← All Insights

CFOs Do Not Have an AI Budget Problem

AIOffice of the CFO

Accordion’s published framework puts three numbers on the problem: 91% of PE-backed CFOs say they do not know how to size an AI budget, 83% of portfolio companies have at least one AI pilot running, and only 18% of those pilots connect to an actual value-creation lever. Accordion does not disclose the methodology for those three figures on that page, so I would not underwrite the entire market on them.

Its separate 2026 PE AI Adoption Benchmark is more transparent. Accordion and Wakefield Research surveyed 150 AI, data, and technology Operating Partners at PE sponsors. Forty-one percent said their firms were deploying AI across multiple portfolio companies without an operational playbook.

Accordion’s prescription is largely right: document the business case before funding, calculate cost from real usage drivers, quantify the workforce impact, and review the result quarterly. The ridiculous part is that companies still treat AI return as if it cannot be measured.

Measuring AI spend and return is not unusually hard. In many ways, it is easier to measure than the consulting projects, software implementations, and additional headcount companies have approved for decades. The model providers meter usage. The workflow has a baseline. The labor has a loaded cost. The output has a cycle time and an error rate.

If a CFO still cannot connect those pieces, the problem is not AI.

The problem is that management does not understand how the work gets done.

The cost is already sitting in the meter

OpenRouter automatically returns usage with every model response. OpenAI provides token counts in each response and a usage dashboard with project filters and exports. Anthropic exposes cost reporting by model, token type, service tier, and workspace.

This is not a black box. Every call can carry a team, synthetic user, workflow, client, or cost-center identifier. Finance can know which model was used, how many tokens it consumed, what the call cost, and which business process produced it. Separate API keys and projects can make the allocation even cleaner.

Compare that with a traditional transformation program. A company hires consultants, buys another enterprise software package, pulls half the leadership team into weekly steering meetings, and hopes somebody can explain the return a year later. The labor gets scattered across cost centers. The meetings disappear into calendars. The opportunity cost never reaches the model.

AI gives you a receipt for the machine work.

The missing receipt is for the human work it replaced, and that is management’s job.

The allocation is straightforward even when the vendor charges a fixed license instead of a metered token bill. Assign the license across the workflows using it, based on seats, runs, or another consistent usage driver. Bundled enterprise products may not expose token usage by workflow, but they still provide seat or activity reports that finance can allocate consistently. Require every system to log the cost center, workflow, user, model, and run identifier. Reconcile those logs to the vendor invoice each quarter. A bundled price changes the denominator. It does not make the cost unknowable.

Privacy does not excuse the measurement either. The ledger needs a synthetic workflow identifier, usage, cost, and outcome. It does not need the client’s name, the prompt, or the confidential output. Where detailed logging is restricted, finance can allocate from aggregate usage and audit a representative sample. In regulated workflows, compliance should define access, retention, and sampling before deployment.

Start with the task, not the technology

A real AI budget starts at the ground level. It is the same discipline I described for putting AI against recurring finance work like invoicing and collections: map the workflow before anything else. Pick a workflow and capture five things before the system touches it: volume, cycle time, loaded labor cost, error rate, and headcount capacity, measured as available FTE hours or utilization.

Then run the same measures after deployment.

Take a client proposal. If a team used to spend a week gathering material, structuring the argument, building the deck, and cleaning up the presentation, and a well-designed AI system now produces a sophisticated first version in minutes, the benefit is not theoretical. Count the hours removed. Multiply them by the loaded rate. Measure how many more proposals the team can produce and how quickly a client receives one.

The operating case fits on one page. Annual gross benefit equals baseline hours minus post-AI hours, multiplied by loaded hourly cost and annual volume. Add avoided compensation or incremental contribution from higher capacity. Then subtract implementation, licenses, model usage, and ongoing review. Payback is the implementation cost divided by monthly net benefit.

The same math applies to underwriting memos, diligence summaries, customer-support research, contract reviews, reconciliations, sales preparation, and board materials. The unit changes. The method does not.

Quality needs the same discipline. For a recommendation, track how often the human accepts it without changes, edits it, or rejects it. For a reconciliation, track exceptions and corrected errors. For a proposal or board deck, audit a consistent sample against the same factual, commercial, and presentation standards. “Quality improved” is not a measure. Acceptance, rework, defects, and review time are.

