Table of Contents
01
Calculating: The 3 critical questions
02
Building: The 8 how-to steps
03
Scoring: The 8 KPIs
HOW TO BUDGET FOR AI IN 2027

The AI budget nobody knows how to build is due soon

8 steps to get there
Download the Report

Budget conversations for 2027 need to start now, this month, while there’s still time to build a real, defensible number.

91%
of PE-backed CFOs say
they don't know how to size an AI budget

Eight in ten CFOs across the broader market are already planning to grow AI spend by more than 15% over the next two years. A growth target isn’t a sizing methodology, and that gap is the root of the problem these steps are built to fix.

80%
of CFOs market-wide plan
to grow AI spend more than 15% over the next two years

Download the full report.

Our contact form is currently blocked by your cookie preferences. Please change your preferences to continue.

Calculating: the 3 critical questions

A number built the traditional way, (take last year’s spend, add a percentage), doesn’t hold up under three specific questions:

  1. What does this dollar return?
  2. What does it cost as usage grows?
  3. Has anyone reviewed that number since it was originally approved?

Most 2027 drafts will fail at least one of those questions, because most AI budgets are still built on the same assumptions as a software license renewal.

A defensible number replaces those assumptions with three mechanisms. 

  1. Documented business case — math and rationale before funding, so spend is justified before approval, not after deployment.
  2. Usage-driven calculation — built from usage drivers (token cost, adoption rate, headcount) so the number moves with the business, not fixed until next year.
  3. Quarterly review cadence — so a number that’s right in September doesn’t go stale by December.
Why it matters

Running these formulas by hand every quarter is its own burden, so the more durable fix is connecting the inputs, headcount, spend, tool usage, and benefit assumptions into one shared source that recalculates them automatically as the business moves.

Why it matters

Building: the 8 how-to steps

01

Know where your sector sits before sizing anything

02

Tag every dollar to a key value lever

03

Fund the data foundation as its own capital investment

04

Fund the tech, workflow redesign, & change management 

05

Put a name on every seat in the room

06

Ask for the acceptance rate

07

Quantify the workforce question

08

Test every line item against a diligence question that hasn’t been asked yet

Step 1: Know where your sector sits before sizing anything

Manufacturing adoption jumped from 27% to 52% in a year. Healthcare climbed from 23% to 48%. Both moved because of labor shortages and margin pressure. Retail sits at just 39%, and that gap is worth taking seriously on its own terms.

Pull your sector’s adoption rate from the Accordion x Ramp AI Index and compare it against your own portfolio company’s current state before drafting a single dollar figure. If your company sits meaningfully behind its sector’s pace, build a catch-up allocation into the 2027 budget as its own line item, distinct from a generic AI increase. If it’s ahead, budget for consolidation and defensibility instead.

Step 2: Tag every dollar to a key value lever

EBITDA is the most common lever. Revenue per FTE works just as well for portfolio companies built around growth. The lever that matters is the one tied to the deal’s value creation plan. Name it before assigning dollars.

Uncertainty about sizing turns directly into pilots, lots of them, funded because doing something feels safer than admitting the sizing question hasn’t been answered. The rest are activity standing in for a number nobody built.

83%
of portfolio companies
have at least one AI pilot running today
18%
of those pilots
connect to an actual value lever

Pull every current AI line item, from vendor contracts to internal headcount to token spend, into a single list. Sort that list into two columns: dollars tied to a named value lever with a target number attached, and dollars that aren’t. Anything in the second column needs a lever assigned and a business case attached before the draft is due.

For the token line specifically, model it with a real usage driver: token cost per thousand, multiplied by calls per user per day, multiplied by days, multiplied by headcount, plus license fees. That number moves with headcount and usage, which keeps it current every time the plan updates. Before locking it in, assign a model tier to every task, since a routine data pull doesn’t need the same model as a complex forecast, and matching tier to task is the single biggest lever on that formula. Ask any vendor how they convert a proven, repeated pattern into a deterministic rule, which controls both the hallucination risk and the token bill.

Step 3: Fund the data foundation as its own capital investment

Sixty-six percent of PE sponsors say the biggest thing missing from their portfolio CFOs today is personal ownership of data infrastructure, funded and treated as a capital investment. No model performs well in a fragmented data environment. This foundation depends on three things working together:

  1. Clean, connected data. Fields populated consistently, source systems genuinely connected.
  2.  Wiring engineering. Engineering that can wire agents into how finance operates day to day.
  3. Private equity expertise. Knowing what a CFO needs at month 18 of a hold versus month 48.

That combination is rare inside any single team, whether that team sits inside finance, at a partner firm, or at a vendor. And it’s rarer than most budgets account for.

