How CFOs are Actually Using AI

👋 Hi, it’s CJ Gustafson and welcome to Mostly Metrics, my weekly newsletter where I unpack how the world’s best CFOs and business experts use metrics to make better decisions.

I realized there was a TON of material on how you can build AI into your product roadmap.

But there was so little on what we could do to improve efficiency internally.

Today I’m going to walk you through how CFOs are actually using AI — not just the stuff you see in pitch decks or on LinkedIn, but real use cases that are improving operations inside finance teams, and the broader org.

Many of you have probably felt a temperature change in the board room where the board is trying to sift through who’s best in class within their portfolios at applying AI use cases internally.

And based on the survey I’m about to share with you, 64% of respondents plan to allocate 5–10% of their 2025 opex to AI tools and talent. A sizeable chunk

And 57% expect to see moderate efficiency gains within the next 12 months

BUT

40% aren’t tracking ROI at all. Makes the first part of the equation more difficult, right?

And only 18% have gone org wide and implemented AI workflows across all departments.

So we’re clearly in the early innings, where expectations are outstretching reality.

And there’s a gap. And it’s hard for people to peg if they are ahead or behind. So let’s demystify that today

Here’s a sneak peek of the report (the ENTIRE deck + analysis is available below):

Below we’ll dig into:

Real quick. The required n equals slide.

So one of the benefits of having a lot of CFOs on my podcast is I can survey people who are far smarter and more experienced than myself.

So I surveyed 42 CFOs across venture and PE-backed companies who have come on my podcast. Companies ranged from $5 million to over $500 million in ARR.

As you can see, nearly half the respondents were in the $100M to $500M in revenue range, so running sizeable operations

We asked 18 questions about their AI usage internally. No to power the product, but to run the business better.

A bunch of CFOs are building predictive models that flag customers at risk of leaving. This lets customer success teams act early — before that QBR goes sideways.

The downfall of this application is that it often feels like a black box. Deterministic models are much more understandable than probabilistic ones that juggle 50+ variables.

(You know that if you’ve ever built a simple deterministic FP&A budget vs probabilistic one.)

What’s scary is we can simultaneously feel confident in the output, while uneasy about how the sausage is made.

To provide some color on actual churn prediction use cases, teams are feeding inputs like

Into tools like Gainsight, Vitally, or ChurnZero.

Some are even layering in macro signals — think hiring freezes, industry slumps, or headcount data from LinkedIn.

Now, here’s where the nuance comes in: it’s not just about a churn score — it’s about generating a prioritized action list.

A churn score is just a number.

A 0.82 risk score for an account tells you something’s wrong, but it doesn’t tell the team what to do.

A prioritized list says: here are the 10 accounts most likely to churn this week, here’s why, and here’s a recommended save play.

And in the setups that are actually driving impact, this isn’t a side dashboard. These prioritized lists are being fed directly into Salesforce or HubSpot, so the CS team sees it in their native workflow.

The risk score becomes part of the account object — paired with last contact date, usage anomalies, and a link to the playbook.

Tools like Catalyst are great for this kind of embedded loop, and platforms like Zapier or Tray.io can automate follow-ups — task creation, Slack alerts, renewal workflows.

That’s how you go from model to action.

So you’re not just modeling churn — you’re operationalizing the save.

In summary:

It’s not about more data. It’s about delivering the right data, to the right person, at the right time.

Another big use case is AI-enhanced lead scoring. Instead of sending reps down every dark alley, finance and sales ops teams are using tools like Rev, 6sense, and Clearbit to clean up their funnels and sharpen their ICPs.

The goal is to focus reps on accounts that are actually more likely to close.

Here’s how it works:

This isn’t about building a better spreadsheet of accounts — it’s about changing the shape of the funnel:

And just like with churn prediction, these scores aren’t trapped in a dashboard. They’re fed directly back into Salesforce or HubSpot, so reps see a rank, a reason, and a recommended action — all inside their normal workflow.

It turns outbound from guesswork into targeting. And it turns your RevOps team into a force multiplier for the go-to-market engine.

