Last week, Secret CFO stressed the importance of AI-native core infrastructure: "I'll say this until I'm blue in the face — the biggest wins from AI in finance will come from aggressively implementing AI-native core infrastructure."
We couldn't agree more.
Campfire is the AI-native finance platform built for exactly this moment. No legacy ERP retrofitted with a chatbot. It’s purpose-built AI that lives inside your financial workflows from day one.
See what that actually looks like in practice. Book a demo today.
This chart from Private Equity monster Carlyle caught my eye this week:
Note: the survey asked where management teams expect to see offsets for AI spend, in the future tense, not where they're seeing it today.
The message I took from this: where businesses are already deploying big into AI, they are not looking to replace their software costs. They are expecting a radical reshaping of their overhead base: people, services, infrastructure, and software.
And you'd be forgiven for thinking the math is simple right now… AI is cheap, people are expensive… surely we should just swap one for the other?
Last month, Jack Dorsey cut 40% of Block's workforce and blamed AI. Block had also significantly overhired post-COVID, but that detail got lost in the headlines.
The AI ecosystem has exploded like a gold rush; chaotic, self-reinforcing, and with the smartest (and loudest) people in the room knowing that the safest path to gold is selling shovels. I know that’s a cliched line, but if ever the analogy was apt, it’s here.
Everyone in this ecosystem has a vested interest in framing AI as abundant, limitless, and basically free.
The model is simple enough: land grab as much market share as possible now, hope the unit economics improve over time, and if they don't… well, that's what pricing power is for. VCs put up capital. Chip guys sell chips. Model companies build data centers and collect record-breaking revenue that still falls well short of covering costs. Everyone talks their book.
There is a relentless, circular velocity to it.
Big Tech CEOs, AI Founders, and VCs meeting for their monthly PR strategy alignment session… probably.
I nearly pasted a couple of AI Finance LinkedInfluencer profile pictures into the middle of this image, but decided against it in the end. So, I’ll leave that to your imagination…
But eventually that circular economy needs to be punctured by real, delivered value and long-term net productivity from outside itself.
It's not an unreasonable thing to believe there will be real productivity… it's just not here yet. And when it arrives, it'll be CFOs quantifying it.
Us delivering the message on earnings calls and in board reports. A solid green bar on a year on year EBIT bridge with a big ‘net AI savings’ label on it. While a baying crowd waits outside the papal chimney for white token-smoke.
So… how the hell do we do that?
Welcome to the final part of this four-week series: The No BS Guide to AI for CFOs
In week 1, we made the case that it's time to stop dabbling with AI and start deploying in your finance team.
In week 2, we got under the hood, what AI should and shouldn't touch, and how to put guardrails in place.
Last week, we worked through the technology decisions: what to buy, what to build, what to borrow, and how to keep your data safe while you do it.
This week, we get into the (literally) trillion-dollar questions CFOs are asking about AI today:
What is this genuinely going to cost? And not just now, but later too.
How much can you really expect to save on payroll? And what happens to the people?
What are the unmeasurable benefits, and the costs nobody is talking about?
How do you build a board-ready business case when the numbers keep moving?
How do you know if it's working?
ROI takes a back seat… for now
Let's skip to the punchline. Trying to do detailed ROI math on AI right now is basically pointless.
Forcing traditional business case logic onto AI investment, project by project, will be an exercise in false precision. Not to mention, too slow.
It's a bit like Arthur Dent in The Hitchhiker's Guide to the Galaxy, lying in the dirt in front of the bulldozer about to demolish his house. Clutching his planning objection forms, while the Vogon constructor fleet is already in orbit above him (with a much bigger plan for Planet Earth.)
OK, it’s not quite like that, but you get the point…
For now, CFOs need to facilitate learning budgets for the organization with AI. And trust that it'll be worth it in the long run.
That means that while the true economics are still forming, you should think of AI investment as a below-the-MFCF-line bet… what we called a ‘Moonshot Pot’ in our recent series on cashflow and capital allocation. Capital ring-fenced for asymmetric experiments.
