The CFO's Guide to AI Agent Investment (Worksheet Included)

A CFO’s guide to AI agent investment comes down to one rule: fund evidence, not ambition. Release money in tranches, and make each tranche buy an answer you can check.

Costs run wider than the model bill. Benefits run narrower than the vendor deck. Below are the funding model, the accounting, the board questions, and a fill-in worksheet.

Why AI agent investment needs gates, not a budget line

The base rates are poor. Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. In IBM’s May 2025 study of 2,000 CEOs, only 25% of AI initiatives had delivered expected ROI. McKinsey’s 2026 State of AI survey, fielded May to June 2026 with 1,719 respondents, found 37% attribute any EBIT impact to AI, about the same as a year earlier. Only about 6% are high performers, who attribute 5% or more of EBIT to AI and call its value significant.

The mechanism matters more than the numbers. An agent’s accuracy on your data, your exceptions, and your systems is unknown until someone tests it. So the biggest risk in the investment is a fact nobody has yet. A pilot buys that fact cheaply. A full rollout buys it expensively.

Treat each tranche as an option payment. You pay a little to learn whether the next, larger payment is worth making. IBM also found that 64% of CEOs say the risk of falling behind pushes them to invest before they understand the value. Gates keep that from becoming the default.

Size tranche 1 so a kill is a routine decision, not a board conversation. Why AI pilots fail covers the usual scoping mistakes, and how we ship pilots in one week shows how small a first tranche can be.

The full cost stack your budget has to carry

Most agent business cases price only the build and the model. Here is the full stack I’d require.

Build and integration. Prompts, tool connections, permissions, and error handling for every system the agent touches. Integration is the line that grows most once real data shows up. See the real cost of building AI agents in-house.

Model inference. Billed per token, a chunk of text about the size of a short word. As of October 15, 2026, OpenAI’s default model, GPT-6.1 Sol, costs $2 per million input tokens and $10 per million output tokens on its pricing page, with cache reads at $0.10 and cache writes at $2.50. The flagship, GPT-6 Astra, is $10 and $50. On Anthropic’s pricing page, Claude Opus 5.5 is $4 and $20, and Claude Haiku 5.5 is $0.10 and $0.50 for prompts up to 100,000 tokens. Agents loop (plan, call a tool, read the result, decide again), so they burn more tokens than chat. Anthropic reports its agents use about 4 times the tokens of chat, and multi-agent systems about 15 times. In McKinsey’s 2026 survey, one in five respondents said operating costs, including tokens, constrain their AI use.

Evaluation. A test set of real cases with known answers, rerun whenever the prompt, the tools, or the model version changes.

Monitoring and maintenance. Reading failures, tracking cost per case, and retesting after vendor model changes. This line never reaches zero.

Human review. Price it at a loaded rate. BLS data for June 2026 puts private-industry employer compensation at $46.89 per hour worked, with wages at 70.0% of that. Divide a role’s hourly wage by 0.70 for a rough loaded rate.

Security and controls. Access reviews, audit logs, and approvals for any action that moves money. The enterprise AI security checklist lists what to ask.

The failure mode to catch early: a business case where tokens are the only run-cost line. In the worksheet below, tokens are 8% of monthly run cost. The monthly run-cost model breaks out the lines people forget.

How to value the benefit without inflating it

Agent benefits come in four forms. Only two are reliably cash.

Cash saved. Spend that stops: outsourcing, temp staff, overtime, agency fees. MIT NANDA’s The GenAI Divide report (July 2025) found top performers’ savings came from cutting BPO and agency spend, including $2 million to $10 million a year of eliminated BPO in customer service and document processing. Those gains came “without material workforce reduction.”

Error cost avoided. Price each error today (write-offs, rework, wrong credits). Multiply by the error-rate change the pilot measured, not the one a vendor promised.

Capacity freed. Hours the team no longer spends on the work. They become cash only if you avoid a planned hire, cut contractor hours, or the owner commits them to work with its own return. Otherwise the hours dissolve into the week and the P&L doesn’t move.

Cycle time. Speed pays only through a mechanism you can name. In receivables, that mechanism is working capital. Hypothetical: cutting days sales outstanding (DSO) by 3 days on $60 million of annual revenue releases about $493,000 once ($60M / 365 x 3). At an 8% cost of capital, that is worth about $39,000 a year. Real, but smaller than most speed claims imply.

My rule: build the funding case on cash saved and error cost avoided. Show capacity on its own line, with a named owner and a plan for the hours. If the investment clears your hurdle rate only when capacity counts as cash, it is a yellow, not a green.

Opex or capex: how agent spend hits the financial statements

I’m not your auditor, and this area is changing. Confirm the treatment with your audit firm before you set the budget.

Model and platform fees are usually expensed. PwC’s software costs guide (updated November 2025) says service fees in a cloud computing arrangement are generally expensed as incurred. Per-token API bills for a running agent and rented agent platforms fall here. The exception: fees consumed while building an asset that qualifies for capitalization can sometimes be capitalized as part of that asset’s cost.

Hosted platform implementation can be capitalized. Under ASU 2018-15, eligible implementation costs for a cloud service contract are capitalized and amortized over the hosting term, as KPMG summarizes. PwC notes that this amortization sits in the same line as the service fee, not in depreciation and amortization. The cost still runs through operating expense and EBITDA, just spread over the term.