I am seeing this in my own work. Sophisticated proposals and decks that once took a week can now reach a reviewable first draft in minutes. When the system also handles the gathering, first-pass analysis, follow-up, and quality checks, a capable manager supervising twice the work becomes a legitimate operating target that can be tested against throughput and quality.

That is measurable value.

The productivity evidence already exists

The external research does not tell a CFO what return a specific workflow will produce. It proves that the return can be measured.

An NBER study covering 5,179 customer-support agents found that access to a generative AI assistant increased issues resolved per hour by 14% on average and by 34% for novice and lower-skilled workers. The researchers measured the unit, the baseline, and the change.

A field experiment with consultants found that people using GPT-4 completed 12.2% more tasks, worked 25.1% faster, and produced results rated more than 40% higher in quality on tasks inside the model’s capability frontier. A controlled GitHub Copilot experiment found that developers with the tool completed a programming task 55.8% faster.

Those figures are not universal forecasts. They are examples of the discipline every CFO can impose: define the work, measure the output, compare it with a control or baseline, and convert the difference into labor capacity and dollars.

Put the workforce plan inside the AI plan

The discomfort starts here. Companies want an AI budget without a headcount budget attached to it.

That makes no sense.

For every funded initiative, the CFO should require one sentence naming the task or role that changes. Does the system eliminate a position, avoid a planned hire, allow one manager to cover a larger team, or move an employee into higher-value work? Then put the dollars next to it: severance, retraining, avoided compensation, incremental revenue capacity, and model run cost. AI is a labor model change, not a feature layer, and the budget has to show it.

The cost side needs the same honesty. Include data cleanup, workflow design, integration engineering, security review, evaluation, employee training, and ongoing human review. Leaving those items out does not improve the return. It creates a fake one.

If ten people perform a workflow today and the redesigned system needs six, the business case cannot stop at software savings. It needs to show the four positions removed or avoided, the implementation cost, the ongoing inference cost, the review labor that remains, and the payback period. The model bill can be tied directly to the workflow. The headcount effect can be tied directly to the operating plan.

There is no reason for the connection to be vague.

The hard part is not the arithmetic. The hard part is making the decision and executing it. Companies delay the headcount question because it is politically difficult, then describe the return as unmeasurable. That is not financial discipline. It is avoidance wearing a spreadsheet.

Budget by workflow, not by AI

The phrase “AI budget” is part of the problem. It encourages finance to create one large bucket, mix software licenses with model usage and experiments, and then ask what the bucket returned.

Nobody should budget that way.

Budget the work instead. Each workflow should have a named executive owner and a one-page operating case:

  1. What does the process cost today in labor, systems, errors, and elapsed time?
  2. What will implementation and ongoing model usage cost?
  3. Which headcount is eliminated, avoided, retrained, or redeployed?
  4. What throughput, cycle-time, and quality target must the system reach?
  5. When does the initiative stop if it misses those targets?

Accordion’s own framework makes this point. It recommends connecting time saved per task, loaded labor rate, task frequency, token usage, acceptance rate, and revenue per employee. That is enough to build a defensible return calculation. The providers supply the usage data. Finance already owns the labor data.

The governance should be equally plain. Finance reconciles the usage and labor data quarterly. The workflow owner signs off on the operating result. The sponsor or board reviews any initiative that misses its target or avoids the promised workforce decision. A pilot that cannot survive that review loses its budget.

What remains is operating judgment.

The CFO’s real job here

The CFO should not be asked to predict which model wins or understand every technical decision. The CFO should insist that every AI dollar attaches to a process, every process has a baseline, every baseline has an owner, and every owner has a target that reaches the P&L. The sponsor and board must own the political decision when that target requires a real change in headcount. Finance cannot measure its way around leadership unwillingness.

Some benefits will show up as lower headcount. Others will show up as avoided hiring, more output per employee, faster client response, fewer errors, or a shorter cash cycle. Not all of them arrive in the same quarter. All of them can be measured.

The companies that do this well will not have the largest collection of pilots. They will have a short list of workflows with known economics, accountable owners, and systems that get better as more work passes through them. Their managers will lead larger spans. Their teams will deliver more at dramatically lower cost. Their headcount plans will move with the technology instead of trailing it by a year.

The companies that do not will keep funding experiments and calling the return unclear.

That is not an AI problem.

It is a management problem.

Sources