Workflow automation ROI = (benefits - costs) / costs x 100 over a set period, and payback = one-off costs / monthly net benefit. Value time saved at a loaded hourly cost (US private-industry employers averaged $46.89 per hour worked in June 2026, per the BLS), count only time that is actually redeployed, and include build, training, platform, AI model and maintenance costs. In the hypothetical invoice example below, the same n8n workflow returns 535% in year one at 1,000 invoices a month, 71% under pessimistic assumptions, and loses money in year one at 100 invoices a month. Volume and honest time estimates decide the result, not the tool.
Checked on October 1, 2026 against Bureau of Labor Statistics data, the n8n, Zapier, Make and OpenAI pricing pages, and the studies cited below. This update replaces the January 2025 version, whose headline ROI averages, benchmark ranges and first-person case study were not traceable to sources, and whose formula double-counted costs.
The formula
ROI (%) = (total benefits - total costs) / total costs x 100, over a stated period, usually the first year or three years. Forrester's Total Economic Impact method defines it the same way: "ROI is calculated by dividing net benefits (benefits less costs) by costs" (Forrester TEI, appendix). Two companion numbers make a business case easier to judge:
For the model-cost side of the ROI math, see our LLM API pricing comparison.
- Payback period (months) = one-off costs / (monthly benefits - monthly running costs). Iowa State University's Ag Decision Maker describes it as the time "required for the cash flows generated by the investment to repay the cost of the original investment," and notes it ignores the time value of money.
- Net present value for multi-year cases: discount future net benefits back to today. Forrester notes that organizations typically use discount rates between 8% and 16%.
| Benefits | Costs |
|---|---|
| Hours saved x loaded hourly cost x share of that time put to valuable use | Build: internal hours or a contractor |
| Errors avoided x cost to fix each one | Training and process changes |
| Revenue you can attribute with evidence, such as faster lead response | Automation platform subscription or hosting |
| Hiring you can avoid as volume grows | AI model usage (tokens) |
| Maintenance, monitoring and fixes |
Step 1: Measure the process as it runs today
- Volume: items per month, from your systems, not memory.
- Minutes per item: time a sample over a week or two. Estimates of routine work are often wrong.
- Errors: how often something goes wrong and how long it takes to fix.
- Exceptions: the share of items that don't fit the normal path. These usually stay manual after automation.
- Who does it: the role and pay level, for the next step.
Step 2: Put an honest price on an hour
An hour of work costs more than the wage. In the Employer Costs for Employee Compensation release of September 9, 2026, the Bureau of Labor Statistics reported that "total employer compensation costs for private industry workers averaged $46.89 per hour worked in June 2026." Wages were $32.82 (70.0%) and benefits $14.07 (30.0%). As a rough rule for an average benefits mix, divide an hourly wage by 0.7: a $30 wage becomes about $42.86 loaded. Use your own payroll data where you can, since the BLS figure averages every private-sector job.
Time saved only counts if it is used. If the freed hours go to other valuable work, or let you avoid a hire, count them. If they disappear into the day, they aren't savings. The UK Department for Business and Trade's evaluation of Microsoft 365 Copilot illustrates the gap: "The evaluation did not find evidence that time savings have led to improved productivity." Put a "share of time put to use" input in your model and set it below 100% unless you have a plan for those hours.
Step 3: Count every cost
| Cost | Where the number comes from (checked October 1, 2026) |
|---|---|
| n8n Cloud | Starter: EUR 20 a month billed annually for 2,500 executions. Pro: EUR 50 a month billed annually for 10,000. An execution is one run of a whole workflow, however many steps it has. |
| n8n self-hosted | The Community Edition is free under the Sustainable Use License for internal business use; you pay for hosting and your own time to run it. |
| Zapier | Professional: $19.99 a month billed annually, or $29.99 monthly, for 750 tasks. A task is a successfully completed action; triggers don't count. |
| Make | Entry plan from $9 a month for 5,000 credits. Each module action usually uses one credit. |
| AI model | Per token. OpenAI GPT-6 Luna costs $0.10 input and $0.50 output per million tokens; GPT-6.1 Sol $2 and $10. |
| Build, training, maintenance | Your hours at loaded cost, or a contractor's rate. |
Sources: n8n pricing, Zapier pricing, Make pricing and OpenAI pricing. Billing units differ between tools, so price the same workflow in each; our n8n vs Zapier vs Make comparison covers the trade-offs. For AI steps, our ChatGPT API workflows guide shows how to estimate token costs per task.
