How CFOs Measure AI ROI: Three Questions

The CFOs getting burned on AI aren’t the skeptics. They’re the ones who said yes to everything. How CFOs measure AI ROI comes down to three questions.

How CFOs measure AI ROI: three questions before anything renews

You push back on AI spend and the press calls it skepticism. As if asking what changed in the business when you added a tool makes you a blocker.

That question is the job.

A tool that doesn’t move a metric, shorten a cycle, or kill work somebody used to do by hand isn’t an investment. It’s a subscription.

Ask the uncomfortable question now, or explain the seven-figure line item in two years when nobody remembers what it was supposed to do.

Before anything renews or expands, three questions:

What metric moved
What cycle got shorter
What work stopped

If the answer is “people really like it,” that’s adoption, not value.

Our costs are up, not down

A CFO asked me last week what AI had done to our cost line.

I lead AI transformation. This is the question I’m supposed to win.

So here’s what I told him: our costs are up, not down.

We’re investing more every quarter. Tools, engineering time, my own hours. Productivity is clearly higher. But headcount isn’t lower and no line item has dropped. Every gain we’ve made, the business absorbed.

His answer surprised me. Same story at his company. Engineering ships faster now, so QA became the bottleneck. They’re running AI everywhere. Nobody’s number has gone down.

Two finance leaders, comparing notes, admitting the thing the vendor decks skip.

That evening I got on a call with my tech lead and we made a Q4 commitment. Not a roadmap. A target. Costs coming down and revenue going up.

Because a roadmap is not a savings plan. A number with a date is. If nobody commits to a target, the gains keep getting absorbed and the question keeps getting more uncomfortable.

Productivity was the easy part. Turning it into dollars, on both lines, is the transformation. And that starts with a commitment, not a deck.

66% expect ROI. Only 14% report value today

According to RGP’s 2025 CFO survey, 66% of CFOs expect significant AI ROI within two years. Only 14% report meaningful value today.

I’m one of the 14%. And I almost wasn’t.

I’m tired of AI bots. Tired of superagents that are supposed to do your job for you. I’ve tried them. I’ve bought them. I’ve sat through the demos where everything works perfectly.

I get pitched AI tools constantly. AI for accounting. AI for forecasting. AI for spend management. Every demo is flawless. I’ve yet to see one that survives real operations.

So I built my own.

Claude Code is the backbone. Everything runs through it. A personal knowledge base with RAG search. A presentation maker that generates branded decks. Competitor scraping. Salesforce queries. Portfolio analysis. A skill that archives my email using AI triage. A humanizer that catches AI writing patterns before anything leaves my desk. Over 70 custom workflows plugged into one assistant that I talk to in my terminal every day.

I’m not an engineer. I’m a CFO who figured out enough to be dangerous.

But here’s the honest part. None of it is scalable. Not yet. Everything is built around me. My workflows, my prompts, my context. It works because I built it for exactly one person.

I’m genuinely more productive. In my own work, AI changed the speed of everything. But in admin, operations, the human layer? It’s moving. Just not at the speed it should.

So I understand the 86%. The bottleneck was never the tools. It’s the change.

But for CFOs who are already there? Closing the gap between your productivity and everyone else’s isn’t a side project. That’s the job now.

The number that matters is how little it takes

Grant Thornton just reported that CFOs are boosting AI spending even with the economy flashing warning lights. Most people read that as a bet on the future. I read it differently. The number that matters isn’t how much we’re spending. It’s how little it takes.

We gave every person on our team access to real AI tools. The whole bill runs a few hundred dollars a month. Not enterprise licenses. Not a seven-figure transformation program. A Claude subscription and the discipline to actually train people to use it.

Here’s what that few hundred dollars bought us.

We built an agent that scans every check that comes in and allocates it to the right opportunity automatically. Cash application used to eat hours every week. Now it mostly runs itself.

