Most AI projects do not stall because the model is weak. They stall because nobody can say, in money, what changed. MIT NANDA’s much-quoted study found that roughly 95% of generative AI pilots showed no measurable impact on P&L (Fortune). Critics point out that “no P&L impact” is not the same as failure (methodology critique), and McKinsey finds that only 39% of companies report any EBIT impact from AI at all (McKinsey, The State of AI). Whichever number you prefer, the lesson for an owner or a CFO is the same: if you cannot calculate the return before you build, you will not be able to prove it afterwards.
This guide gives you a method that works for small and mid-sized companies: the formulas, the cost lines people forget, a simple ramp-up model and two worked examples built on real 2026 benchmarks. You can run it in a spreadsheet in an afternoon, and it will tell you whether a project deserves a pilot — or not.
Why AI ROI is different from ordinary software ROI
A classic software purchase has a licence fee and a fairly predictable benefit. AI automation behaves differently in four ways:
- Costs move with volume. Tokens, voice minutes and per-outcome fees grow with every conversation or document. In Germany, 54% of companies already using AI name unpredictable costs as a barrier (Bitkom).
- Automation is never 100%. An agent resolves a share of cases; exceptions still land with people. Your ROI depends on that share, not on the demo.
- There is a ramp-up. Weeks of building, testing in shadow mode and tuning come before the full benefit.
- Benefits are split. Some use cases cut the cost per unit, others recover revenue you are currently losing. The two need different formulas.
Large companies have learned this the hard way: 74% now include a cost review in their AI approval process and 43% run explicit token budgets (KPMG Q3 2026 AI Pulse). Smaller companies need the same discipline, just in a lighter form.
The core formulas
Keep the maths simple and the inputs honest. Five formulas cover almost every SMB use case:
- Monthly net benefit = monthly gross benefit − monthly running cost
- Payback period (months) = one-time cost ÷ monthly net benefit
- First-year ROI = (first-year benefit − first-year cost) ÷ first-year cost
- Cost-side benefit = volume × automation rate × (manual cost per unit − residual cost per automated unit)
- Revenue-side benefit = additional conversions × average order value × gross margin
Two rules make these formulas trustworthy. First, calculate per unit of work — an invoice, a ticket, a call, a lead — never per tool. Second, use the fully loaded cost per unit: salary plus taxes and benefits, software seats, management time and rework, divided by the units actually processed.
Benefit lines worth counting
| Benefit type | How to measure it | Watch out for |
|---|---|---|
| Labour cost avoided | Minutes per unit × fully loaded hourly cost | Hours “freed” but not redeployed are capacity, not cash |
| Outsourcing or BPO avoided | Invoices from your current provider | Notice periods and minimum volumes |
| Revenue recovered | Lost leads or calls × conversion × margin | Use gross margin, never revenue |
| Errors and penalties avoided | Rework rate, late-payment fees, wrong refunds | Needs a measured baseline |
| Speed | Response time, cycle time | Convert to money only via conversion or penalties |
The full cost stack
| Cost line | Typical 2026 range | Source |
|---|---|---|
| Process audit / discovery | $1.5–2k fixed, 1–2 weeks | TheCrunch |
| Single automation (n8n/Make + LLM) | $3–15k | TheCrunch |
| Custom AI agent | $10–50k+ | Lets-Viz |
| Integrations in a RAG project | 40–60% of the budget | ment.tech |
| Platform outcome fees | $0.50–2.00 per resolution | Fin, Macha, Resolve247 |
| Model usage per agent at scale | $50–500 per month | Lets-Viz |
| Voice AI, all-in per minute | $0.09–0.31 | Techsy |
| Support and improvement retainer | $2–8k per month (mid-market) | Lets-Viz |
Do not forget the internal line: the process owner’s time, the weekly quality review, knowledge-base updates and the people who handle exceptions. If you are comparing pricing structures, our guide to outcome-based AI pricing shows how per-resolution and per-conversation fees change the maths.
Step by step: building the business case
- Pick one process and one unit. “Customer service” is not a unit; “an order-status ticket” is.
- Measure a baseline for 4–8 weeks. Volume, handling time, fully loaded cost, error and rework rate and, for revenue cases, loss or conversion rate. Take it from your systems, not from memory.
- Set a realistic automation rate and define “done”. Resolved, booked, posted to the ledger. HFS Research expects the definition of a “resolved” case to become the next battleground in SLAs (HFS), so write yours down now.
- Cost all three layers. One-time, running and internal.
- Model the ramp-up month by month. Build time, shadow mode, partial rollout, full run.
- Run three scenarios. Conservative, base and optimistic. If the conservative case does not pay back in a time frame you accept, stop or narrow the scope.
- Agree the measurement with finance before go-live. Same metric, same source, same period length as the baseline. This is also what separates projects that reach production from those that stall — more in why AI pilots stall.
Worked example 1: invoice processing in the back office
The back office is where MIT’s researchers found the strongest returns (Fortune), and invoice processing has unusually good benchmarks: an average of $10.89 per invoice, versus $2.78 for best-in-class teams with automation (Parseur). HFS reports companies that raised their touchless accounts-payable share from 27% to 71% within six months (HFS).
Assumptions for a mid-sized distributor:
- 2,000 supplier invoices per month at a fully loaded $10.89 each.
- After automation, a residual cost of $3.50 per invoice for exceptions, approvals and spot checks — deliberately above best-in-class.
- One-time build and integration: $25,000, inside the $10–50k custom-agent range.
- Running cost: $1,500 per month for model usage, hosting, monitoring and support.