Protected from the kind of short-term return focused decision making that kills innovation before it starts.
It can't live in that zone forever, but for now, it’s where it belong.
At some point though, the experiments need to graduate into the P&L or get killed. Set a budget, define what you're trying to learn, and put gate posts in so you know what each experiment has delivered before you decide whether to scale or stop.
So while we shouldn't paralyze innovation with a false hurdle rate based on predicted impacts, capturing the results of the net costs and benefits of the actual impact is crucial. Because the further we get up the learning curve on AI, and the size of the commitments needs to grow, the more we will expect these to become targeted, financialized bets.
A generational shift in business cost structure
So, why is measuring this so complicated for finance?
CFO’s measurement systems are typically built for two types of expenditure:
Recurring expenses that are fundamental for servicing the current level of revenue
J-curve investments that are made now in the expectation of a return later.
But AI deployment is not like any other ‘J curve’ investment I’ve seen. It’s more of a generational shift in fundamental cost structures of businesses.
I’ve been in industries with aggressive trends towards automation, seeing a rapid shift from ‘variable’ labor cost to CapEx or fixed overhead, e.g., manufacturing robotics. But while they feel uncertain at first, those are comparatively easy to quantify once you’ve proven the logic.
AI is different. We aren’t talking about a straight switch from labor costs to capex, or even labor into software. There is so much more going on.
We are going to get into that in detail now across five buckets:
Software and AI costs
Headcount costs
CapEx and infra costs
Shadow costs
Shadow benefits
Let's get into each of those five buckets and see just how complex the financial impact of enterprise-wide AI adoption really is:
Software and AI Costs
There are so many moving parts here. Let me get forensic.
Token costs. Like it or not, you are going to need a token budget. Tokens are the new currency of AI (tiny imaginary units that almost nobody in finance understands yet). Controlling token spend is about to become one of the biggest cost governance challenges facing finance teams.
Replaced software. AI will retire some legacy software costs, or at least facilitate an upgrade path. That should mean a saving, but it also means exit costs, transition periods, and the uncomfortable overlap phase where you're paying for both the old and the new.
New software. It's not as simple as a straight swap of software costs for tokens. AI adoption enables new software adoption too (especially in finance, as I covered last week). Setting up governed, scalable AI environments creates real demand for new AI-native platforms and tooling.
Governance infrastructure. Prompt libraries, skills, projects, audit frameworks. If you don't build this in-house, you'll buy it from someone else. Either way, it costs money.
The agentic multiplier. Once you move into agentic workflows, a single routine can call the model multiple times. Gartner recently reported that agentic workflows consume 5-30x more tokens than a standard model call.
Tokenomics. Today, frontier models charge ~$2 per million input tokens and ~$15 per million output tokens. One token is roughly 3-4 characters of text. The foundational model providers are selling tokens at an outrageous loss right now. They will bring unit costs down over time, but they will also use pricing power if and when they need to.
Shadow AI. You'll accumulate ungoverned tools outside the core business case, too. Expensed ChatGPT subscription upgrades. C-suite leaders are going rogue to get the tools they want.
SaaS inflation. The legacy SaaS tools staying in your stack will find a way to use their own AI features to inflate your costs over time. Microsoft already added $3 per user per month for Copilot in M365. This is just the beginning.
Learning budget vs steady state. Two distinct phases with very different cost profiles. The learning curve inefficiencies on AI costs will run for years. So the shape of these costs over time will be more volatile than we are used to
Wowsers.
If, after reading that, you can tell me where your AI and software costs are heading, and when (other than up), you are a lot better than I am.
And that's just software.
Headcount Costs
So will payroll costs go down?
Well… it should, but only if it’s well controlled. Here’s what I see:
Replacing existing work. The most obvious dynamic. The theory that AI will eventually replace existing human work is robust. This should emerge as a downward force on payroll costs, but there is a lot of nuance here.