66%
of PE sponsors say
the biggest thing missing from their portfolio CFOs today is personal ownership of data infrastructure, funded and treated as a capital investment.

Run a data readiness check across the systems any prospective AI tool will touch, before allocating a dollar to that tool. Confirm the fields it needs are populated consistently, the source systems are genuinely connected, and a single owner is accountable for that data’s quality. Budget for any gap ahead of the AI tool itself, as its own line item, then put a name and a dollar figure against each of the three legs and be honest about which ones your own team can staff versus which ones need to come from outside it.

Step 4: Fund the tech, workflow redesign, and change management together

Confidence about AI’s impact on exit value is real:

For every technology line item in the draft, add two more next to it before submitting: one dollar figure for redesigning the workflow that tool touches, and one for the training and adoption work needed to get the team to use it. A technology line item without both of those isn’t ready to submit yet.

Step 5: Put a name on every seat in the room

Forty-one percent of sponsors scaling AI across multiple portfolio companies say they have no operational playbook behind it. Sixty-five percent of PE partners say there’s a moderate to severe gap between how their CFOs describe AI progress and what the sponsor observes on the ground, and a budget with nobody named to own the outcome is exactly what produces that gap. Five roles need a name:

Role 1: Owns the outcome and answers for it

Role 2: Defines the problem and prioritizes it

Role 3: Connects the work to a specific value lever

Role 4: Can say in 5 minutes if frontline behavior will change

Role 5: Accountable for closing the reporting gap sponsors flag

Most finance teams can fill one or two of those seats internally. Filling all five, especially the ones that require real PE operating experience rather than just technical skill, is where budgets tend to quietly assume capacity that doesn’t exist. The structure underneath those five names matters too: give AI spend its own budget by function, with a single named owner, governed at the CFO level.


Write the five names down before the budget goes final. Where a name is missing, treat that as a hiring, reassignment, or outside-partner line item, not a footnote. Budget for the roles themselves, not just the tools. Governance and AI-literate hiring sit outside the tool and data-stack spend, easy to overlook next to a flashier line item, and they’re what makes the other seven steps executable.

Step 6: Ask for the acceptance rate

Every AI recommendation flowing through a finance workflow should carry a human acceptance rate: how often it gets approved as-is, how often it gets edited, how often it gets rejected outright. This is one of the few numbers a vendor can’t spin, and it’s also one most vendors haven’t been asked for yet.


Request this number in writing as a condition of any new or renewed vendor contract, broken out by workflow so each function’s rate stays visible on its own. Build a line into the budget review template that tracks this number next to the dollar figure every quarter, alongside a variance flag for any line item where actual spend runs more than 15% over budget. Acceptance rate, the token formula, and Rev/FTE CAGR are a starting set. A mature AI budget review eventually tracks a fuller scorecard, spend as a share of OpEx, cost per execution, adoption rate, and a few others, but these three are the ones worth having in place before September. A low acceptance rate signals a usage problem. The team hasn’t adopted the tool deeply enough to trust it, and that gap shows up directly in the ROIC calculation Step 8 depends on.

Step 7: Quantify the workforce question

Economists and technology leaders, including 15 Nobel laureates and the chief economists at OpenAI and Anthropic, have warned that AI could reshape the economy on a scale larger than the Industrial Revolution, compressed into a fraction of the time.

200
economists and technologists
warning AI could out-pace the Industrial Revolution in a fraction of the time.

For every AI initiative in the budget, write one sentence naming which specific role or task it changes, and whether that change reduces headcount need, reshapes a role, or adds a new one. Attach a dollar figure to whichever of those three is true, whether that’s a severance line, a retraining budget, or a new hire. Before calculating the change, capture the baseline: tasks per FTE, cycle time, and error rate for the function before AI touches it. Measure the same three after deployment. Then track revenue per FTE over time: current revenue per FTE divided by the baseline figure, raised to the power of one over the number of years measured, minus one. A result at or above 10% is the number that turns this section from a sentence into a defensible line item.

Step 8: Test every line item against a diligence question that hasn't been asked yet

A line item earns its place in the 2027 budget if it survives the three questions below, and the way to test that is to run a mock diligence session before the real one happens.

86%
of sponsors already expect buyers
to price AI-enabled finance capability into valuation within two years

Before calculating ROIC, build the benefit side of that equation properly: time saved per task, multiplied by the loaded rate of the role doing that task, multiplied by how often it happens. That’s the annual benefit. Calculate ROIC for every major line item next: annual benefit divided by investment plus run cost. Put a real number on both sides of that equation before moving forward. Sit down with each major AI line item and answer, out loud, in the room:

  1. What value lever does it connect to and who is accountable for hitting it?
  2. What existing process does it replace and is there money set aside to retire that process?
  3. What will a buyer’s diligence team ask about this line-item eighteen months from now?

