The Setup
Every AI sales pitch shows the ROI. Most of them inflate it.
Common tricks:
- Calculating savings on time the team would never actually save (the rep saves 2 hours/week but those hours don't translate to cancelled headcount)
- Hiding the operations cost ($25K build looks great until you find out the bot costs $4K/month to run)
- Cherry-picking a 6-week window where the bot was new and impressive
- Ignoring opportunity cost (the team that's installing AI is the team not doing something else)
Real AI ROI is positive in most well-scoped SMB cases. But the math has to be honest.
The Five Inputs
To calculate AI ROI for an SMB use case, you need five inputs:
1. Build cost (one-time)
What the vendor charges, plus any internal time to provide data, do reviews, run UAT.
2. Operations cost (recurring)
- Model provider costs (typically $200-$5,000/month for SMB scale)
- Hosting (typically negligible at SMB scale, $50-$500/month)
- Vendor retainer for ops (typically $0-$5,000/month depending on scope)
- Internal time to maintain (typically 2-5 hours/month at $25-50/hour loaded)
3. Savings (recurring)
Three honest categories:
- Time saved that translates to fewer hours billed/paid (cancelled headcount, contractor reduction, overtime elimination)
- Revenue lifted (faster response → higher conversion, deflection → ability to serve more without hiring)
- Error reduction (fewer mistakes that cost money to fix)
The trap: counting "time saved" that doesn't actually become dollars. If your team has unused capacity, freeing up 2 hours/week doesn't save money. It just gives them more slack.
4. Adoption ramp
AI doesn't hit full savings on day one. Realistic ramp:
- Month 1: 20-30% of full savings (still configuring, edge cases discovered)
- Month 2-3: 60-80% (system tuned, team trained)
- Month 4+: 90-100%
5. Risk factors
- Probability of build delays (add 20% to timeline as buffer)
- Probability of adoption failure (10-25% for well-scoped projects, 50%+ for over-scoped)
- Probability of vendor flake (varies by vendor)
A Real Example: AI Receptionist for a Dental Practice
Build: $18,000 one-time. 5 weeks.
Operations cost:
- Model provider: $400/month
- Vendor retainer: $2,500/month
- Internal review: 1 hour/month at $40 loaded = $40
- Total ops: ~$2,940/month
Savings:
- Front desk hours reduced from 2.0 FTE to 1.5 FTE = $30,000/year saved in loaded cost ($2,500/month)
- After-hours call capture (previously zero) adds an estimated 8 patients/month at $400 LTV = $3,200/month additional revenue
- Total savings: $5,700/month at full ramp
Net at full ramp: $5,700 - $2,940 = $2,760/month net positive
Cumulative cash flow:
- Month 0: -$18,000 (build cost)
- Month 1: -$18,000 + $1,000 (30% ramp) - $2,940 = -$19,940
- Month 2: -$19,940 + $3,800 (70% ramp) - $2,940 = -$19,080
- Month 3: -$19,080 + $5,100 (90%) - $2,940 = -$16,920
- Month 4-12: full ramp +$2,760/month × 9 = +$24,840
- End of year 1: +$7,920 positive
Payback window: Month 11.
Year 2: +$33,120 positive (clean year, no build cost).
This is honest ROI. Year 1 is mildly positive after accounting for build cost. Year 2 and beyond is where the win compounds.
A Bad ROI Example (Just To Show The Pattern)
Same dental practice, but the AI vendor sold them on a generic SaaS chatbot at $99/month.
Build: $0 setup (it's a SaaS product). Operations: $99/month.
Reality:
- Bot can't actually book appointments (no real integration with their scheduling software)
- Bot can answer hours and location, nothing else
- Front desk staff still answer all real calls
- After-hours capture: still zero
- Savings: ~$0/month
Net: -$99/month forever. They cancel after 4 months. Bot was a tax, not an investment.
This is the trap. Cheap AI tooling that doesn't actually solve the workflow is more expensive than custom-built infrastructure that solves it.
How To Evaluate Any Vendor's ROI Claim
When a vendor presents ROI:
- Demand the model. A spreadsheet with the inputs and assumptions. If they don't have one, the ROI is marketing.
- Stress-test the savings. If they claim "$5,700/month in time savings," ask what specific person becomes part-time, contractor goes away, or hiring is paused. If the answer is hand-wavy, the savings aren't real.
- Add the operations cost. Vendors sometimes "forget" to include this. Always ask for the all-in monthly cost.
- Apply a realistic ramp. No project hits full savings in month 1.
- Calculate cumulative cash flow. Most SMBs care more about "when does this go positive" than "what's the year-3 IRR."
The Honest Read
For most well-scoped SMB AI projects:
- Payback window: 4-12 months
- Year 1 ROI: 50-150%
- Year 2 ROI: 200-400%
- Year 3+: compounding gains as the system gets refined
For poorly-scoped projects:
- Payback window: never (negative cash flow indefinitely)
- Reason: scope didn't match the actual workflow problem
The first set looks like every AI project. The second set looks like every AI project that failed.
The difference is the scoping. Which is exactly what an AI Readiness Audit is designed to solve.