Software you build follows ASC 350-40. Under current guidance, preliminary work (weighing options, proving feasibility) is expensed. Costs directly attributable to coding and testing in the application development stage are capitalized. Training people and converting data are expensed. PwC notes that US GAAP has no guidance specific to AI software.

ASU 2025-06 moves the trigger. The FASB issued it on September 18, 2025. It drops the project stages. Capitalization starts once management has authorized and committed to funding the project and completion is probable. That threshold is not met while significant development uncertainty remains: novel features not yet resolved through coding and testing, or significant performance requirements still being substantially revised. It is effective for annual periods beginning after December 15, 2027, Early adoption is allowed in any interim or annual period whose statements have not yet been issued, applied from the start of that year. As of October 2026 the FASB has not changed those dates. For a calendar-year company, adoption is required on January 1, 2028. A company can still adopt early during 2026, effective January 1, 2026, if its 2026 statements are not yet issued, or adopt as of January 1, 2027. So decide this quarter which rules your 2027 projects follow. In PwC’s Example SW 3-2, a company approves the budget for new AI-enabled features on February 1. It starts capitalizing on July 1, once coding and testing resolve the novel functionality and the performance requirements stop changing substantially.

Here is the useful part. Gated funding and the new standard point the same way. Plan to expense the pilot. A funding gate does not make costs eligible by itself; the accounting criteria still decide. But the gate memo that releases tranche 2, with measured accuracy and settled performance requirements, is strong supporting evidence for when capitalization starts. Write it once, for both audiences.

The worksheet: inputs, formulas, and a worked example

Copy these tables into a spreadsheet. Every example value is hypothetical except the GPT-6.1 Sol prices and the BLS rate.

The example: a B2B company gets 4,000 billing disputes a month. Its team works 3,000 at 18 minutes each. An outsourcing firm (BPO) handles the other 1,000 at $14 each. Today 3% of cases end in a wrong credit costing $150 on average. Outsourced volume gets cut first.

Table A. Inputs

K assumes 5 calls per case at 10,000 input and 800 output tokens each, at GPT-6.1 Sol prices. Half of each call’s input is a shared prefix: the first call writes it to the cache and the next four read it. That is $0.1045, plus 20% for retries. O is 30 engineering hours at $95, plus $600 of tooling and $300 of evaluation runs.

Table B. Monthly costs and benefits

Table C. Decision outputs

How to read it. Cash alone pays back in about 13 months at steady state, before any ramp-up. Counting capacity cuts that by more than half, which is why capacity needs its own owner. Faster published payback benchmarks often measure something else.

Now the sensitivity that matters. At A = 20%, net cash falls to $7,400 a month and payback stretches to about 19 months. Above 25%, the BPO volume is fully covered. Past that point, outsourcing savings stop growing. Cash still rises a little from avoided errors (going from 25% to 35% adds about $225 a month, net of extra review), but most of the gain turns into capacity. In this example the BPO contract sets the cash ceiling, not the agent. Model price barely matters: tripling K cuts net cash by only $1,000 a month.

Questions your board and audit committee will ask

Boards rarely ask about the model. They ask about cash, downside, and control. Have these answers on paper first.

One point audit committees catch that project teams miss. If the agent touches anything that feeds the financial statements, its prompt and model version are part of a control. Changing either one is a change to that control. It needs the same testing and sign-off as any other system change; the change control guide for prompts and models shows the record. Put that in the operating plan before go-live.

Red, yellow, green funding criteria

Score every request against this table before money moves. The thresholds are my defaults, so tune them.

Why 300 cases? At 90% observed accuracy, 300 cases give a 95% confidence interval of about plus or minus 3.4 points. At 100 cases, it is about 5.9 points, which is too loose to plan review staffing against.

The decision rules. Any red means no money beyond tranche 0. One or two yellows means fund only the next tranche, with a cap and a dated re-review. All green means release the next tranche. Never release two tranches at once. Tranche 2 is where the integration surprises live.

Bottom Line

Fund AI agents in tranches, and make each tranche buy a measured answer.

Count every cost at its full rate. Count only cash and error savings in the funding case.

Plan to expense the pilot, and write the gate memo so it also supports the capitalization decision.

Next step: fill in Table A for one workflow, with the workflow owner in the room.

FAQ

Do AI model prices only go down?

No. OpenAI priced GPT-5.2 at $1.75 per million input tokens and $14 output in December 2025, up from $1.25 and $10 for GPT-5.1. The token count can move too: Anthropic says Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text. Keep price and tokens per case as input cells, and run a 2x case.

Is the MIT “95% of organizations get zero return” figure reliable?

Treat it as directional. The research ran January to June 2025 and drew on 52 organization interviews, 153 survey responses from senior leaders, and a review of over 300 public AI initiatives. The authors note that outcomes are self-reported and that success definitions varied. Use it as a reason to gate funding, not as a forecast for your project.

What useful life should we use for capitalized agent software?

There is no AI-specific rule. PwC’s guide lists obsolescence, competition, and rapid changes as factors when reassessing amortization periods for hosted arrangements. Model providers retire versions on their own schedules, so ask your auditors whether a shorter life than a typical ERP system is more defensible.

What should we report if a pilot gets killed?

Report the spend, what it measured, and the threshold it missed. Under a gated model, that is a control working. If pilot costs were expensed as incurred, as they usually are, the P&L hit landed when you spent the money. If any were capitalized, expect an impairment conversation.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect