The CFO's Guide to AI Investment: How to Evaluate ROI When Vendors Will Not Tell You the Math

If you're a CFO at a mid-market company, you've probably been pulled into an AI conversation in the last six months. Your CEO is excited. Your operations team is interested. Sales wants better forecasting. Customer success wants better churn prediction. Somebody is going to ask you to approve an AI investment soon.

Here's what I've noticed across thirty-plus of these conversations: the vendor pitches are full of capability and empty on financial commitment.

You'll hear "transformative impact" and "exponential ROI" and "step-change in productivity." You'll see slides full of customer logos and case studies that conveniently don't include the actual dollar amounts spent or saved. You'll get a pricing structure that's deliberately vague - "scoped to your needs," "enterprise pricing," "let's talk about it after the audit."

This is on purpose. The vendor doesn't want to commit to a number because their model doesn't survive the math. If you make them put real numbers on the table, half the AI services industry can't take the engagement.

This post is for you, the finance leader who's about to be asked to sign a check. It's the framework I'd use if I were sitting in your seat.

The three categories of AI investment value

Most AI vendor pitches conflate three completely different kinds of value. As a CFO, you need to separate them, because they have different evidence requirements, different timelines, and different levels of defensibility.

Hard cost savings (Year 1, defensible)

This is the only category that should drive an initial AI investment decision. Hard cost savings are dollars you can point to in the P&L. They come from three sources:

FTE displacement or capacity unlocked. The agent does the work that one or more employees were doing. Either you reduce headcount, or you redirect existing headcount to higher-value work without hiring new heads. Calculate this as: hours per week the agent handles × hourly fully-loaded cost × 50 weeks. Be honest about what fraction of the role the agent actually replaces - most agents handle 40-70% of a role's tasks, not 100%.

Error reduction. The agent prevents costly mistakes. Examples: invoices paid twice, contracts renewed at unfavorable terms, leads lost to slow response, returns processed incorrectly. Calculate this as: average cost per error × error rate × volume × reduction percentage. If you don't know your current error rate, your real first AI project is putting tracking in place to measure it.

Cycle time compression. The agent makes things happen faster, which has dollar consequences. Faster lead response means higher conversion. Faster month-end close means earlier visibility into financial reality. Faster invoice processing means earlier payment discounts captured. Each compression has a measurable dollar impact if you do the math honestly.

Strategic value (Year 2 and beyond, harder to defend)

This is the category that gets vendors in trouble, because it's the easiest to over-promise. Strategic value includes:

Decision velocity that lets you act on opportunities competitors miss. Capacity to scale without proportional headcount growth. Better customer experience that improves retention. Ability to take on work you would have declined before.

These are real, but they're hard to measure in advance. Don't let a vendor anchor your investment decision on Year 2 strategic value. The Year 1 hard savings have to justify the deal on their own. Strategic value is the upside.

Vanity value (zero, run away)

These are the "values" that don't show up anywhere measurable. "Innovation positioning." "Future-readiness." "Competitive parity." "Industry leadership."

Vendors who lean on these are usually selling you on the feeling of doing AI rather than the outcomes of it. Any vendor pitch that's heavy on vanity value and light on hard savings should set off alarm bells. You're being sold a logo for your slide deck, not an operational improvement.

The seven questions every vendor should answer before you sign

When a vendor pitches you an AI engagement, here's the question battery that separates real builders from people selling vapor.

1. What is the projected dollar payback in Year 1, broken down by source?

If they can't itemize the savings - hours saved by role, errors avoided by category, cycle time compression by workflow - they haven't done the work. A real vendor walks you through a spreadsheet during the discovery call. A bad vendor hands you a slide that says "ROI: 5x" with no math behind it.

2. What's the payback period, and what's your commitment if we don't hit it?

Most vendors will project a payback period. Few will commit to it. The right answer is something like "12 months, and if the math doesn't work out, here's the structure we use to make it right." If they have no skin in the game on whether you actually realize the savings, the projection is just marketing.

3. What does the engagement actually cost in Year 1, including hidden costs?

The headline price is the build fee. The hidden costs are: integration work your IT team has to do, change management work your operations team has to do, model API usage fees, infrastructure hosting fees, ongoing optimization fees, and the cost of the production owner you'll need to designate. Get all of these on the table before you sign. Many vendors quote a $50K build and forget to mention the $8K/month in operational fees the agent will rack up on top.