The capture problem. Just because there is a theoretical productivity benefit doesn't mean it will become a delivered one. Initially, these gains will be hard to see and hard to capture, because they'll be incremental. If you make everyone 10% more productive, does a team of 10 become a team of 9? Or do 10 people just have 10% more ‘thinking time’? Hybrid working has made the visibility of partial productivity savings even harder to pin down.
Mix shift. A 30% hours saving doesn't mean a 30% payroll cost saving. AI will target entry-level work first, meaning the average cost of every headcount saving will likely be lower than the org/department average. In the short term, there will be new review and governance burdens on management teams learning how to oversee AI-produced work. That creates an upward force on average cost per head even as volume falls.
Central governance cost. The hub in the hub-and-spoke model costs money. An AI center of excellence, a platform team, and a FinOps function for AI spend.
Specialist premium. The people who remain become more valuable and more expensive. Rate pressure goes up as the market for AI-fluent finance professionals tightens.
Learning and change management. Consistently the most underestimated cost in any technology transformation. And it will be bigger than AI for any other tech change, because there is a whole new language and way of being to internalize.
Workflow obliteration. This is the big one, but it will come later. AI will eventually lead to the obliteration of whole workflows, and probably even whole functions. Middle management globally is full of roles that are essentially just the cost of moving information and coordinating people. There is radical savings potential here, as organizations redesign entire org charts around AI-native processes.
In sum, you can see a path certainly to a radical productivity from AI across the long term, almost universally. But do I actually think there’ll be a net headcount reduction? Probably not.
More likely those savings get reinvested in faster deployment, more productivity and more growth. Or in some cases, just … wasted.
But one thing is for sure … we will need more robust headcount controls and reporting than ever to ensure we are both giving the business the oxygen it needs to implement, to offset the token costs, and harvest the cost savings.
We will need to understand and be able to bridge these ups and downs to bring the insight to the business it needs to make these decisions
CapEx and Infrastructure
We could also see some movements in the IT CapEx line. This kinda depends on how you approach AI, whether you will rely on custom AI builds or focus more on implementing AI powered software.
But the data sovereignty question will ask real questions of cloud vs on-premises, and if there is a trend back towards on-prem that would mean CapEx, lots of Capex.
Shadow Costs
And what about those things that are much harder to see?
Switching costs. AI-native platforms create new lock-in. Somewhere on the journey you will make big platform bets (possibly without realizing that’s what they are).
Compliance and audit costs. AI adds new requirements for explainability, auditability, and model risk management. Material, growing, and difficult to forecast.
Sprawl. The volume of ungoverned custom-built tools is exploding even as the cost of any single one falls. Pushing the total system risk (and governance cost) up.
Getting it wrong. A hallucination in a board pack. A privacy incident from shadow AI. An agent executing the wrong workflow 10,000 times before anyone notices. This is the thing we are all worried about
Shadow Benefits
While I do expect there to be a significant net durable productivity benefit, when you account for all of the above, I don’t think in $ terms it will be as big as people are speculating.
But I also don’t think directly attributable cost savings are the real prize from AI. The real prize lurks deeper, is much bigger, but naturally is harder to capture:
Speed. The business that makes better decisions faster than its competitors has a structural advantage that compounds over time. AI compresses the gap between information and action across every function and every layer of the organization. And speed is still the most underrated weapon in business. This is a lesson that is consistent in almost every major founder biography.
Reinvestment for growth. In practice much of the so called savings from AI, more likely … will be redeployed back into the business to grow harder and fast. All the more reason why they’ll be hard to measure.
Customer experience. Faster responses, fewer errors, more personalized interactions at scale. It might be hard to attribute, but it doesn’t mean it’s not there.
Prototyping and iteration speed. New products, new markets, new propositions, tested and iterated in weeks rather than years. Reducing the cost and time of an experiment, will mean better products faster.