A line item that can’t survive that mock session out loud does not get a budget line.

Sizing an AI budget is a set of decisions: which pilots get promoted to funded line items, which capabilities get staffed instead of assumed, and which risks get named out loud before a sponsor names them first. The CFOs walking into September with a real number are the ones who ran their 2027 budget through these steps before anyone asked them to.

Most finance teams can execute some of these steps entirely on their own. Few can execute all eight without pulling in engineering depth, data infrastructure work, or PE-specific domain expertise they don’t currently have on staff. This is the same three-legged problem showing up across every portfolio company right now, and the CFOs who address that early are the ones who get ahead of it instead of discovering it in the diligence room.

Scoring: the 8 KPIs

Eight steps build a defensible AI budget. What follows is the discipline that keeps it defensible after the plan is approved.

Each metric traces back to one of the eight steps, translated into a number a CFO can report to a sponsor, a board, or a diligence team on a quarterly cadence. The foundational six can be built now, from data finance already holds. The advanced two are worth building toward over the next year, so the CFO walks into that diligence room with answers already in hand.

Where Accordion comes in

Accordion partners with PE sponsors and portfolio company finance teams to build the AI budget discipline these eight steps outline and turn that discipline into measurable results. The focus is sizing AI spend against a real value lever, embedding AI into the workflows that matter most, and delivering impact that shows up in cash flow, margins, and valuations.

 

Budgeting season is close. If you’re ready to build a number that holds up in the room, now is the time to start a conversation with Accordion.

Tell us where you need help

FAQs

How do you build an AI budget for 2027?

A defensible 2027 AI budget replaces last year’s spend-plus-a-percentage with three mechanisms: a documented business case before funding, a usage-driven calculation built from real drivers (token cost per 1K × calls per user per day × days × headcount, plus license fees), and a quarterly review cadence. Eight steps operationalize this: benchmark your sector’s adoption rate, tag every dollar to a named value lever, fund the data foundation as a capital investment, fund technology alongside workflow redesign and change management, name an owner for all five accountability seats, require acceptance rates from vendors, quantify the workforce impact, and stress-test each line item against a mock diligence session. For PE-backed portfolio companies, the discipline matters more than the dollar figure — 83% run at least one AI pilot, but only 18% of those pilots connect to an actual value creation lever.

What KPIs should a portfolio company CFO track for AI spend?

Six foundational AI KPIs can be built now from data finance already holds: Initiative ROI (annual benefit ÷ investment + run cost), % of AI spend tied to a named value lever, AI spend as a share of OpEx, driver-based AI cost forecast, AI efficiency gains reflected in the headcount plan, and % of AI initiatives with a formal control owner. Two advanced metrics are worth building toward over the next year: time to first measurable EBITDA impact, and AI portfolio concentration (top three initiatives’ AI spend ÷ total AI spend). Each traces back to a specific budgeting step and translates into a number a CFO can report to a sponsor, board, or diligence team quarterly.

How do you calculate ROI on AI spend?

AI ROIC is annual benefit divided by investment plus run cost — but the benefit side has to be built before the ratio means anything. Calculate annual benefit as time saved per task × the loaded rate of the role performing it × how often the task occurs. For the workforce dimension, capture a baseline first (tasks per FTE, cycle time, error rate) and measure the same three after deployment, then track revenue per FTE CAGR: current revenue per FTE divided by the baseline, raised to the power of one over the years measured, minus one. A result at or above 10% turns a workforce claim into a defensible line item. Accordion helps portfolio company finance teams put real numbers on both sides of that equation.

Why do most AI budgets fail sponsor and diligence scrutiny?

Most AI budgets fail because they’re built on software-license-renewal assumptions and can’t survive three questions: what return does this dollar produce, what does it cost as usage grows, and has anyone reviewed that number since approval. The symptoms are visible across the market — 66% of PE sponsors say portfolio CFOs lack personal ownership of data infrastructure funded as a capital investment, 41% of sponsors scaling AI across multiple portfolio companies have no operational playbook, and 65% of PE partners report a moderate-to-severe gap between how CFOs describe AI progress and what the sponsor observes. The fix is a mock diligence session run before the real one: for each major line item, answer out loud which value lever it connects to and who owns it, what process it replaces and whether money is set aside to retire that process, and what a buyer’s diligence team will ask about it eighteen months from now.