You can see how we’ll eventually want to hire people who are responsible for AI GTM operations. Kind of like a GTM engineer (which Kyle Poyar has identified as a new key hire to manage growth ops)

Early indicators show lead like this can lift win rates by 10% to 15% and lower CAC by 5% to 10% in the process.

Sounds small, but those wins compound over time, which is something we’ll come back to.

Everyone wants data access. But no one wants data disputes. And CFOs are tired of being the bottleneck.

We’re all trying to reduce ad-hoc data requests — but the bigger goal is avoiding the dreaded ‘dueling dashboards’ moment.

When someone walks into a meeting with a different number than what you’ve prepped, you’ve already lost the room. The conversation stops being about strategy — it becomes a data debate. And now you're on defense.

That’s why modern CFOs are investing in self-service BI tools like Looker, Mode, Sigma, and Canvas, but with one key twist: guardrails.

The idea is to give access without losing control.

This means:

And it’s evolving. I was recording with Sarah Riley, the CFO of DBT Labs, recently (recording coming soon!), and she essentially said:

“We’re trying to build tooling that lets you vibe code your way to the answer.”

It’s not about drowning in dimensions — it’s about giving operators the feel of fluency, without the steep learning curve. I personally don’t know Python, but I want to be able to query a data table like I do.

The outcome is bigger than fewer Slack pings. It’s about governed exploration that aligns the org around one version of the truth — while still letting teams move fast.

I’d be remise if I didn’t hit on some of the more practical and common use cases

These aren’t the splashy use cases that show up in investor memos — but they’re quietly saving finance teams hundreds of hours a quarter.

We're talking about:

These are the types of workflows where AI excels — high-volume, low-variance, rules-driven tasks that were built for automation.

1. First, on the expense classification front:

Tools like Ramp, Brex, and Navan now use AI to auto-categorize spend based on merchant, team, and description patterns.

These models get smarter over time, reducing manual GL tagging and flagging out-of-policy spend.

The result? Fewer finance fire drills at month-end.

2. Anomaly Detection

AI embedded in platforms like the players I mentioned plus procurement tools like Zip or accounts payable platforms like Onestream can detect patterns that deviate from norms — a sudden vendor spike, a duplicate charge, or even shadow IT.

This cuts audit time and improves compliance without needing a human to stare at rows all day.

Now, don’t get me wrong. You still want a human to do the final check. No one is pumped about 98% payables accuracy. But it makes it a lot less to review

3. Dynamic Cash Flow Forecasting

Traditional models are static. AI-powered forecasting (think OneStream, Planful, and Pigment) ingest real-time expense, sales, and cash data.

They update forecasts continuously — so you’re not presenting stale scenarios to the board.

4. Close Acceleration

Tools like OneStream, FloQast, and Numeric automate document parsing — scanning invoices, contracts, and receipts, and tagging them to the right GL accounts or workflows.

This shaves days off the close cycle, especially when your team’s small and the volume’s high.

These aren’t moonshot use cases — they’re compound efficiency plays.

And the best part? You can try most of them without an enterprise license or 6-month implementation cycle.

So those are the tools. But let’s talk about the psychology.

Most CFOs I talk to feel behind when it comes to AI — like everyone else is out there running fully autonomous finance orgs while you’re still stuck in spreadsheets.

But that feeling? It’s not backed by the data.

In our survey, 37% of CFOs aren’t doing much. Only 23% identified their orgs as AI first.

So less than a quarter are what we’d call advanced.

The vast majority are still piloting tools, experimenting with use cases, or even just trying to figure out where to start.

So if you’ve tested one model? You’re ahead of the curve.

You can join the 40% who are taking a more deliberate approach and adopting single use cases where they make the most sense

AI efficiencies can be categorized into two broad buckets.

Will this reduce headcount?

Or will this make my existing headcount more efficient?

Where AI replaces roles: Support, sales, and finance (high-structure, high-volume tasks)

Where AI boosts speed: Engineering, product, GTM (creative or technical work with leverage)

Knowing which bucket a tool falls into helps you model costs, reassign headcount, and avoid overpromising.

