
The most immediate operational gains from AI tooling inside the finance function are temporal. The time to produce a monthly close, an additional scenario, a variance walk, or a deal model has compressed. The frontier-lab examples cited so far suggest order-of-magnitude reductions on specific workflows. Two compositional effects follow.
The first effect is external. In sectors where AI adoption has been faster, the operational cycle has compressed in step. A competitor that took six weeks to respond to a market move now takes ten days. A finance function running at the old cycle time, in an industry where peers are running at the new one, is supplying decisions to the executive team on a delay that did not previously register as a disadvantage. Whether a given industry has reached this inflection point is an empirical question that finance leaders are now in a position to ask, and one that is increasingly worth asking explicitly rather than treating as background.
The second effect is internal, and is the one most likely to be underweighted today. The historical finance discipline has been organized around triage. Every analysis costs time, every additional scenario costs more time, and the portfolio of questions a function could explore was bounded by the staff and the calendar it had available. Anyone who has tried to commission an additional scenario in the last week of a planning cycle has felt that constraint directly. As the cost of running an additional analysis falls by an order of magnitude, which is what the frontier-lab examples suggest is now achievable, the binding constraint shifts. It is no longer how much can be analyzed, but which questions are worth asking. For the first time in a generation, the finance function may be in a position to genuinely explore rather than triage. That is a structural change in the role’s bandwidth, and most finance organizations have not yet absorbed its implications.
The transferability question is the rate at which the cycle compression spreads from the frontier outward. Inside frontier labs, the effect is visible at full strength. In a typical large-cap, it is partial and uneven, and it tends to land first in the parts of the finance function that have already been digitized. The transferable observation is that the value of the gain is not the time saved on any individual workflow, but the change in the function’s effective scope of inquiry. The CFO who treats AI tooling as a speedup to existing tasks captures the smaller share of the benefit. The CFO who treats it as an expansion of the question set captures the larger share. The two stances look similar on a one-quarter view. On a three-year view, they diverge sharply.
The strategic shift the role may be making
The strategic role of the finance function may be about to take its next step forward, and frontier labs are the place where that step has already been forced. The CFO has historically been the steward of the past: closing the books, reporting performance, defending the variance against plan. That work is not going away, but in the settings where AI tooling has been adopted earliest, it is being compressed and increasingly automated. The capacity that is being freed up is not destined for the cost line. It is destined, in those organizations, for the part of the role that has always mattered most and has always been hardest to make time for: deciding where the next dollar of capacity, capital, and attention should go.
Return on compute is one of the metrics that anchors that work at the frontier, and it is the metric most likely to extend outward from there. It belongs alongside return on invested capital, return on R&D, and the conventional unit economics of the business, as one of the principal lenses through which AI-related spending is evaluated. The CFO who can produce a coherent return-on-compute view, even at modest scale, across customer-facing workloads, internal productivity, and capability development, will be operating with an instrument that most peers do not yet have.
The broader shift, if it generalizes from frontier-lab finance functions to the rest of the profession, is that the finance function may be in a position, for the first time in a generation, to shed much of the technical minutiae that have consumed most of its bandwidth, and to redirect that bandwidth toward the questions that drive strategic growth and tactical resource allocation. The reconciliations, the formatting, the manual close, the variance walkthroughs: these are being absorbed by tooling that is already adequate today and improving rapidly. What remains, and what expands once the cost of exploring an alternative falls far enough to make exploration routine, is the work the title was always meant to describe: allocating capacity under uncertainty, classifying spending categories the legacy chart of accounts cannot describe, judging where the marginal dollar produces the most leverage, and acting as the principal counterparty to the rest of the executive team on every meaningful resource decision.
The finance organizations that complete this shift, if the frontier-lab pattern carries over, will look different in three years. They will have shorter reporting cycles and longer analytical horizons. They will track compute alongside capital expenditure as a first-class category. They will manage pricing dynamics in both directions and treat their own technology stack as a working test case. They will spend less time describing what happened and more time deciding what to do next. None of this requires being at the AI frontier. It does require the willingness to retire the parts of the playbook that are no longer earning their place, and to put something better in their stead. What exactly that better playbook will look like in detail, in a generalist large-cap setting rather than a frontier-lab one, none of us yet know. That is part of the work, and one reason it is worth watching the frontier closely while it is still ahead of the rest of the field.