Ninety-one percent of private-equity-backed chief financial officers cannot size their artificial-intelligence budget, and 80% expect it to rise. Eighty-three percent of portfolio companies are running at least one AI pilot, and fewer than a fifth of those are tied to a defined value.
Finance teams have treated AI as a software line item. Justin D'Onofrio's argument is that it behaves nothing like one: token spend is consumption-based, and an approved initiative with no cap on it can take the whole budget with it.
"You can approve spend for an initiative, uncap that token spend behind it, and just have a complete budget blowout if you're not managing the process successfully."
D'Onofrio has spent 25 years at the intersection of finance and technology, runs the performance management team at Accordion, and is the person portfolio-company CFOs are calling three months before their 2027 budget is due.
The full interview is covered here so you can skip it. 57 minutes of audio, 22 minutes of reading.
Here are the 14 lessons that matter.
👤 Guest: Justin D'Onofrio, Managing Director of the Performance Management team at Accordion, who has spent 25 years working at the intersection of finance and technology in financial planning and analysis
🎙️ Host: A partner at ParkerGale Capital, the Chicago private equity firm that invests in middle-market software companies and produces the FunCast
📰 Published: 16 September 2026 on YouTube
🔴 YouTube | ⏱️ 57 min | ✅ Time saved: 35 min
Key Takeaways
The failure mode in AI pilots is ownership, not technology
D'Onofrio's line is that the pilots are not failing so much as belonging to nobody
Budget for the outcome, not the tool, because the tool will change
A line item naming the model tells a board nothing; one naming a 20% productivity gain can be tracked
Token spend is consumption pricing, which most FP&A teams have never had to forecast
One services company pulled a single month's bill running to many hundreds of thousands of dollars
Uber spent its full-year AI budget in four months, which is the proof that sophistication does not help
Nobody is taking costs out yet — the metric everyone watches is revenue per head
The data cleanup AI requires is contextual, not numerical
Policies, procedures and strategy documents, so the model has qualitative context for a quantitative call
Every experiment gets a leash and a clock: kill, fund or scale at 90 days
2027 is the reckoning year, and per-person spend caps are already appearing
1. Teenagers at the Mall
The host opened with the line he uses to describe where most companies are on AI spending, and then with the board meeting he had sat through the day before.
His description of the current state: corporate cards have been issued, there is no expense policy and no travel-and-entertainment budget, and some people are at the teenagers-at-the-mall stage of the AI journey.
The board meeting is the specific version of the problem. Half an hour went on it and produced no answer: the chief technology officer described the good work under way and the work to come, and the board asked how much it would cost and how anyone would know it was working.
The host's framing of why this is hard is that AI spend has no attribution model yet. Marketing spend has one — conferences, leads, top of funnel, conversion — and it took twenty years to build. AI spend has nothing equivalent.
The cause, on his account, is a deliberate decision a year ago. Firms told their teams to get AI-pilled, use everything and experiment; now everyone is discovering what that cost.
The exposure is not only large model bills. It is a Claude subscription on a corporate card, or a Perplexity account somebody forgot and has not logged into in six months.
What makes this urgent is the renewal calendar. Budget season brings the renewals, and with them a fight over which tools renew, which projects get the AI dollars, and which get starved for failing to prove anything.
D'Onofrio confirmed this is what Accordion is being called on to do now. His framing of the inbound: "And the most common one that we're getting question wise is how do I budget for AI?" And: "Every finance team is in the budgeting cycle and candidly a lot of folks just don't know how to approach it and don't know where to start."
2. Even Uber Can't Size It
D'Onofrio's context-setting is that the problem is not a mid-market problem.
"The one that was going around the past couple months has been Uber spending their full-year AI budget in four months."
"I mean we're talking about a sophisticated engineering forward organization not being able to size the prize with AI."
He also pointed to headlines about Meta shifting a large AI investment toward leasing rather than internal use.
"So, I think that the key takeaway is that even the mature folks, the folks that have been using and leveraging AI and making big big investments in it really still don't have that playbook unlocked yet."
Further down the market it is worse: a recent vendor survey found about 34% of mid-market finance planning teams are using AI in their process without knowing the impact, the return or how to budget for it.
The report behind the episode is Accordion's, and the host read its headline numbers. Ninety-one percent of private-equity-backed CFOs cannot size their AI budget and 80% expect it to go up. "83% of portfolios have one or more AI pilots going on" and fewer than a fifth are tied to value. Sixty-five percent of sponsors say data infrastructure or data ownership is missing, which matters because a model fed bad data will confidently return a wrong answer.