4. What happens to the cost structure in Year 2 and Year 3?

This is where vendors often hide the meat. The Year 1 build fee is one-time. Year 2 and 3 are recurring - monitoring, optimization, model fees, support. Get the multi-year total cost of ownership before you commit. A "cheap" Year 1 engagement that turns into $200K/year in Year 2 isn't actually cheap.

5. What metrics will we track to know if it's working, and who's accountable for them?

Real vendors define success metrics before the build, not after. They tell you what they're going to measure, what good looks like, who reports on it, and how often. Bad vendors hand-wave on metrics until the engagement is over, then declare victory based on whatever happened to look good.

6. What does the exit look like if we want to stop?

This question terrifies bad vendors. The right answer is: you keep the agent, you keep the code, you keep the data, you keep the integrations, you stop paying us. Bad vendors lock you in - proprietary platforms, hosted-only architectures, "managed services" you can't unwind. If the answer to "what if we want to leave" is anything other than "you take everything with you," you're being trapped.

7. Show me three references at companies my size, doing similar work, that I can call directly.

Not testimonial quotes. Not case study PDFs. Phone numbers. References at companies of similar revenue, in roughly similar industries, doing roughly similar AI work, who are willing to take a 30-minute call from you. If a vendor can't produce three of these, they haven't done this work enough times for you to be their guinea pig.

The math you should do before any vendor pitches you

Before you take a single vendor meeting, do this exercise. It will save you ten hours of bad sales calls.

Step one: Pick the five workflows in your business that are most expensive, most repetitive, and most dependent on a single person showing up. Rank them.

Step two: For each one, calculate the all-in annual cost: salary of the people involved, errors that workflow produces, opportunity cost of slow execution, and whatever else is real for your business. Be honest. Most workflows cost more than you think because the costs are spread across multiple line items.

Step three: For the top one or two, calculate what 60% improvement would be worth annually. This is roughly what a well-built AI agent should deliver.

Step four: That number is your maximum reasonable Year 1 investment. If a vendor's price exceeds it, the math doesn't work. If their price is way below it, ask why - they may be underestimating the work or planning to make the money back in Year 2 fees.

This exercise puts you in control of the conversation. You're no longer evaluating vendor pitches against each other. You're evaluating them against the actual financial reality of your business.

What "good" looks like in a vendor proposal

A vendor proposal that should make you feel comfortable signing has these elements:

Itemized cost: build fee, operational fees, expected model usage costs, integration costs, all on one page. Itemized value: hours saved per role, errors avoided per category, cycle time compression per workflow, with dollar figures attached to each. Payback projection with a commitment: X months, and if we miss, here's the remediation. Multi-year total cost of ownership, not just Year 1. Defined success metrics with reporting cadence. Named production owner (yours and theirs). Reference list with phone numbers. Exit clause that lets you take everything with you.

If a vendor proposal doesn't have all of these, push back until it does. The good vendors will appreciate that you're treating this like a real investment. The bad vendors will get frustrated and walk away - which is also a useful filter.

The honest cost-benefit framing

Here's the version I'd give my own CFO if I were the operator buying this.

A well-scoped AI engagement at the mid-market level is going to run somewhere between a meaningful capital expenditure and a small acquisition. It's not cheap, but it's not enterprise-software-priced either. The right build pays back in 9-14 months on Year 1 hard savings alone, with strategic upside in Year 2 and beyond.

The wrong build doesn't pay back at all, because it never makes it to production. That's not a financial problem - it's a vendor selection problem. The 73% of AI projects that fail aren't failing because the math was bad. They're failing because the vendor couldn't ship.

Your job as CFO is to make sure both the math and the vendor are right. The math is in the framework above. The vendor selection is in the seven questions. Run both, ruthlessly, and the AI investment becomes a normal capital allocation decision instead of a leap of faith.

The category needs more CFOs asking these questions. Right now, most AI deals are being signed on enthusiasm. The good vendors will welcome the rigor. The bad ones will avoid you. That's the outcome you want.