The 5 Things CFOs Need to Get Right About AI Spend

In my first post on AI investment from the CFO’s vantage point, I argued that AI is following a familiar pattern — hardware first, then software, then services — and that the services phase is just beginning. In the second, I made the case that the companies who win with AI won’t be the ones cutting costs. They’ll be the ones using it to grow. 

Both posts prompted the same follow-up question, usually from a peer, usually within the first thirty seconds of a conversation: 

How do I actually budget for this? 

It’s one thing to believe AI is a generational opportunity. It’s another to translate that conviction into a capital allocation framework and a set of metrics you can defend in a board meeting. 

Most enterprises are measuring AI at the wrong altitude. They are counting tokens consumed when they should be measuring value created. That gap is about to become the most expensive accounting error in the enterprise. 

Here’s the way I’ve come to put it: every company deploying AI at scale is already running an internal token economy. They just don’t have the financial discipline around it yet. The winners over the next three years will be the ones who treat tokens like capital, not compute. 

Now, the humility. As I said in a recent conversation with John Furrier on theCUBE, this environment is changing so fast that there is a great deal I simply don’t know. I’ve been an investment banker and an investor; I love putting key variables into a model. Tokens resist that. 

What follows isn’t a finished playbook. It’s five AI-era reality checks I keep running into, and how we’re approaching them at Ensono, both as our own first customer and as an advisor to our clients. 

• You probably don’t know what you’re spending. AI spend is scattered across SaaS bundles, direct API calls, cloud bills and shadow purchases — which is why most CFOs can’t actually answer the question. 

• Token economics is not cloud FinOps with a new noun. Cost per token is the wrong altitude. The unit that matters is cost per completed outcome. 

• Forecasting works differently when the ground keeps moving. The technology changes faster than annual budget cycles, so precision is the wrong goal. 

• The ROI question requires courage. Growth investments are harder to model than cost cuts — and that bias is costing companies their best opportunities. 

• Governance is what makes speed possible. It’s typically framed as an acceleration killer. AI flips that script. 

Let’s dive into each one. 

Reality Check #1: You Probably Don’t Know What You’re Spending 

Most CFOs cannot tell you what their organization is really spending on AI, and that isn’t a failure of financial controls. AI spend is distributed, embedded and largely invisible inside existing budget lines. 

• SaaS add-ons. AI capabilities are being bundled into subscriptions you already pay for, making it hard to isolate the AI premium. Expect more of this to go the Broadcom route: bundling that doesn’t ask you to take the AI — it requires it. 

• Direct API consumption. Developer teams call model providers directly, often on departmental budgets that never roll up to a central view. 

• Hyperscaler bills. AI workloads sit inside cloud invoices that already take real effort to parse and allocate. 

• Shadow AI. The one that should worry you most. Employees using consumer accounts carry real data risk, and paid subscriptions bought outside the process never appear in any budget at all. 

You can’t manage what you can’t measure, and ungoverned AI spend accelerates faster than anyone models. As I’ve told my own team, the fix isn’t a dashboard. What I want to see is a weekly email showing spend by platform and by group, so I can hold people accountable. Visibility only matters when it lands on a named owner. 

Reality Check #2: Token Economics is Not Cloud FinOps With a New Noun 

Cloud costs are complex, but they’re ultimately predictable at the unit level. Aa virtual machine of a given size costs a known amount per hour. With tooling and discipline, you can model it. 

Token-based pricing doesn’t behave that way: cost depends on the model, input and output length, whether you’re doing retrieval, and a dozen other variables, and can vary by orders of magnitude depending on how an application is built. 

Token efficiency and operational efficiency are not the same thing. A user submits one request. Behind it, an agent may make dozens of model calls, search multiple data sources, invoke tools, retry failed steps and ask another model to validate the answer. Measuring cost per token in that world is like measuring a factory by the price of electricity. 

The Right Unit of Measurement is Cost Per Successfully Completed Business Outcome. 

Our Chief AI Officer, Jim Piazza, frames this as a stack, not a list. Each level is built from the one below it, and most organizations are measuring at the bottom while believing they’re measuring the top: 

• Token cost — what the model consumed. 

• Interaction cost — what a single exchange cost, built from however many tokens it took. 

• Workflow cost — every model call, tool, data retrieval and validation step required to finish the task, built from however many interactions it took. 

• Cost per outcome — whether the task was completed successfully, at whatever workflow cost that required. 

• Enterprise value — avoided incidents, faster resolution, better customer experience, added capacity, new revenue: the return the whole stack exists to produce. 

There’s a cheaper fix hiding underneath all of this. Most automation doesn’t need a frontier model, and a lot of it doesn’t need a large language model at all. Teams are reaching for LLMs by default because they’re top of mind right now, not because the task actually requires one. 