We’re building a customer service agent that pulls from our SOPs to answer questions. It’s at roughly 80% accuracy today, and we think we can push it past 90%. These aren’t easy questions, so we keep a person reviewing every response until we trust it. We’re not there yet, and I’m comfortable saying that out loud.

The trap most teams fall into is thinking AI is a chatbot you pay for. It isn’t. The real leverage is using AI to build deterministic tools, then adding a thin layer of AI on top to help you read what those tools are telling you. Think Excel on steroids, not a magic answer box.

So the cost was never the barrier. For most companies the barrier is fluency. The tools are already far more capable than what people are actually doing with them.

When the headline says CFOs are spending more on AI, don’t picture a giant budget line. Picture a few hundred dollars and a team that finally knows how to use it.

Budget off the metered rate, not the sticker

It’s tempting to judge your AI rollout by how much your people use it. More usage means adoption is working.

But usage is exactly what gets expensive.

A friend was building automation skills with Claude and kept slamming into his usage limit. The “add more usage” toggle sat right there. Looked harmless. Five dollars.

I told him to never click that button.

Here’s what almost everyone misses. The $100 and $200 plans are subsidized. Anthropic and OpenAI are selling you a loss leader to get your team hooked. The second you click that button, you are on API pricing: ten to fifteen dollars per million tokens. One power user on a heavy plan can burn close to a thousand dollars of real consumption a month. You’re paying a hundred for it.

So the budgeting trap isn’t that AI is expensive. It’s that the price you see and the price you’ll actually pay are two different numbers. And the gap only shows up once your people get good at the tool.

I run two of the $200 plans. I know exactly where that ceiling sits.

If you operate a finance function, here’s the move. Don’t budget AI off the sticker. Budget off the metered rate, because that’s what scaling really costs. And if you’re a small company, there’s good news: Team plans start around $25 a seat, so giving people room to ramp up is still cheap.

The surprise on next year’s bill won’t be some new line item. It’ll be the old one repricing itself the day adoption finally clicks.

Deliberate, not skeptical

The first thing I asked my new AI agent to do was book a grooming slot for my dog. It came back and told me, confidently, that it was done. It had booked it in 2025.

CFO Dive says skeptical CFOs will help win the AI arms race. I don’t buy the word skeptical.

The one who wins isn’t skeptical. They’re deliberate. Two different things, and the difference is the whole game.

I’m in these tools most of the day. Not hesitant about them at all. But deliberate about three things: what I let AI do, what stays deterministic, and what gets checked before it touches anything real.

So I don’t use AI to forecast. I use it to build the forecasting tool, and the tool is deterministic. Excel on steroids, coded in Python. A thin layer of AI on top helps me read what the forecast is saying. The math never depends on a model’s mood.

Our customer service agent pulls from our SOPs. It’s around 80% accurate and climbing. It answers exactly zero customers on its own. A human reads every response, and we’re nowhere near loosening that.

None of it is doubt. It’s placement. You don’t win this by doubting the tools. You win it by knowing exactly where to put them.

The success rate matters less than what happens when it can’t decide

My accounting team’s support queue dropped 90% in three weeks. People keep asking what tool we used. Wrong question.

For years, processing the lockbox was a 3-hour grind. Pull checks, eyeball each one, match to the right Salesforce opportunity. When something didn’t match cleanly, a ticket. Every weird check became a “what do I do with this” case. The queue never emptied.

Three weeks ago we launched a script. Three hours per batch became under two minutes.

That’s not the win.

The win is what we did before we launched it. We wrote down the matching logic the team already trusted. The signals. The edge cases. Boring documentation work. The kind nobody posts about.

Then the harder part. We designed what happens when the script can’t decide.

When it sees a $3,911.23 check and there’s already a $3,911.23 check applied to that opportunity, it doesn’t post. It opens a case explaining what it tried, what it saw, why it stopped. A human picks it up. Trust stays intact.

Skip that step and you don’t get a 90% drop in tickets. You get a 90% increase in trust problems.

What does your AI do when it can’t handle something? That answer matters more than the success rate.

How CFOs Measure AI ROI: Three Questions