Your AI Spend Just Hit the P&L. Can Your CFO Answer These 12 Questions?

Last month I watched a CTO ask for another $4M in AI tools.

The CFO asked one question: “What did the first $2M produce?”

The room went quiet.

Neither of them was wrong. They were both missing the same thing - a framework for governing AI spend the way you’d govern any other line item that grew 400% in 18 months.

What follows is that framework. Twelve questions every CFO should be able to answer about their company’s AI program - before the board asks them first.

The 12 questions

1. What’s the thesis, and what’s the three-year P&L curve? Cost-out, capability, or new product. If your CTO can’t draw the curve, you’re funding optionality, not a plan. Optionality has a price - pay it deliberately or not at all.
2. Where does the productivity dividend land? Agents will free 25–35% of engineering hours over the next 18 months. You can spend it three ways: ship faster (revenue acceleration), ship more (capability expansion), or ship with fewer people (margin expansion). Each has a different NPV. The cheapest path almost never has the highest - it just looks safest on a quarterly basis.
3. What’s the cost of NOT resetting comp bands? AI-leveraged engineers ship 3–5x. They know it. Your competitors know it. Rewrite the bands or model the retention hit. One of those numbers is larger than you think - build it before HR builds it for you.
4. Can your data answer a top-20 business question in under 60 seconds, with citations? Revenue by segment. Churn by cohort. Gross margin by product. If it can’t, the AI you layer on top will hallucinate confidently. Treat the data foundation as a funding gate - every initiative downstream pauses until it clears.
5. What’s the asset on your AI balance sheet that compounds? Answer: the unified, governed, queryable data layer - your “brain.” Models get cheaper and better every quarter. Your data layer gets more valuable every quarter you run it, because every query, correction, and label is institutional memory the next vendor can’t replicate. Build it where you control it.
6. What can agents decide without a human in the loop? Agents are already making decisions - pricing tweaks, refund approvals, support escalations, code merges. The question isn’t whether, it’s whether you’ve authorized which. If the policy lives in a Slack thread instead of a controlled document, your exposure is unbounded. Your auditors will ask. Your D&O carrier will ask after the first incident.
7. Where are your eval controls? No agent ships to production without an eval suite. Think SOX controls, not engineering testing - same rigor, same governance. Cheap to build. Brutal to skip. Your CAE should be in this conversation.
8. Have you force-ranked the AI portfolio by ROI? Score every candidate: (labor displaced + revenue unlocked) / build cost. Fund the top quartile. Defund the bottom half publicly so nobody wastes another quarter on a vanity agent. Most AI roadmaps are subsidies for vanity - yours shouldn’t be.
9. Is engineering cycle time actually dropping? If you’re spending real money on AI coding tools, idea-to-merged-PR should fall ~40% in 12 months. If it’s flat, you bought a tool, not a transformation. Cycle time is a leading indicator of throughput. Throughput is a leading indicator of revenue. The tool budget is the test.
10. Have you priced the AI cyber tail? Prompt injection, shadow AI usage, data exfiltration through agents, model supply chain. One incident dwarfs the entire hardening program - and most cyber insurance policies haven’t been rewritten for this class of loss. Pull your exclusions language before you assume you’re covered. The tail is yours.
11. Can you pull AI gross margin on a Monday morning? Per-feature cost attribution. Model routing for margin. Vendor concentration limits. If you can’t pull the number, you’ve already lost control of it.
12. Is adoption tied to leader comp? Active agent usage per team. Percentage of in-scope workflows migrated. Retention of AI-leveraged engineers. If you don’t tie it to comp, you don’t have ROI - you have hope. Hope doesn’t show up on the gross margin line.

What changes when this works

The CFOs winning this conversation aren’t the ones with the most AI fluency. They’re the ones who got comfortable asking dumb questions in front of their CTOs until the answers stopped sounding like abstractions.

Two patterns from the rooms where this is going well:

The CFOs in those rooms stopped approving AI budgets. They started approving AI P&Ls.

And the AI program that doesn’t survive a board slide doesn’t survive the year.

Where to start

If you’re a CFO and one of these 12 made you flinch - that’s the one for your next 1:1.

If you’re a CTO reading this thinking “my CFO would never ask that” - mine did last week. Yours is either three weeks behind, or three weeks ahead. Find out which.

What’s missing from your diligence list?