Forget which model you’re arguing about. Route each task to the smallest, least expensive method that can complete it reliably: deterministic logic where that’s enough, a small model where it isn’t, a frontier model only where the task genuinely demands it. Routine classification, extraction and summarization rarely need a frontier model; complex reasoning and high-risk decisions might. 

Jim’s line for this is using the most powerful model for every request is the AI equivalent of sending a limousine to deliver every package. Case in point: the San Francisco software company employee whose AI-scheduled out-of-office replies cost $10,000 a day. 

On the other hand, we’ve seen what discipline is worth. One of our internal sales analytics tools cut its cost per full analysis by roughly 80%, not by removing features but by routing each task to the right method and tightening how data was passed in. Same output quality. A fraction of the bill. 

Reality Check #3: Forecasting When the Ground Keeps Moving 

Even if you solve visibility, forecasting AI spend remains genuinely hard, and not only because of pricing complexity. 

The technology moves faster than planning cycles. Most enterprises budget annually; AI capabilities evolve monthly. The model you planned to use in Q4 may be superseded by something better and cheaper by Q3. Good news for efficiency, bad news for multi-year modeling. 

Unit costs may fall, or they may not — inference costs are dropping and open-weight models are gaining share, but newer, more capable models cost more, and agentic workloads consume far more per request than chat ever did. I’d be suspicious of anyone who claims to know which force wins. Usage is just as unpredictable, and increasingly it’s triggered by machines rather than people: one employee request can set off a chain of model calls, validation and action, so budgets, permissions and stopping conditions need to be built into the agent itself, not applied after the invoice arrives. 

My recommendation? Abandon the pretense of precision. Build conservative, base and optimistic scenarios, define the triggers that move you between them, refresh quarterly rather than annually, and carry real contingency until you have 18–24 months of actual consumption data to anchor against. 

Reality Check #4: The ROI Question (and the courage it requires) 

This is the one that generates the most anxiety in conversations with my peers. And the market is starting to reflect it. Recent research from Cognizant found that two-thirds of the Global 2000 have not yet demonstrated measurable AI productivity gains, and that one in four have paused or abandoned deployments. 

Companies are spending a lot of money on tokens and the AI stack, then stepping back to ask whether they’re getting the value. That’s a fair question, and traditional ROI frameworks answer it badly. 

Cost reduction ROI is straightforward. Automate a process that required ten FTEs, calculate the labor savings, net the technology cost. Not elegant, but quantifiable. 

Growth ROI is harder. Use AI to launch a product, enter a market, or deliver a service that couldn’t exist before, and the calculation requires assumptions about market size, adoption and competitive response. The upside is far larger. The defense in a capital allocation meeting is far weaker. 

This creates a perverse incentive. Investment committees systematically favor cost reduction projects over growth projects because they model more cleanly. A clean model wins the room even when it’s the wrong bet. I think that’s a mistake, and cloud is the proof. The enterprises with the tidiest cloud business cases were modeling data center savings. The ones that generated the most value used cloud to build new businesses. Netflix’s return on cloud was never measured in racks avoided. It was measured in subscribers. 

The CFOs who learn to model and defend growth-oriented AI investments — not just cost reduction ones — will hand their organizations a real strategic advantage. 

Here’s the framework I’m using: 

Investment Type	Primary Metric	Secondary Metrics	Time Horizon
Cost Reduction AI	Cost per outcome	FTE productivity, error rates	6–12 months
Productivity AI	Revenue per employee	Cycle time, output quality	12–18 months
Product AI	New revenue generated	Customer acquisition, retention	24–36 months
Market Expansion AI	Addressable market growth	New segments, geographic reach	36–60 months

The further down the table you go, the harder the ROI is to model and the larger the potential return. A balanced portfolio holds projects in all four rows, not just the top one. 

Which brings me to the part I feel most strongly about: you have to have the courage to accept that not every AI project will have a defined ROI. If we believe something is core to a differentiated strategy, we fund it — even without knowing precisely how to price the token cost. Demanding a clean business case for every AI initiative is a reliable way to guarantee you only ever build the unambitious ones. 

That isn’t a license to spend without discipline. It’s an argument for a portfolio: mostly investments you can measure, and a deliberate minority you fund on conviction, with a clear owner and a defined point at which you decide. 

Reality Check #5: Governance is What Makes Speed Possible 

The final check is a reframe around governance. It usually gets presented as the brake pedal. In my experience, it’s closer to the thing that lets you take the corner faster. 

When AI arrived, we didn’t open the doors. We closed off direct access to external models and built our own internal tool so associates could work with AI inside our walls, because our clients’ data — and our own — are the crown jewels. We wrote policies and brought the topic to our audit committee early, and they appreciated it, precisely because there was so much we didn’t know yet. 