3. Budget the Outcome
The central reframe of the episode is that the line item is in the wrong unit.
"The spend for AI is not just a new software line." It is not a recurring fee negotiated up 5% or 10% a year.
"You can approve spend for an initiative, uncap that token spend behind it, and just have a complete budget blowout if you're not managing the process successfully."
He called it a new muscle for teams that have never dealt with consumption-based cloud pricing.
On the reported pilot failure rate: "From my experience, it's not that the pilots are failing, it's just that they don't really belong to anyone."
"So the ROI isn't defined upfront, what folks are looking to get out of that investment." Nor is the spend tracked once the investment starts.
The fix is to change what the line item names. "The tools can change. There's going to be a new model that saves you money, but making sure that you're thinking about the budget for an initiative AI-related in the context of the outcome, not the tool."
The host's restatement, which D'Onofrio accepted, is that a budgeting question here is a question about value in disguise. Instead of a line reading Claude, the line reads the outcome — for example, accounting team productivity up 20% from a given deployment — which makes the investment case easier to argue and to track.
The framework's first move is an owner. Every initiative, whether it is an investment in finding out what AI can do or an investment in improving a specific workflow, gets tagged to a named person in the business.
4. Three CFOs in Three Weeks
Asked for stories, D'Onofrio said he would give three examples from the previous three weeks and worked through two of them. His point is that maturity has nothing to do with company size.
"And I would say too that the interesting thing is the maturity for budgeting AI doesn't seem to be necessarily linked to company size."
The first was a media company CFO who said the firm was prioritizing AI spend. Asked how it was tracking and breaking that out: "And the answer was we're just not doing it yet." D'Onofrio thinks that is where most companies sit — they know they want to spend, and have no mechanism.
The second was a services company that pulled up a single recent month's Claude bill running to many hundreds of thousands of dollars and asked what had happened. The firm had historically spent very little on AI enablement, so nobody had been watching. Accordion was brought in to find the underlying causes.
"But you know both of these point to the state of affairs today which is we are investing in AI but we are not mature enough to manage it."
The three questions the host says are on every sponsor's mind: what does this dollar return and how is that attributed; what does it cost as usage grows, particularly where models are exposed to customers who can run up an uncapped bill; and has anyone reviewed the number since it was first approved.
5. Where the Tokens Leak
The host's own exposure, as a software investor, runs through the product; D'Onofrio's clients' exposure runs through the desktop.
For a software company the questions are how much AI sits in the product and how many tokens customers burn using it, plus the tokens spent building it, plus the general spread of Claude, Perplexity and OpenAI subscriptions across departments — and then every AI-assisted product on top of that.
Outside software, D'Onofrio says the first lever is personal productivity, and it is a legitimate investment with a condition attached. Companies early in the journey typically deploy a model to the internal finance organization to learn what the platform can do. He encourages that, provided there are guardrails on the spend and a stated expectation of what comes out of it.
The second lever is the data environment, on the principle that AI is only as good as the data it reads. Without a good data environment the output is garbage in, garbage out, which is why he sees clients spending heavily to shore up data infrastructure.
The host's reply was that this is the hard part, and he has the scar tissue. Portfolio companies have built internal servers that let any employee ask any question about any topic, and data governance is where it breaks: everyone has a file server holding version 1, version 1.1, version 1F and a file called the final version, and the model does not know which to pick.
One portfolio company has built an access layer over every model its employees use. Employees go through a company portal with a chatbot that picks the model by cost and question complexity, and serves a cached answer at zero cost where the question has already been answered.
The host's warning about the alternative: hand a few hundred employees a license or an API key each and the unexpected bill runs into hundreds of thousands of dollars, which can show up as basis points of gross margin nobody planned for.
6. Know Your Sector, Tag It
The report's eight steps start with benchmarking and with the tagging discipline finance teams already have.
Step one is knowing where your sector sits before sizing anything. "So you tend to see trends within industries of where the value capture is for AI and then within those industries you also want to look at the function that you're investing that spend in."
"So finance or accounting is going to have a different return than maybe customer service for example." He called this the due-diligence step: find out what comparable companies are actually capturing.
Step two is tagging every dollar, and the host's addition is granularity. Do not tag to engineering; tag to the specific project inside engineering — paying down technical debt, building new products, automating testing.
