Many AI proposals stall in the CFO review because nobody answered the questions before the meeting. The CFO usually isn’t against AI. The proposal was written to impress a board, not to survive a P&L. These are the nine questions that decide those conversations, and what an honest answer to each looks like.
We help businesses scope and build AI projects through our AI integration service, so weigh that. Every fact below about Anthropic, Gartner and the research was checked against their own pages and the papers themselves in October 2026.
Short answer: A CFO wants to know what the spend changes on the P&L, by when, and who answers for it. Bring a measured baseline for one process, a named owner, the full cost, a pilot with pass/fail criteria and an exit plan. The API bill is usually the smallest line: in our worked example, 10,000 short tasks a month cost $50 on Claude Sonnet 5.5 at Anthropic’s published prices. The hard part is proving value, and firm-level research finds little measurable impact so far.
What the Numbers Say
As of October 2026, from Gartner, NBER, Workday and Anthropic:
| Topic | What the source says |
|---|---|
| CFO confidence | 36% of CFOs feel confident they can drive enterprise AI impact |
| Where finance AI goes | 45% leans to productivity, 20% to decision quality |
| Firm-level impact | About 9 in 10 executives saw no AI effect on jobs or productivity over 3 years |
| Rework | Nearly 40% of AI time savings lost to fixing output |
| Claude Sonnet 5.5 | $2 input, $10 output per million tokens |
| Claude Opus 5.5 | $4 input, $20 output per million tokens |
| Claude Haiku 5.5 | $0.10 input, $0.50 output per million tokens (prompts up to 100,000 tokens) |
| Batch API | 50% off input and output tokens |
| Cache hits | 0.05x the base input price on Opus 5.5 and Sonnet 5.5 |
| Model retirement | At least 60 days’ notice for publicly released models |
Sources: Gartner CFO priorities (Dec 2025), Gartner finance AI portfolios (Jul 2026), NBER “Firm Data on AI”, Workday research, Anthropic pricing, Anthropic model deprecations.
Why AI Budget Approval Is Harder Now
From “impressive output” to “which process does this fix”
A polished demo used to be enough to win interest. Now many finance teams have funded pilots that never showed up on the P&L. The bar has moved from “can this do something impressive” to “which process does this replace or speed up, and what does that save us”.
That’s the right question. The people proposing the spend often haven’t answered it yet.
The data behind the AI ROI gap
CFOs’ doubts have a basis. In a Gartner survey of more than 200 CFOs in August 2025, only 36% felt confident they could drive enterprise AI impact, despite rising AI investment (source). A later Gartner survey of 204 finance leaders found 45% of finance AI investment leans to productivity and 20% to decision quality, which Gartner says can leave boards seeing limited strategic impact (source).
In an NBER survey of nearly 6,000 senior executives in four countries, nine in ten reported no AI impact on employment or productivity over the past three years (source). Task-level studies show bigger gains, which is why CFOs read the P&L, not the productivity dashboard.
The 9 Questions CFOs Are Actually Asking
1. What measurable outcome does this produce, and by when?
“Efficiency gains” is not an answer. A CFO wants one metric, such as cost per transaction or error rate on a defined task, with a date attached. They also want the baseline: today’s number, measured and dated, so the result can be checked.
Include the cost of doing nothing. What does the manual process cost per year? That’s the number to beat. If you can’t state the outcome in two sentences, you’re not ready to ask.
2. Is this automating a working process or a broken one?
AI on a broken workflow produces the same problems, faster. A CFO wants to know whether the process is documented, stable and measured. “We’ll fix the process as part of this” sounds like two projects on one budget line.
Fix the process first. Then automate it.
3. What does the vendor’s track record look like outside demo conditions?
A demo shows what a tool does once, with chosen inputs. It doesn’t show delivery. CFOs want live systems that people use, not proofs of concept. Ask for references from clients running the system in production, and how long it has been live.
When people ask about our agency track record, we point them to live work, not wireframes. Ask any vendor for the same.
4. Who owns this when something goes wrong?
AI systems get things wrong, and output drifts when the data feeding them changes. The CFO wants to know who holds the contract, what the support terms say, who the internal owner is and what happens when quality drops. “The vendor handles it” is not an answer. A named contact, the support clause and an internal owner who can pause the system is.
5. What’s the data quality and readiness score right now?
AI output is only as good as its data. Records are often incomplete, inconsistent or spread across systems. A CFO will ask what data the system needs, what yours looks like and what closing the gap costs. That cost is often missing from vendor proposals.
Budget for the data work as its own line. Skip it and the output becomes unreliable, the team stops trusting it and the pilot fades out.
6. Does this integrate into existing workflows or create new manual steps?
Some AI tools work and still slow the team down. Output that someone reformats by hand adds steps that eat the gain. Workday’s research found nearly 40% of AI time savings are lost to rework: correcting errors, rewriting content and checking output (source). CFOs care about net time saved.
Map the whole workflow before the meeting, including where people still check the output. If you can’t map it, the integration isn’t ready.
7. What are the governance and security guardrails?
If the system touches customer data, financial records or anything under GDPR or CCPA, the CFO needs the governance plan before approval. Who can use it? Is output reviewed before customers see it? Is the data used to train anyone’s model?
Check the vendor’s own documents. Anthropic says it doesn’t use inputs or outputs from its commercial products, including the API, to train models by default (source). It also says a current model ID is a fixed snapshot, updates ship under new IDs, and infrastructure changes can still cause minor behaviour differences (source). Pin the model ID, test before switching and keep monitoring.