Step 4: Run the numbers (a hypothetical example)
Everything in this section is hypothetical. It shows the method, not a result you should expect. A finance team keys supplier invoices into its accounting system by hand, and considers an n8n workflow that extracts the fields with an AI model and posts them for a person to check.
| Input | Assumption |
|---|---|
| Volume | 1,000 invoices a month |
| Today | 6 minutes each by hand = 100 hours a month |
| After automation | A person checks each invoice for 1.2 minutes (20 hours), and the 10% that fail are done by hand at 6 minutes (10 hours) = 30 hours a month |
| Hours saved | 70 a month, 840 a year |
| Loaded cost | $46.89 an hour (the BLS private-industry average) |
| Build | 40 hours at an assumed $90 freelance rate = $3,600 |
| Training | 8 staff hours = $375.12 |
| Platform | n8n Cloud Starter (EUR 20 a month billed annually), budgeted at $300 a year to cover currency conversion and tax |
| AI model | GPT-6.1 Sol at 3,000 input and 300 output tokens per invoice = $9.00 a month at standard rates, budgeted at $20 a month for reasoning tokens and retries |
| Maintenance | 3 hours a month = $1,688.04 a year |
Year-one results:
| Line | Amount |
|---|---|
| Benefits: 840 hours x $46.89 | $39,387.60 |
| One-off costs: build + training | $3,975.12 |
| Running costs: platform + model + maintenance | $2,228.04 |
| Total year-one costs | $6,203.16 |
| Net benefit | $33,184.44 |
| ROI: $33,184.44 / $6,203.16 | 535% |
| Payback: $3,975.12 / $3,096.63 net per month | 1.3 months |
Now change the assumptions that matter most:
| Scenario | What changes | Year-one ROI | Payback |
|---|---|---|---|
| Base | As above | 535% | 1.3 months |
| Pessimistic | Only half the saved time is put to use, the build takes 80 hours, maintenance takes 6 hours a month | 71% | 5.8 months |
| Low volume | 100 invoices a month, same setup | -34% | About 25 months |
Three lessons fall out. Volume decides: the same workflow loses money in year one at a tenth of the volume. Honest time estimates decide: halving the value of saved time cuts ROI from 535% to 71%. The tool price barely matters: the platform and model together cost $540 a year here, against $39,387.60 of benefits in the base case. The UK Treasury's Green Book warns of optimism bias, "the demonstrated systematic tendency for practitioners to be over-optimistic about key assumptions in appraisal," so plan on the pessimistic case and treat the base case as the upside.
Copy this calculator into a spreadsheet
INPUTS A1 Items per month A2 Minutes per item today A3 Review minutes per item after automation A4 Exception rate (share still done by hand, 0 to 1) A5 Loaded cost per hour A6 Share of saved time put to valuable use (0 to 1) A7 One-off costs (build + training) A8 Platform cost per month A9 AI model cost per month A10 Maintenance hours per month FORMULAS B1 Hours saved per month = A1*(A2 - A3 - A4*A2)/60 B2 Benefit per month = B1*A5*A6 B3 Running cost per month = A8 + A9 + A10*A5 B4 Net benefit per month = B2 - B3 B5 Year-one ROI (%) = (12*B2 - (A7 + 12*B3)) / (A7 + 12*B3) * 100 B6 Payback (months) = A7 / B4 (only meaningful if B4 > 0)
Add a row for errors avoided (errors per month x minutes to fix x A5) if your process has a measurable error cost, and run three columns: pessimistic, expected and optimistic.
What published research says about returns
Most studies measure AI and automation broadly rather than one workflow, and each has limits worth knowing:
- McKinsey's 2026 State of AI survey of 1,719 respondents found that 37% attribute at least some EBIT impact to AI use, unchanged from 2025, and 6% qualify as high performers, as reported by The Register. Self-reported.
- IBM's 2025 CEO Study of 2,000 CEOs: "only 25% of AI initiatives have delivered expected ROI over the last few years, and only 16% have scaled enterprise wide" (IBM). Self-reported, against the CEOs' own expectations.