D'Onofrio's three required tags: "I think the three key tags are who's the assigned owner within the org, give them ownership, make sure they feel empowered to track and measure that over time in conjunction with the finance team." The second is the EBITDA lever — revenue increase, productivity saving or cost takeout. The third is the overarching program the spend rolls up to.
The prior step is a conventional one done properly: build the return-on-invested-capital and net-present-value case for the investment, exactly as a company would for capital expenditure planning. What is new is the owner tag and the explicit EBITDA lever.
His worked example of a case that adds up is the monthly board pack. A client says it takes 40 hours a month to assemble and analyze. Save 30 of those 40 and there is a clear return case for the tooling that does it.
The host's version of the same opportunity: close the books faster, automate collections, systematize the reporting that goes to the sponsor. He named an internal tool used across some ParkerGale portfolio companies, Maximor, which posts journal entries and has cut the close significantly at one large business with a small team.
7. Fund the Data Foundation
Step three is the one the host called the hardest, and the one that starts before budgeting season.
The host disclosed his own interest: ParkerGale owns Profisee, a master data management business that cleans up internal data.
His stat for why this matters: there is far more private corporate data sitting on servers worldwide than there is data on the public internet — and the large language models were trained on the public internet.
The practical obstacle he named is the departed employee. The finance analyst who left six months ago has material on a file server and in email accounts nobody knows exists, which makes a genuinely clean golden record very hard to reach.
D'Onofrio's business data confirms the demand. "The interesting thing is one of the biggest upticks that we've seen in business at accordion is we have an offering called a data platform which is effectively you know the old school data warehouse" — with current tooling on top — and clients are asking for that foundation in large numbers.
"I wouldn't say that's specific to AI. It's just been exacerbated by the need to deploy AI."
His examples of why clean data is not abstract: to ask a model which department has the most room for efficiency gains, the department structure has to be defined; to ask why vendor spend is up, vendor naming conventions have to be clean and consistent.
The genuinely new requirement is contextual rather than numerical. "The one thing that I'll add here is you know the historical sort of data cleanup exercise has largely been an exercise in actual data so call it you know numbers dollars. The thing that AI is introducing into the mix here is cleaning up your contextual data."
"So what are your policies and procedures? Where are your strategy documents? For AI to make a quantitative decision, it's helpful for it to have qualitative information to go along with that." Clients are tidying up knowledge repositories alongside the data platform.
The host's upside on that: a clean internal knowledge base plus a chatbot is the fastest way to push new strategy, rules and regulations through an organization without repeating them in every office.
8. Fund Change, Not the Tool
Step four is to fund the technology, the workflow redesign and the change management as one thing.
"Yeah, I think that the key takeaway there is, this is not specific to AI, but the implementation if you will of the technology is not the only portion of the change." The questions are whether people adopt it and whether the team's skills shift to support it.
"The tool can be great, but if nobody's using it, you're going to get zero value out of it."
The host's own conclusion, reached a year ago, is the cheapest intervention available: teach people how to use the basic tools already in front of them. That was the single most effective thing the firm did across its portfolio and internally.
D'Onofrio's addition is calibration, using Ethan Mollick's idea of AI's jagged frontier — the models are very good at some things and poor at others, and users need to learn where the edges are rather than generalizing in either direction from one experience.
The host's closing point on culture: the tools can do a great deal, but if people are not bought in and the reason has not been explained, that is the constraint. He has come to see this as at least as much a cultural challenge as a technical one, and said it usually takes one or two people in each department deciding to make it work.
9. A Name on Every Seat
Step five returns to the owner, which D'Onofrio treats as the cheapest and most-skipped control in the process.
"Making sure there's an owner. That's one of the easiest things you can do. Assign someone the ownership, the governance responsibility, the backing of the return on investment case, the adoption."
"That can be someone in the C-suite if it's a really strategic investment. That could be someone on the ground if it's more of a tactical sort of productivity investment."
He used the phrase throat to choke, said he hates it, and made the point anyway: somebody has to be in charge of tracking and realizing the value. Making that a gating step is the part almost nobody does at volume.
The host's amendment is the important one. ParkerGale's internal rule is that if you have the information you have the authority — so the owner also gets the authority to call a timeout, and an incentive to come back and say the money should stop because the organization is not ready.
His reason is a failure mode he has watched: an AI-enthusiastic chief executive hands over budget and authority, the owner does not want to disappoint, and the problem surfaces six months later when a private equity associate asks a question at a board meeting.