On the headcount reduction front, survey participants identified tools like 11x, Clay, Ema, and Unify on the GTM front to help with prospecting, and reducing the number of people they had to hire

And on the customer support side they identified tools like Decagon, Intercom, and Forethought to help with tier one support tickets.

And on the finance and accounting side tools that were mentioned include Navan for travel booking, Tabs for AI driven AR, and Onestream for planning and close. These are leading to a 10% to 20% reduction in headcount for high volume orgs

When it comes to making employees more efficient…

On the engineering side respondents identified tools like GitHub Copilot, Devin, Replit, Cursor and Bolt as favorites. This is boosting junior engineers by 30% to 50% per sprint and increasing code coverage and testing automation by 50% to 80%

On the product side we saw Replit, Bolt, and Lovable cutting down handoff time and allowing PMs to spend more time with customers.

And on the marketing side tools like Hubspot, Descript, Surfer, and Copy.ai are helping with multi touch attribution, content editing, SEO optimization and and programmatic content.

So where will this go over the next 12 months? Are we ramping up our spend on tooling?

To revisit a metric I gave you earlier, 5 to 10% of total OPEX seems to be the sweet spot for AI investment next year.

I was shocked that 17% actually said nothing.

And you saw very few saying they’d be spending more than a quarter which is also telling.

And if we go a layer deeper, where is that 5% to 10% coming from?

I hate to break it to software vendors trying to sell into orgs, but AI isn’t creating a net new budget.

There isn’t an AI budget fairy.

Only 20% of respondents said they were coming up with net new dollars

The other 80% said it was either being reallocated from existing budgets and coming from somewhere else, like the R&D budget

So we’ve covered the use cases, vendors, and psychology.

Let’s bridge the gap and hit on communication.

That’s the part we don’t talk about enough: how you communicate your AI strategy internally is just as important as the tech itself.

If you're positioning AI as a headcount reducer — you'd better get ahead of the fear. Because the moment you roll out a tool that automates tasks, people will jump to conclusions.

Even if you're just trying to eliminate wasted work, it can feel like you're eliminating them.

On the flip side, if you're using AI to boost productivity, but don't tell your team what success looks like, it turns into this vague initiative where no one knows what’s expected. It becomes a novelty, not a lever.

Communication sets the tone for adoption:

As CFO, you have to frame the story:

This is where you have to be clear on what’s the upside for the team, not just for the company?

People want to know what it means to them personally.

Because the difference between ‘This tool saves time’ and ‘This tool might replace me’ is often just how you talk about it.

The numbers made it clear that we should redefine what being “ahead” looks like.

PSA: Being ahead doesn’t mean building your own LLM.

It means getting 20 hours a month back by automating something tedious.

It means building confidence in the forecast. It means moving faster with the same headcount.

Because here’s the honest truth: most of the gains you’ll get from AI in the first year will feel… modest.

We’re talking 10%, maybe 20% improvements in a few key workflows — faster close, cleaner forecasts, fewer manual reconciliations.

But those small wins? They compound.

They free up hours. They create momentum. They unlock confidence.

And over time, they add up to real leverage.

Compound effort is the greatest force in business. But for it to work, you have to start. And the earlier you start, the more that effort builds on itself.

This isn’t about swinging for a 10x in year one. It’s about stacking 10% wins year after year — and letting them do the heavy lifting.

Gearing towards a close – why isn’t adoption happening as fast as we’d like?

The three most common responses were

What’s cool is 94% of the reasons are controllable variables.

Yes, there’s a lot of work to do, but you can choose to do it.

So here’s your playbook — and it’s not complicated.

Your team, your execs, your board — show that this stuff can create leverage.

Then review, iterate, and kill what’s not adding value.

That’s the strategy.

But here’s the mindset:

AI isn’t replacing you. It’s replacing wasted time.

The first gains might be 10%. Maybe 20%. They’ll feel incremental at first.

But those wins stack.

If you remember one thing from this talk: