The Conversation Is Coming Whether You Are Ready or Not
At some point in the near future, your CFO is going to look at the AI line item — which is growing, and which is probably not clearly attributed to anything — and ask what it is producing. This is not a hostile question. It is a governance question that every finance leader responsible for a growing budget category is required to ask. The engineering leader who cannot answer it is not in a performance problem. They are in a communication and instrumentation problem that, left unresolved, tends to become a budget problem.
The standard advice for justifying AI spend to a CFO is to show ROI. This advice is technically correct and practically useless for most engineering and AI leaders in 2025. Clean ROI numbers on AI spend require clean attribution of AI outcomes, which in turn requires measurement infrastructure that most organizations are still building. The gap between "we know this is working" and "we can prove what this is worth" is real, and it is not a gap you can close by working harder on the slide deck.
This post is not about manufacturing ROI numbers you do not have. It is about how to have the CFO conversation honestly, productively, and in a way that advances rather than damages the relationship — even when your attribution is incomplete and your measurement infrastructure is still under construction.
Understand What the CFO Actually Needs
Before building the conversation, you need to understand what the finance leader is actually trying to accomplish. CFOs asking about AI spend are not usually trying to cut it. They are trying to govern it. The concern is not typically that AI costs too much; the concern is that the organization does not have the management infrastructure to ensure that what it is spending is being spent well. These are different problems with different solutions.
A CFO who believes the organization is spending AI dollars with appropriate controls, clear attribution, and some signal about outcome — even if that signal is partial — is generally satisfied. What creates concern is the sense that no one knows where the money is going, that there are no controls on growth, and that accountability is diffuse. Your job in the CFO conversation is not to prove that every dollar is well-spent. It is to demonstrate that the organization has the management discipline to know what it is spending, to improve its accountability over time, and to escalate before costs get out of hand.
The Data You Need Before the Conversation
You cannot have a credible CFO conversation without specific data. The following is the minimum viable dataset for that conversation. If you do not have some of these, acknowledge the gap explicitly — do not present partial data as complete data.
Total Spend and Trend
What is the all-in AI spend for the past four quarters, broken down by quarter? This includes API costs (OpenAI, Anthropic, Google, and others), seat licenses (Copilot, Cursor, other tooling), infrastructure costs that are AI-specific (GPU compute, vector database hosting), and any AI-specific third-party services. Many organizations discover during this exercise that their actual AI spend is significantly higher than their API costs alone, because the seat license and infrastructure components are often budgeted in different cost centers.
Attribution Breakdown
Of your total AI spend, how much can you attribute to specific teams, products, or use cases? Present the attributed portion with team-level detail. Present the unattributed portion explicitly as unattributed, with a brief explanation of why attribution is incomplete and a timeline for closing the gap. Unattributed spend that is acknowledged and on a remediation path is manageable. Unattributed spend that is presented as attributed is a liability.
The Top Use Cases and Their Status
For your top three to five AI initiatives by spend, provide a one-paragraph description of what each initiative is, what it is intended to accomplish, and what signals you have about whether it is working. If you have hard outcome data, include it. If you have leading indicators (adoption rates, user retention, error rates), include those. If you have only qualitative signal, say so and describe what measurement you are building.
Controls and Governance
What mechanisms exist to prevent AI costs from growing unexpectedly? Budget controls, approval workflows for new AI integrations, automated alerts on spend anomalies — any of these are meaningful to a CFO. The absence of controls is not disqualifying if you are building them, but you need to be specific about what is being built and when it will be operational.
The Framing That Works
The CFO conversation about AI spend is most productive when framed as a maturity conversation rather than a defense conversation. The distinction matters. A defense conversation positions the engineering leader as someone who needs to justify past decisions. A maturity conversation positions them as someone who is building the governance infrastructure that will make future decisions more defensible.
The maturity framing works because it is accurate. AI cost governance in most enterprises is genuinely in an early stage of development. The tooling, the instrumentation patterns, and the organizational practices are all still evolving. Finance leaders who are operating in good faith understand this. What they need to see is that the engineering organization is treating governance as a serious investment, not an afterthought.