I’d offer that as the principle: effective governance starts by admitting what you don’t understand and setting boundaries around it, not by pretending you have every answer. 

And critically, we weren’t saying no to AI. We required AI training for every associate, put an AI goal in every annual plan, and let people build their own agents on our platform. Curiosity is one of our core values. You can’t ask for it and then lock the door. But distributed innovation without governance becomes the next generation of shadow IT: public cloud created shadow IT, AI creates shadow AI. The operating model has to pair centralized accountability with distributed creativity. 

Two more points worth making:  

The regulatory floor is rising, and boards have noticed. AI governance is now a board-level topic, and the frameworks here build on risk management and compliance capabilities you likely already have. 

The real skills gap isn’t in finance or IT. It’s the seam between them. Roughly 46% of leaders cite skill gaps as a major barrier to AI adoption. IT sees the usage but doesn’t own the financial framework; finance owns the budget but can’t read the consumption drivers. Reading an AI bill is a new literacy, not an engineering discipline. The interpretation and the governance belong to finance, even if engineering explains what’s driving the number. 

No single executive solves this alone. The CFO, CIO, CTO, CAIO and Chief Legal Officer have to operate as one system, because AI cuts horizontally across all of them. The old advice was to stay in your lane. In this environment, staying in your lane is how you get lapped. 

Practically, here’s what I’d put in place: 

• Stand up a cross-functional AI governance group: finance, legal, IT, HR and the business units. 

• Inventory what’s already in production. You’ll likely be surprised. 

• Allocate cost at the project level, so every initiative has a named owner and a budget. 

• Set spend alerts and caps thoughtfully. Enough to stop a runaway agent, not so rigid you put your engineers in a stop-work situation. 

• Define acceptable use before you need it, not after an incident. 

• Build audit trails for AI-assisted decisions, particularly in regulated functions. 

The CFO’s Opportunity 

The CFO as chief “no” officer doesn’t work anymore. In the AI era, our job is to partner on capital allocation: measure growth investments, set governance early enough to enable bold moves and get ahead of costs before the bills land. 

At Ensono, the deciding factor was the cost of inaction. Two years ago, we hired a technology leader before we had a playbook for what an AI organization should even look like. That role is now our Chief AI Officer. The result was a service model that moved from monitor, react, repair to predict, prevent, optimize, built on ten years of our own operational data. Incidents are down as much as 40%. Resolution time is down up to 66%. And when we brought those capabilities to clients, we went 25 for 25 on “can I buy that?” That’s growth, not headcount reduction. It came from a hypothesis, a governance structure that let us move and a willingness to be wrong in public. 

Some enterprises will navigate this well; some won’t. The former group will be the ones that spent most deliberately, treating AI as a growth engine rather than an efficiency program. History says the services phase is coming, and that it will be large. History also says the value goes to the clearest strategy and best execution, not the first movers. 

For CFOs, that’s the opening. Take it with humility. And take it now. 

This is Part 3 of a multi-part series on AI investment patterns. Read Part 1 and Part 2, and watch the full conversation with theCUBE on the changing role of the CFO in the AI era. 

---

Frequently Asked Questions:

How should CFOs budget for AI?

CFOs should use scenario planning rather than relying on a single fixed AI forecast. Build conservative, base and optimistic spending scenarios, define the triggers that would move the organization between them, refresh forecasts quarterly and maintain contingency funding while actual consumption patterns emerge. 

How should companies measure the true cost of AI?

Companies should look beyond token consumption and measure cost per successfully completed business outcome. That means understanding the full cost of an AI workflow, including model calls, data retrieval, tools and validation, and connecting that cost to the business value the workflow produces. 

How can companies control AI costs as usage grows?

AI cost control starts with visibility and accountability. Organizations should identify AI spending across SaaS platforms, APIs, cloud infrastructure and employee-purchased tools, assign costs to individual projects or business groups and establish clear budget owners. Routing tasks to the least expensive method capable of completing them reliably — deterministic logic, a small model, or a frontier model, as the task demands — can also reduce costs significantly. 

How should CFOs calculate ROI on AI investments?

AI ROI should reflect the objective of the investment. Cost-reduction initiatives can be measured through cost per outcome and productivity, while growth-oriented initiatives may require measures such as new revenue, customer acquisition, retention or addressable-market growth. CFOs should build a portfolio that includes both readily measurable initiatives and strategic investments whose returns may take longer to materialize. 

What role should the CFO play in AI governance?

The CFO should connect AI spending with accountability and business value. That means working across technology, legal, HR and business teams to establish project-level budgets, spending controls, ownership and appropriate guardrails. Strong AI governance should enable responsible experimentation rather than simply restrict it.