Talent. The organizations that move decisively on AI will attract the best people. The ones that don't will lose them to the ones that do. This dynamic is already playing out and it will accelerate. You can bet that is why Ramp have been so public about their approach to AI… it’s a talent magnet.
Competitive moat. Proprietary data, proprietary workflows, proprietary institutional knowledge. All of it can become valuable and more defensible when it is embedded in systems that compound over time.
Optionality. The organizations building AI capability now are creating options that don't yet exist.
So … about that ROI case.
If you've read all of that, can you tell me how to build a robust ROI model for AI implementation today? Well… you are a far better finance professional than I am.
You could spin in circles chasing false precision.
The honest answer is that you have to lean into this as an undeniable, and inevitable structural change in the shape of your P&L, one that our traditional measurement systems are not built for.
So what can you do?
Baseline first. Before you can measure the impact of AI, you need to know what you're measuring against. Most organizations don't have a clean baseline of current process costs, cycle times, and error rates. Without it, you can't attribute any improvement to AI with confidence. Good old fashioned time and motion studies are about to become very popular again.
Allocate capital to experiments. Like I said earlier … accept that real AI adoption means deploying capital with no clear return promise in the short term. Some investments just go like that. Ring-fence a budget, define what you're trying to learn, and treat it as a Moonshot Pot, not a line item waiting for a payback period.
Stage gates. The Moonshot Pot needs kill/continue criteria built in. Explicit points at which you evaluate whether an experiment graduates to the P&L, gets more funding, or gets stopped. Without stage gates, experiments quietly continue forever. Every CFO knows this from CapEx. But, it applies equally here.
High-fidelity cost capture. Adding an "AI spend" account to your chart of accounts is not going to be enough. You need to capture AI spend in the most granular detail possible. And it’s not just the costs of AI, but throughout the P&L. Atomic measurement of your cost base is going to become very important; by tool, by team, by process, by person, by workflow, by token…
Name an owner. Who owns the AI cost line? Token costs are appearing on expense reports, corporate cards, and software invoices simultaneously. Someone needs to own the total number to make sure someone is looking at the whole damn picture while you put some control around it.
Build guardrails around token spend. The potential to get the 2026 equivalent of a 2006 mobile roaming charge surprise is very real. Put hard limits on total spend as a minimum. You don’t want to wake up to find an agent doomloop ran overnight and burned through your quarterly token budget before breakfast
Benefit capture. While benefits are hard to forecast for all the reasons above, capturing and reporting the costs and benefits of AI implementation as they land is crucial. It will help you learn what is working, double down on it, and report the success (or failure) honestly.
Net Net:
As CFO, you could crush every AI experiment before it leaves the nest by applying traditional ROI-based approval hurdles. Or just as easily, you could force the business into false financialization. (One of my least favorite qualities in a CFO.)
Both would be wrong… this is a generational shift in overhead structures.
You need to allocate capital (money and time) to AI implementation and treat it as a learning budget. Then scale your spend appetite as you learn what works and what doesn't. As the business starts to deliver more concrete net benefits, you can commit more capital. But you have to give this thing some oxygen first.
And on that note, we end this series The No BS Guide to AI for CFOs.
Go forth and AI…
Next week, we’ll dive back into the cashflow and how to street fight your way to a better cashflow profile with a new series; Working Capital Warface
If you’re looking to sponsor CFO Secrets Newsletter fill out this form and we’ll be in touch.
:::::::::::::::::::::::::::::::::::::::::::::::::::
:: Thank you to our sponsor ::
:: Campfire ::
:::::::::::::::::::::::::::::::::::::::::::::::::::
What did you think of this week’s edition?
If you enjoyed today’s content, don’t forget to subscribe.
Disclaimer: I am not your accountant, tax advisor, lawyer, CFO, director, or friend. Well, maybe I’m your friend, but I am not any of those other things. Everything I publish represents my opinions only, not advice. Running the finances for a company is serious business, and you should take the proper advice you need.