The specific framing: "We are at Stage [X] of AI cost governance maturity. Here is what we know, here is what we do not yet know, here is what we are building, and here is when we expect to have [specific capability] operational." This framing is honest, demonstrates self-awareness, and converts a potentially adversarial conversation into a collaborative one about resource allocation for governance infrastructure.
The ROI Question: What to Say When You Cannot Give a Clean Number
The ROI question is coming. Here is a framework for answering it when your data is incomplete.
| Situation | Appropriate Response | What Not to Say |
|---|---|---|
| You have measured ROI on some use cases | Present measured cases with methodology. Acknowledge unmeasured cases. Give timeline for measurement. | Extrapolate measured ROI to cover unmeasured spend. |
| You have strong leading indicators but no outcome data | Present the leading indicators explicitly as leading indicators, not as ROI. Describe the causal theory. Quantify what outcome measurement is being built. | Present adoption metrics as ROI proxies without labeling them as proxies. |
| You have only qualitative signal | Name the qualitative signal, attribute it to source (team feedback, user satisfaction), and describe the measurement framework being built to convert it to quantitative data. | Describe qualitative signal in language that implies quantification you do not have. |
| You genuinely have no signal on a use case | Say so. Explain why the investment was made (strategic, exploratory, infrastructure). Describe what signal would indicate success and when you expect to have it. | Omit the use case from the ROI discussion, leaving a gap in the attributed spend. |
The Questions That Will Come and How to Handle Them
"Why Is This Growing So Fast?"
AI spend tends to grow faster than other technology spend categories for structural reasons: model prices change, usage expands as adoption increases, and new use cases generate new costs that were not in the original budget. A credible answer to this question acknowledges these drivers directly and connects growth to specific initiatives rather than presenting it as an unexplained trend. If you cannot attribute growth to specific initiatives, that is itself a data point about attribution maturity that belongs in the governance section of your conversation.
"How Do We Know We Are Not Duplicating Across Teams?"
This question is often not rhetorical — many organizations genuinely are running duplicative AI initiatives at the team level because there is no central inventory of what is in production. If duplication is a real risk, acknowledge it and describe the audit or inventory process you are running. If you have already done this work and can show it, that is a strong demonstration of governance competence.
"What Happens If We Cut This Budget in Half?"
Do not answer this question hypothetically without real data. The honest answer describes which specific initiatives would be affected, what the operational impact would be on each, and which ones you would deprioritize based on current ROI signal. An answer that treats AI spend as a monolithic budget that can be proportionally cut demonstrates that you have not thought through the allocation — which is a worse signal than having a deliberate prioritization framework, even if that framework is imperfect.
Building the Ongoing Relationship
The CFO conversation about AI spend is not a one-time event. It is a recurring relationship that reflects the organization's evolving AI governance maturity. Engineering leaders who treat each conversation as a standalone pitch — arriving with a new set of numbers to defend — tend to have harder conversations over time as the numbers grow. Engineering leaders who treat the CFO relationship as a governance partnership — providing regular, consistent, honest updates on both spend and governance progress — tend to build trust that gives them more operational latitude.
This means communicating proactively when spend increases, not just when it is time for a budget review. It means flagging governance gaps before they become incidents. It means treating the attribution work as a shared investment in organizational capability rather than a compliance exercise. The CFO who feels like a partner in AI spend governance is a fundamentally different stakeholder from the CFO who feels like an auditor.
Platforms like Oberhahn exist specifically to make this relationship more productive — by providing the attribution data, trend visibility, and governance documentation that converts the CFO conversation from an adversarial numbers negotiation into a mature enterprise technology governance discussion. The quality of that conversation is a direct function of the quality of the underlying instrumentation.
The Uncomfortable Truth About Attribution Debt
Every quarter you operate without full AI cost attribution, you accumulate governance debt. The spend continues to grow. The attribution gap grows with it. The CFO conversation gets harder, not easier, because the denominator of unexplained spend increases. The organizations that invest early in instrumentation and attribution infrastructure — even when the numbers are still incomplete — are building toward a conversation that gets easier over time. The organizations that defer that investment are building toward a conversation that becomes harder to have honestly.
The CFO is not your adversary in this. They are asking the same question your board will ask, your auditors will ask, and your executive team should be asking. The answer to that question is built in your measurement infrastructure, not in your slide deck. Build the infrastructure. Then have the conversation with the data it produces.