8. Can this scale beyond one department, or is it a point solution?
A CFO will ask whether the spend serves one team or becomes shared infrastructure. Neither is wrong, but the framing must be accurate. Overselling scale is a quick way to lose credibility.
A point solution with a tight case is easier to approve than a vague platform.
9. What’s the exit plan if this vendor folds or pivots?
Vendors get acquired, change markets or shut down, and models get retired. A CFO will ask what happens to your data, workflows and investment if either goes away. Cover data export, termination clauses and who inside the business can keep things running.
Anthropic notifies customers with active deployments at least 60 days before retiring a publicly released model, and publishes tentative retirement dates (source). Our post on avoiding AI vendor lock-in covers the contract side. If a vendor says “that won’t happen”, that’s your answer.
What Will the AI Itself Cost to Run?
CFOs want the spend broken down. Model usage is the easiest line to estimate, because it’s priced per token, and Anthropic bills on actual monthly usage (source).
Our own worked example, from Anthropic’s published prices: a workflow reads 10,000 emails a month and returns a category and summary for each, at 1,500 input and 200 output tokens per email. That’s 15 million input and 2 million output tokens a month.
| Model (our example) | Monthly cost |
|---|---|
| Claude Opus 5.5 | $100 |
| Claude Sonnet 5.5 | $50 |
| Sonnet 5.5, Batch API | $25 |
| Claude Haiku 5.5 | $2.50 |
Prompt caching cuts repeated instructions to 5% of the input price on Sonnet 5.5, plus a write charge each time the cache is stored (source).
Two cautions. Anthropic’s rough guide is about 4 characters per token, but newer models use a tokenizer that produces about 30% more tokens for the same text (source), so count tokens on your own samples. And usage is rarely the big number. The build, data work, review time and maintenance are, as our guide to the full cost of an AI integration sets out.
What Gets AI Proposals Rejected
The “headcount avoidance” inflation problem
Headcount avoidance values AI savings as if the tool replaced a salary. CFOs discount it. The honest number is recovered capacity:
- hours saved per week, measured
- times working weeks in the year
- times the person’s loaded hourly cost
- minus time spent checking and fixing output
Workday found that many organisations fill saved time with more workload (32%) (source). Unless freed hours go somewhere the P&L can see, the CFO will call them capacity, not savings.
Pilot projects with no rollout plan
“Let’s run a pilot” can mean “spend now, find the case later”. Pilots get approved when the rollout criteria are set at the start: which metric, over what period, at what threshold. If the proposal says “pilot, then reassess”, expect to be asked what “reassess” means.
Define success criteria before asking for pilot money. Our post on running an AI pilot walks through it.
Vendors who can’t answer questions 3 and 4
If a vendor can’t show live deployments, or can’t say who is accountable when the system fails, the proposal isn’t ready. That’s the minimum bar. A CFO who watched one rollout fail on “the vendor will handle it” won’t fund a second on the same terms.
How to Build an AI Proposal That Survives CFO Review
Lead with income statement impact, not productivity claims
Frame the case as revenue protected, cost reduced or margin improved, not hours saved. If you don’t translate the claim into a line item, the CFO will, more conservatively. Gartner advises CFOs to measure AI by enterprise impact rather than pilot counts or hours saved, with metrics tied to enterprise objectives (source).
Document assumptions, CFOs will stress-test them
Name every assumption, say where it came from and test how sensitive the result is to it. What if adoption is half what you planned? What if the data work takes twice as long? Do this in writing first, so the meeting starts from your analysis.
Propose a phased approval: pilot budget vs. scale budget
The fastest route to a yes is a smaller yes first. A capped pilot with clear success criteria and an agreed rollout threshold is an easier decision than full deployment. Show both numbers, but ask for approval on the pilot alone.
Related Posts
- Realistic AI productivity gains: task studies versus firm-level results
- The business case for a custom AI tool: what to put in front of leadership
- How to run an AI pilot: testing before a full build
- AI vendor lock-in risks: keeping your exit options open
Frequently Asked Questions
What financial metrics do CFOs prioritize when evaluating AI investments?
Income statement impact: cost reduced, revenue protected and margin improved. Productivity claims count only once translated into money. Expect questions on the baseline, payback and the full cost, not just the licence or API fee.
How do CFOs differentiate between an AI pilot and a full deployment approval?
A pilot is time-boxed and budget-capped, with success criteria and a rollout decision agreed up front. A deployment approval funds production infrastructure, ongoing usage, support and change management. CFOs treat them as separate decisions.
What’s the biggest reason AI budget proposals get rejected?
Vague value: headcount avoidance, undefined productivity gains or outputs that can’t be tied to a line item. Weak vendor due diligence is another, especially no evidence of delivery outside a demo.
How should a business case for AI be structured for a CFO?
Start with the problem, the process and what it costs today, measured and dated. Model the after-state with conservative, named assumptions. Include pilot cost, rollout cost, data readiness, governance, the exit plan and references. Put the payback calculation on page one.
What do CFOs mean by “headcount avoidance” as an AI ROI metric?
Valuing AI savings as if they removed a full-time role. The real value is recovered capacity, not a removed salary line. CFOs discount these claims because the role rarely disappears and the time gets absorbed by other work.
How should SMBs approach AI vendor due diligence before taking a proposal to a CFO?
Check production references, support terms, data export and termination clauses, data-use terms and a named accountable contact. For the model, read the provider’s own pricing, data-use and retirement pages. CFOs respect proposals that show this homework.