- Deloitte's October 2025 survey of 1,854 executives in Europe and the Middle East: "Most respondents reported achieving satisfactory ROI on a typical AI use case within two to four years," compared with 7 to 12 months for typical technology investments (Deloitte).
- Gartner predicted in June 2025 that "over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls" (Gartner). A forecast, not a measurement.
- Forrester's Total Economic Impact study of Microsoft Power Automate (July 2024) modeled a 248% risk-adjusted ROI over three years, with payback in under six months, for a composite organization of 30,000 employees built from interviews at six companies. Microsoft commissioned it, and the composite is modeled, not a real company.
The pattern: commissioned studies of single products report high returns for modeled companies, while broad surveys find many organizations still waiting for AI to pay off. Neither tells you what your process will return, which is why the measured inputs above matter more than any benchmark.
Step 5: Measure after launch
- Keep the baseline. Compare volume, minutes per item, error rate and exception rate before and after, over at least a month of normal work.
- Log real costs: executions or tasks used, token usage from each AI response, and maintenance hours.
- Track where the time went. If nobody can say, the benefit may not be real.
- Re-run the calculator with measured numbers and report the difference from the plan. That record makes your next business case far more credible.
Mistakes that inflate automation ROI
- Using theoretical time savings instead of measured ones.
- Leaving out review time for AI output and the exceptions that stay manual.
- Valuing hours at loaded cost when nobody plans to use them, or at wage alone when they will be used.
- Automating low-volume processes because they are annoying rather than costly.
- Forgetting maintenance: APIs change, AI models are retired and prompts need re-testing.
- Presenting only the best case. Lead with payback under pessimistic assumptions and propose a small pilot.
Next steps
Pick one high-volume, rule-heavy process, time it for two weeks, and run the calculator with a pessimistic column before you build. To build the workflow itself, see our guides to connecting OpenAI to n8n, self-hosting n8n and AI agent automation. If you would rather build your own AI product than deliver automation projects for clients, AI SaaS Builder covers validating an idea before you build, choosing a pricing model, Stripe billing, and tracking monthly recurring revenue, churn and lifetime value.
Want the full AI SaaS Builder playbook?
A 10-module, 52-lesson curriculum: validate an idea, build on Supabase and Next.js, add AI features with the Claude API, speed up with Claude Code and MCP, deploy on Vercel, launch, and charge with Stripe.
Workflow automation ROI FAQ
How do you calculate workflow automation ROI?
ROI = (total benefits - total costs) / total costs x 100, over a stated period such as the first year. Benefits are usually hours saved x loaded hourly cost x the share of that time put to other valuable work, plus the cost of errors avoided. Costs include building, training, the automation platform, AI model usage and ongoing maintenance.
What is a good ROI for automation?
There is no universal benchmark. Compare the result with your other uses of the money and a pessimistic version of your own numbers. Published figures vary widely: a Microsoft-commissioned Forrester study modeled a 248% three-year ROI for Power Automate at a composite organization, while most respondents to Deloitte's 2025 survey reported satisfactory ROI on a typical AI use case only within two to four years.
How long should an automation take to pay back?
Payback in months = one-off costs / (monthly benefits - monthly running costs). It depends mostly on volume: in this article's hypothetical example, the same invoice workflow pays back in 1.3 months at 1,000 invoices a month, 5.8 months under pessimistic assumptions, and about 25 months at 100 invoices a month.
What hourly rate should I use for time saved?
Use a loaded cost that includes benefits, not just the wage. In June 2026, US private-industry employers paid an average of $46.89 per hour worked in total compensation, with wages at 70.0% and benefits at 30.0%, according to the Bureau of Labor Statistics. Use your own payroll data where you can, and count only time that is actually put to use.
What costs do people forget in automation business cases?
Review time for AI output, exceptions that still need a person, maintenance and monitoring, training, platform upgrades as volume grows, and the work of re-testing when an AI model or API is retired.
Why do automation projects miss their expected ROI?
Common causes are overestimated time savings, low volume, time saved that is never redeployed, and underestimated running costs. For agentic AI specifically, Gartner predicted in June 2025 that over 40% of projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.