The naive calculation: gross saving = (10.89 − 3.50) × 2,000 = $14,780 per month; net $13,280; payback 1.9 months; first-year ROI 312%. It looks great — and it is wrong, because it assumes the system works at full volume from day one.
The realistic calculation adds a ramp-up: months 1–2 are the build (no savings), months 3–4 run at 50% of volume in shadow mode and partial rollout, months 5–12 at full volume. Running costs start in month 3.
| Scenario | Residual cost per invoice | Monthly gross saving at full run | First-year net benefit | First-year ROI | Payback from kickoff |
|---|---|---|---|---|---|
| Naive (no ramp-up) | $3.50 | $14,780 | $134,360 | 312% | 1.9 months |
| Conservative | $5.00 | $11,780 | $66,020 | 165% | Month 6 |
| Base | $3.50 | $14,780 | $93,020 | 233% | Month 5 |
| Best-in-class | $2.78 | $16,220 | $105,980 | 265% | Month 5 |
Two conclusions follow. The project is worth doing even in the conservative case. And the honest payback is “month 5 or 6 from kickoff”, not “under two months” — exactly the kind of gap that destroys trust when finance looks back a year later. For the build itself, see our guide to AI document processing.
Worked example 2: missed calls in a home-services business
Revenue-side cases need the other formula. Home-services companies miss 14–27% of inbound calls, 85% of those callers never call back, and a missed call can mean a lost job worth $230–3,000 (Thoughtly). In a published ServiceTitan case, a plumbing company raised its share of booked calls from under 50% to 80% with voice agents (ServiceTitan).
Assumptions for an HVAC and plumbing company:
- 500 inbound calls per month; 20% are missed (100 calls) and 85% of those callers do not call back: 85 lost opportunities.
- The AI receptionist answers them and books 40%: 34 jobs.
- Average job $400, gross margin 50%.
- Setup $4,000; running cost $700 per month for minutes, platform and monitoring.
Incremental gross profit = 34 × $400 × 50% = $6,800 per month. Net of running costs: $6,100. Payback: within the first month after go-live.
Now the trap. Count revenue instead of margin and you report $13,600 a month — double the real effect. And some callers would have been won back anyway by a quick human callback. The clean way to prove the number is a control group: run the agent only after hours or only on overflow for the first month and compare booking rates. For comparison, a live answering service costs $235–1,640 per month and an in-house receptionist $48–63k per year (AIRA). This is the core of what we build in Lotus Libre Voice; the industry details are in voice AI for home services.
Benchmarks to sanity-check your numbers
| Benchmark | Value | Source |
|---|---|---|
| Invoice processing cost | $10.89 average vs $2.78 best-in-class | Parseur |
| AI support, per resolution | Fin $0.99; Zendesk $1.50–2.00; HubSpot $0.50 | Fin, Macha, Resolve247 |
| Salesforce Agentforce | $2 per conversation or $0.10 per action | Salesforce |
| Human ISA vs AI (real estate) | $3–5k vs $99–500 per month | Rework |
| Senior AI developer in-house | about $8,300 per month, fully loaded | Lets-Viz |
| Time to measurable impact in service | 70% of teams within 60 days | Salesforce |
If your model beats these benchmarks by a wide margin, double-check the inputs before you celebrate.
The one-page business case
Whoever approves the budget does not need your spreadsheet; they need one page they can sign. A good AI business case fits on a single sheet with seven lines:
- Process and unit — for example, “supplier invoices, one invoice = one unit”.
- Baseline — volume, fully loaded cost per unit, error rate, measured over a named period.
- Target — the automation rate and the definition of “done”, plus the guardrails that must not get worse (accuracy, customer satisfaction, escalation quality).
- Costs — one-time, monthly running and internal, with the volume assumption behind each usage-based line.
- Three scenarios — first-year net benefit, ROI and payback month for the conservative, base and optimistic case.
- Risks and mitigations — data quality, integration access, adoption, compliance, and what happens if a vendor changes its pricing.
- Decision points — the go/no-go date after the pilot, who signs it, and the metric that decides.
The last line matters most. It turns the ROI calculation from a forecast into a commitment that can be checked. Companies that agree upfront on how success is measured find it much easier to secure the next budget — and to stop a project early when the numbers do not come in, which is also a good outcome.
Typical mistakes
- Treating hours saved as cash. Unless you avoid a hire, cut overtime or outsourcing, or move people to revenue work, it is capacity. Report it separately.
- Ignoring the build and ramp-up period. As example 1 shows, it moves payback by months.
- Forgetting costs that scale with volume. Per-resolution fees, tokens and minutes look small per unit and large per year.
- Using revenue instead of margin in revenue-side cases.
- Skipping the baseline. Vendor ROI calculators are built on assumptions; your own “before” number is the only one that survives an audit.
- Assuming 100% automation. Plan for exceptions and for the people who handle them.
- Trusting how productive it feels. In a METR study, experienced developers using AI were 19% slower while believing they were 20% faster (METR). Measure, do not estimate.
- Leaving out the cost of errors. A wrong refund or an invented promise has a price; include quality checks and an escalation path in the plan.
Conclusion
Calculating the ROI of AI automation is unit economics plus honesty about timing. Choose one process, measure it, cost every layer, model the ramp-up and test three scenarios. If the conservative case still pays back within a year, you have a project worth piloting; if it only works in the optimistic case, narrow the scope or pick another process.
Most engagements in Lotus Libre Agents start exactly here: one process, one metric, a baseline and a payback date that everyone signs. If you would like a second pair of eyes on your numbers, book a free 15-minute audit — we will tell you honestly whether the case holds.