10. Ask the Acceptance Rate
Step six is the one D'Onofrio flags as a more mature practice, and it is the only real measure of whether a deployed system is working.
"So you you put AI in place to help you make decisions within a process. What does good look like in that?"
His worked example is a model producing a deep dive into vendor spend, with a usable output — an explanation, a recommendation to increase or rationalize spend with a given vendor — half the time, accepted without editing.
"What's a good acceptance rate? What would the team find valuable? And that's going to be case specific."
The premise underneath it is that outputs are never perfect. "What's the acceptance rate that really drives value and where do you have diminishing returns with that?"
The host's extension is that perfection is structurally unavailable, because capability and ambition rise together. As models improve and context windows grow, users ask for more, so accuracy slides back toward 50% and is worked back up to 90% before the next expansion. His comparison is what anyone asked a chatbot to do a year ago against what they ask now.
His framing of the calendar: last year was the tinkering year, this year is the implementation year, and next year is the measurement year.
D'Onofrio's correction to that is the sharpest line in the episode: "And I would call it the reckoning year because I think it's already starting to happen where we kicked off the call with some of the big headlines, but those same organizations that were token maxing, blew out their budgets in four months." Those firms are now putting guardrails in, including a fixed cap on spend per person per month.
11. The Workforce Question
Step seven is quantifying the workforce effect, which D'Onofrio called one of the most important and least discussed.
The method is to take a census of the current workforce and group people into cohorts by business unit, department and job title — finance analyst, customer service representative, engineer — and then plan headcount growth and AI impact by cohort over time.
"And really the goal here is to make sure that your workforce plan, so your growth in headcount or lack thereof has strategy behind it in the context of AI because again some functions within the organization are going to be severely impacted by AI."
His asymmetry: first-line customer support can be pushed to an agent, while back-office functions are less exposed. The industry lens matters here too.
The host's own read, which D'Onofrio endorsed completely, is that mid-market private equity is not cutting finance headcount. The realistic outcome is not hiring as the business grows.
"I think, the trend we're seeing for the companies that are thinking about AI spend is the key metric that most folks are looking at is revenue per head and how that is increasing or flattening over time."
"I have not seen a lot of clients look to do cost takeouts yet with AI." The measurable thing is productivity, and he compared the whole conversation to the one the industry had when robotic process automation arrived.
The host's description of why small finance teams are the natural first target is the ratchet. When private equity arrives, data requests intensify after close rather than tapering; answer fast and more is asked, and a 40-page monthly pack becomes an argument for a 60-page one.
12. The Diligence Question
Step eight is to test every item against a diligence question nobody has asked yet.
"And you're going to start getting this transition to why did you spend this? What was the investment case? What return are we getting here? So, make sure you have that answer up front."
"It's going to come quickly if it hasn't already, particularly in the in the PE space."
The purpose of the preceding seven steps is to make that answer available — including the answer that it did not work and was cut quickly.
The host described the pressure management teams are under from the other direction: sponsors asking for more AI because another portfolio company did something, in a different industry, at a different scale, with a different sales cycle — which is an easy comment to make in a board meeting and a hard one to act on.
His summary of the position: this is a squishy budgeting season, the efficiency and growth results are unknown because they have not happened yet, and the honest sequence is to start with something, make a guess and test against it.
D'Onofrio agreed, and named the structural reason. Financial planning teams are good at forecasting where there is historical data to trend from and poor where there is none, and most companies have no history of AI spend and return. "You're never going to be 100% accurate, you're actually probably going to be more inaccurate than accurate." The answer is to tag it, build the new muscle, track whether the expected value arrived, and kill, fund or scale accordingly.
The host's suggestion for 2027 is to split the year: approve six months, and greenlight the second half only if the first is working, rather than letting a 12-month budget run unexamined.
13. How Accordion Engages
The host asked what a client engagement actually looks like as the phones start ringing before Thanksgiving.
The request Accordion gets most often is generic — a plan for AI spend — and there are two distinct answers to it.
The first is a financial planning toolkit: templates, tracking, the return-on-invested-capital capture structure and the mechanics of modeling token consumption. The point is that the client is not starting from a blank workbook.
"But the important thing is meeting you where you are with the structure, not necessarily the end result or what technology you use for that." If the client plans in Excel, Accordion builds it in Excel; if they run a planning platform, it gets built there.
The second engagement answers a different question — where to spend — and starts with a workshop. Accordion benchmarks the client's internal processes, measures how long things take and what the return is, builds a matrix of investment against return for each AI use case, and delivers a road map of the opportunities it sees.
