The ROI math on AI: how an owner actually measures it

Most “AI ROI” claims are vendor noise. Here’s the math an owner can actually defend to a CFO, a board, or a sponsor — and the discipline that separates real returns from theater.

The framing that works

For each AI workflow, measure three things over a 12-month window:

1. Direct labor hours saved (the easy number)
2. Revenue impact (the harder number)
3. Cost avoidance (the often-real but speculative number)

Then back out the spend: partner retainer + Sprint fees + tool licenses + infrastructure + internal time allocated. The ratio is the ROI.

For mid-market AI work, a defensible target is 3x payback within 12 months. That’s the target we set for the AI Office retainer — the bar it has to clear to pay for itself. It’s a target and a floor, not a contractual guarantee (the 12-month KPI guarantee applies only to Value Sprints with a committed KPI). The math has to support it, not just the marketing.

Direct labor hours saved (the foundation)

The first and easiest piece. Calculate it as:

Hours saved per week × 52 weeks × loaded hourly rate

Loaded hourly rate = (annual salary × 1.4) ÷ 2,000 hours. The 1.4 multiplier covers benefits, taxes, and overhead. For a $100K loaded salary, that’s $70/hour.

Examples:

• AI quote drafting saves the senior estimator 8 hrs/week. $200K loaded × 8 × 52 ÷ 2,000 = $41,600/year.
• AI document processing saves 1 FTE-equivalent. $80K loaded = $80K/year.
• AI customer service handles 40% of routine volume previously handled by 4 reps. 4 × 0.4 × $60K = $96K/year of capacity recovered.

Critical discipline: only count hours that actually get redeployed to value-producing work. If a senior estimator gets 8 hours back and uses them to leave at 4pm instead of 6pm, that’s fine for her quality of life but doesn’t count as ROI. If she uses them to win more complex bids, that’s real.

Revenue impact (the harder one)

For workflows that touch revenue, measure:

Pricing or win-rate change × volume × gross margin × annualization

Notice: gross margin, not revenue. If quote acceleration produces an 8-point win-rate improvement on $5M of inbound bid volume, the revenue impact is $400K. But the ROI impact is the gross margin on that revenue. If gross margin is 25%, the real impact is $100K, not $400K.

This is where AI ROI claims get inflated. Vendors will show you the topline revenue impact and call it ROI. CFOs see through this immediately. Use gross margin and you’ll be defensible.

Cost avoidance (the speculative one)

For workflows that prevent bad outcomes — compliance acceleration, fraud detection, anomaly flagging — measure:

Expected cost of the bad outcome × probability prevented × frequency

This number is real, but discount it heavily. A reasonable rule: if you can’t point to a specific incident in the past 12 months that this workflow would have prevented, the cost avoidance number is theoretical.

Putting it together — a $20M service business example

A $20M industrial services firm runs AI Office Operator for $60K/year and ships two Sprints at $35K each. Year-1 spend: $130K.

Year-1 measured outcomes:

• Quote acceleration: 8-point win-rate improvement, 25% gross margin = $280K of margin captured
• Field-to-office reporting: 1 FTE-equivalent recovered, $80K loaded = $80K of labor recovered
• Senior estimator time: 6 hours/week recovered, $200K loaded, conservatively half used productively = $31K of expert capacity captured

Total Year-1 ROI: ~$391K against $130K spend = 3x payback within 12 months.

This is roughly the math the AI Office payback target is built on — “we pay for ourselves.” Some clients see better; some see worse, depending on execution quality and how disciplined the team is about using recovered time productively. The figures above are illustrative, not a contractual guarantee.

What kills ROI claims when a CFO digs in

1. Hours saved that didn’t get redeployed. “We saved 200 hours” without a story for what those hours produced isn’t real ROI.
2. Revenue counted at the top line instead of gross margin. Inflates the number 3–4x; CFOs see right through it.
3. Cost avoidance with no specific incident behind it. Theoretical numbers don’t hold up.
4. No baseline measurement. If you didn’t measure before, you can’t credibly measure after.
5. One-time gains presented as recurring. A one-time data migration is not a recurring ROI.

What strengthens an ROI claim

1. Baseline measured before the build started.
2. KPI agreed in writing before the build — not retrofitted after.
3. Hours saved tied to specific productive use.
4. Revenue impact counted at gross margin.
5. An auditor or independent reviewer could verify it.

This is what makes AI ROI defensible. It’s also why the 12-month KPI guarantee on Value Sprints exists — to force the discipline of measurement, not just the claim of impact.

Where to go from here

If you’ve got an AI investment running and the ROI math feels soft, that’s a conversation worth having. We’ve helped clients rebuild ROI cases that CFOs and sponsors will actually trust.