The AI ROI Question: How to Think About Returns Before You Build Anything

The CFO asks every CEO the same question before the AI project starts:

“What’s the ROI?”

And the CEO gives the same answer every CEO gives:

“Uh… we’ll know once we’ve built it.”

Wrong answer. Wrong answer twice over. The CFO is right to ask, and you should have an answer before the engineering team writes the first line of code.

Here’s how operators actually estimate AI ROI — without pretending to certainty they don’t have, and without hand-waving past the question.

Step 1: Stop Calculating “AI ROI”

The phrase is a trap. AI isn’t an investment category. It’s an implementation method for solving specific business problems. So you don’t calculate “AI ROI” — you calculate the ROI of the specific problem you’re solving, and AI is one of several ways you might solve it.

Reframe the question this way:

“If we solved problem X — by any method — what would it be worth to the business?”

That’s the number. Now you can talk about whether AI is the right method, but the value of the outcome is independent of the implementation.

Step 2: Pick Your Lever

Every AI project, when you strip the branding, is pulling one of five levers. Be honest about which one you’re pulling — and demand a number on it before you start.

Lever 1 — Labor leverage. Same output, fewer hours.

• Math: hours saved per week × fully-loaded hourly cost × 50 weeks
• Honest version: Account for hours that get reabsorbed into other work. Real savings are usually 60–70% of theoretical.

Lever 2 — Throughput. More output, same headcount.

• Math: incremental units shipped × gross margin per unit
• Honest version: Throughput only matters if you’re capacity-constrained. If demand isn’t there, this lever is zero.

Lever 3 — Quality. Fewer errors, fewer returns, fewer rework cycles.

• Math: current defect rate × cost per defect × volume
• Honest version: You probably don’t know your real defect rate. Estimate, but commit to measuring after.

Lever 4 — Speed-to-customer. Faster quote, faster onboarding, faster decision.

• Math: deals won that would have been lost × average deal size × win-rate lift
• Honest version: The hardest to measure but often the largest. If your competitors quote in two hours and you quote in three days, you’re losing deals you never see.

Lever 5 — Risk avoidance. Fewer compliance failures, safety incidents, errors with regulatory consequence.

• Math: probability of incident × cost of incident × reduction factor
• Honest version: This is real, but only if you’ve actually had incidents. Don’t multiply a hypothetical incident by the cost of an outcome that’s never happened.

Pick one lever per project. Stacking levers in the same justification is how strategy decks get a 4x ROI when reality returns 0.8x.

Step 3: Build the Honest Cost Stack

The ROI denominator is usually wrong because companies count only build cost. Real cost is:

• Build cost: Engineering, vendor fees, infrastructure.
• Integration cost: Hooking the AI into existing systems is often 1.5–2x the build cost.
• Change management cost: Training, documentation, redesign of workflows, the slow productivity dip while people learn the new system.
• Operating cost: Inference cost (this can be material at volume), monitoring, evals, ongoing maintenance.
• Opportunity cost: What else could your team be building right now?

A project budgeted at $80K usually costs $140K to actually deliver. Estimate accordingly.

Step 4: Apply the Realism Discount

Every AI ROI estimate I’ve ever made or seen is too optimistic. Always. So we apply a discount factor:

• Best case = your optimistic estimate (the one in the deck)
• Expected case = best case × 0.6
• Worst case = best case × 0.3

Make the project decision on the expected case, not the best case. If the expected case is still worth doing, you have a project. If it isn’t, you have a research idea — fund it as research, with a research budget, not as production capital.

Step 5: Pay Back Inside 12 Months, or Don’t Build It

For mid-market businesses, the rule I use is brutal but right:

A first AI project should pay back inside 12 months. Year-one ROI should be 3x or better. Year-two ROI should be 8x+ as the system compounds.

If the math doesn’t work at those bars, either:

• The project is too small (find a bigger problem)
• The project is too speculative (treat it as R&D, not production)
• The cost stack is bloated (you’re buying too much vendor)
• The lever is fake (revisit your assumptions)

Don’t bend the assumptions to make the math work. The CFO will catch it. So will reality.

Step 6: Pre-Commit to the Measurement

Here’s the part most companies skip: write down, before you start, exactly how you’ll measure whether the ROI hit.

• “Quote turnaround time” — measured from inbound RFQ timestamp to quote-sent timestamp, weekly average.
• “Engineer hours per quote” — measured by self-reported timesheet, sampled.
• “Quote conversion rate” — won quotes / total quotes, by month.

Pick three metrics. Establish baseline before launch. Measure monthly after. Report quarterly. No moving the goalposts.

If you don’t pre-commit to measurement, you’ll claim victory because the AI launched — not because it produced value. That’s how organizations end up with five “successful” AI projects and no profit improvement.

The Operator’s Real Test

The single best ROI filter is this question:

“If this AI project produced zero value six months in — would I be embarrassed in front of the board, or would I just shrug?”

If you’d shrug, don’t build it. The fact that you’d shrug means it isn’t important enough to fight for when it gets hard. And AI projects always get hard.

If you’d be embarrassed, you have the right project. You’ll fight for it. You’ll allocate the budget, defend the timeline, and ship.

That emotional test is more accurate than any spreadsheet. Spreadsheets lie. The pit-in-your-stomach test doesn’t.

---

Want help running this exercise on a real project? It’s exactly what the AI Roadmap Session produces — written, defensible, board-ready. Book one here.