The host's framing of why both are in demand: every management team was asked at its second-quarter board meeting how it plans to grow faster and spend less in 2027.
The calendar he laid out: a draft budget before Thanksgiving, negotiation between Thanksgiving and Christmas, finalized by year-end, to the banks by the end of January.
14. Referee, Not Dr. No
The closing exchange is about the position the chief financial officer is put in, and the host's worry is that the loudest voice wins.
His concern, particularly in the lower middle market, is that the funded projects are the ones that sound coolest — put AI in the product and sell more of it at a higher price; put AI in customer service and remove half the heads because another portfolio company did — and the CFO ends up refereeing on personality and momentum rather than evidence.
D'Onofrio's first step is to set the strategic buckets with the management team. "I think the first thing to do is you know with the management team senior folks really determine the strategic buckets of spend that you have as your organization." For a software company that usually means product investment to hold or grow share.
Then push the work down. "Because if you're doing all the work for them, you're going to be accountable for the outcomes versus giving them the structure to think properly about how the spend actually unfolds and how you can capture value from that spend."
"That's just the best thing I think a CFO can do at this point is you know teach a man to fish, so to speak."
The host's addition is an early, honest communication: this is new for everyone, the organization has covenants and goals, not everything can be funded, so come with a return case, a tracking method and check-in points — which makes it easy for the CFO to say yes to an experiment.
D'Onofrio's caveat against pure gatekeeping: "You never want to have that always no CFO lens hat on, right?" There should be a deliberate internal enablement fund. "Set that aside. Know those are going to be probably low ROI but good for the organization to kind of understand the muscle that they're going to need going forward. But for the big bets, make sure you have that process in place that we talked about."
The host's reporting format: the first page of every board deck should be the five keys to the investment case and how they are measured, and next year the second page should be the AI initiatives and their tracking. Showing that every quarter, he said, buys the credibility to cut something that is not working.
His view of the incentives is the reassuring part for a CFO. There is credit for doubling down on a project that works, and as much or more for cutting one that does not — whether the technology failed, the scope was wrong, the ambition was too great, or the organization was not culturally ready.
Bonus Insights
D'Onofrio's closing recommendation, which he flagged as the one thing he would add: run a monthly or quarterly reforecast specific to the AI spend buckets, because the cost can balloon and the value capture can change quickly.
The host's response to the teenagers-at-the-mall line was that he would have been dangerous at Crossgates Mall with his parents' credit card in the 1980s.
Accordion's report behind the episode is titled "The AI Budget Nobody Knows How to Build Is Due Soon," and sets out the eight steps.
The show trailed an earlier episode with D'Onofrio's colleague Kyle Romer on how AI is changing finance and accounting teams, which the host said several thousand people have already watched.
ParkerGale is itself engaged with Accordion on this at one of its portfolio companies.
The host said his own firm is going through the same exercise internally for 2027, and that his instruction to his team is to pick a winner and move on.
D'Onofrio's bottom line is that nobody will forecast AI spend accurately in this budget cycle, and that this is not the point: what separates the companies that come out of 2027 with a usable playbook is whether every dollar was tagged to an owner, an outcome and an EBITDA lever, and reviewed often enough to kill what is not working.
Products, Companies & Tools Mentioned
Accordion (D'Onofrio's firm — a private-equity-focused financial consulting business whose performance management team builds the budgeting frameworks the episode describes, and whose data platform offering is seeing its biggest uptick)
ParkerGale Capital (The host's firm: a Chicago private equity group investing in middle-market software companies, and the FunCast's publisher)
Profisee (The master data management business ParkerGale owns, raised as the example of cleaning up internal data before pointing AI at it)
Uber and Meta (The two headline cases D'Onofrio cites — a full-year AI budget spent in four months, and a large AI investment shifted toward leasing)
Claude and Perplexity (The subscriptions that turn up on corporate cards without an owner, and the source of the six-figure monthly bill one services client could not explain)
Maximor (The journal-entry automation tool ParkerGale uses across some portfolio companies, credited with materially shortening the close at one large business with a small finance team)
NetSuite (The host expects it to launch competing accounting automation products in the fall)
Books & Resources Mentioned
The AI Budget Nobody Knows How to Build Is Due Soon – Accordion (The report behind the conversation, setting out the eight steps)
Ethan Mollick's jagged frontier (The idea D'Onofrio uses to explain why people should calibrate where AI is strong and weak rather than generalizing from one result)
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: