How to calculate the ROI of AI automation before you buy

What does ROI mean for AI automation?

AI automation earns its return by replacing repetitive manual work with reliable software: bookkeeping entries, support replies, data entry, and the copying of details from one system into another. ROI asks whether the value of that change (time returned to your team, revenue recovered, errors avoided) is greater than everything the system costs, and how quickly.

Two numbers answer it:

• ROI = (total gains − total costs) ÷ total costs × 100, with both measured over the same period.
• Payback period = one-off cost ÷ monthly net gain, in months.

We prefer payback as the headline number. "This pays for itself in 11 months" is a decision you can make; "71% three-year ROI" needs a spreadsheet to trust. Report both. The glossary entry on automation ROI has the short version; this guide is the full method.

How do you measure your baseline before you buy?

Measure the process as it runs today, for two to four weeks, before anyone builds anything. Without a baseline you can't estimate the gain, and after launch you can't prove it.

For each process you're considering, record:

1. Volume. How many times it runs per week: calls, tickets, invoices, orders, bookings.
2. Minutes per run. Time a sample of 20 real runs with a stopwatch, including switching between tools. Don't estimate from memory.
3. Who does it. The role and its hourly pay, so you can work out loaded cost.
4. Errors. How often it goes wrong (a wrong invoice, a lost lead, a double booking) and what each mistake costs to fix.
5. Customer outcomes. Where relevant: speed to lead, answer rate, conversion rate, show rate.

Most of it is already in systems you use:

If you haven't picked which processes to measure yet, our 40 workflow automation examples include a scoring table for choosing.

How do you value the time saved?

Multiply the hours saved by the loaded hourly cost of the people who did the work, then count only the share of that time that turns into something useful.

Hours saved per month = runs per month × minutes saved per run ÷ 60 × share the system handles.

Loaded hourly cost is pay plus what the employer spends on top of it. The US Bureau of Labor Statistics' Employer Costs for Employee Compensation survey put private-industry costs at $46.89 per hour worked in June 2026: $32.82 in wages and salaries (70.0%) and $14.07 in benefits (30.0%). In establishments with 1 to 49 workers, the average was $37.64, of which $9.76 (25.9%) was benefits.

That gives a quick rule: multiply an hourly wage by about 1.35 using the figures for establishments with 1 to 49 workers, or about 1.43 using the private-industry average (total cost ÷ wages). Better still, divide your own annual payroll cost for the role, including employer taxes and benefits, by the hours actually worked. The BLS figures cover compensation only, not overhead such as office space, equipment or software, so they are a conservative floor.

Realization rate. Time saved is not cash saved unless it goes somewhere. If an office manager saves 20 hours a month and spends them on faster quotes or more sales calls, that time has value. If it disappears into a slower pace, it doesn't. Count freed hours at full value only when they avoid a hire, overtime or contractor costs. Otherwise apply a realization rate, such as 50% to 60%, and label the result "capacity" rather than "savings".

How do you estimate revenue effects conservatively?

Count only the extra gross profit you can trace to the automation, use your own baseline conversion rates, and take the low end of any uplift you expect.

Extra gross profit per month = affected volume × conservative uplift in conversion × gross profit per sale.

Typical sources of revenue effects, across different businesses:

• Faster responses. A real estate team replying to portal leads within a minute instead of hours, or an online store answering pre-purchase questions in chat. Our article on speed to lead covers what the research does and doesn't show, including its age and limits.
• Answered calls. A law firm or a home services company answering after-hours calls instead of sending them to voicemail. Concierge, Worcia's AI receptionist for calls and chat, is built for this; see how Concierge works.
• Fewer no-shows. A clinic, salon or tutoring center using reminders and easy rebooking. See how to reduce appointment no-shows.
• Consistent follow-up. Agencies and contractors following up every quote instead of some. Cadenza, Worcia's sales follow-up automation, runs these sequences.

Three rules keep this honest. Use gross profit, not revenue, because extra sales also bring extra costs. Use your own numbers, not a vendor's average. And if you can't attribute a sale to the automation lead by lead in your CRM, don't count it.

What costs should you include?

Include every cost of building, running and owning the system, one-off and recurring, plus the time your own team will spend on it.

Usage costs depend on volume and the type of work. Language models are billed by the token. Anthropic's pricing page, for example, listed its Sonnet 5.5 model at $2 per million input tokens and $10 per million output tokens (checked October 2026). Voice adds telephony: Twilio listed inbound calls to a US local number at $0.0085 a minute, plus $1.15 a month for the number, and outbound calls at $0.0140 a minute (checked October 2026). Speech recognition, synthetic voice and platform fees come on top and vary by provider. Estimate usage from a pilot with real data, then test a case where it's 10 times higher.

Maintenance is the easiest line to leave out. Automations depend on other tools, and those tools change their forms, fields and APIs. Budget for someone to fix and improve the system for as long as it runs.

How do you calculate payback and ROI?

Fill in this worksheet for each process, one line at a time. It is the same calculation behind our Lab calculators, with a realization rate and error savings added.

A printable version is in the Toolkit.

How sensitive is your ROI to your assumptions?

Change one assumption at a time and watch the payback period. If a small change breaks the case, the case is weak.

Two lessons from this example. First, model usage barely moves the answer; adoption and build cost do. Second, the downside case still pays back, but in nearly three years, so the decision depends on how long the business expects the process to stay the same. That is the conversation to have before you buy.

It also helps to ask the question the other way round: what has to be true for this to pay back within, say, 24 months? In this example, even if the AI resolves only 20% of eligible tickets, payback is 23.3 months, because the bookkeeping and data-entry automations, which need no AI, carry most of the case. When the rules-based parts justify the project on their own, the AI part is upside rather than risk.

To run your own numbers quickly, use the calculator below. Enter only the hours you expect to redeploy, and enter gross profit, not revenue, as "revenue recovered".

What should you compare the automation against?

Compare it with the realistic alternative, not just with doing nothing. A project can beat the status quo and still lose to a cheaper option.

• Doing nothing. The baseline continues: the same hours, errors and missed revenue, growing with volume.
• Hiring. If the real alternative is another person, compare the system's full cost with that person's loaded cost plus recruiting and training time, and the hours they can actually cover. A person brings judgment and flexibility; an automation covers nights and weekends and handles volume spikes without overtime.
• A simpler fix. A template, a booking link or a built-in rule in a tool you already pay for sometimes removes most of the waste for almost nothing. Do that first and measure what's left.
• A rules-only automation. Rules are cheaper to run and easier to test than AI. Use AI only in the steps that need reading or writing free text, and price both designs.

The automation is worth buying only if it beats the best of these.

How do you verify ROI after launch?

Agree the metrics, definitions and targets before launch, measure them the same way afterwards, and recalculate the worksheet with actual figures at set points.

1. Lock the definitions. Write down how each metric is counted, from which system, before launch. A new way of counting is not an improvement.
2. Track weekly for the first 90 days. Volume handled, share resolved without a person, time spent on exceptions, error rate, complaints, and attributed revenue.
3. Compare like with like. Seasonality can fake a win. Where you can, keep a comparison group, such as one location, channel or team staying on the old process for a few weeks.
4. Audit quality. Read a sample of automated replies and entries every week. Look for wrong answers (AI hallucination), wrong postings and customers who gave up.
5. Recalculate at 3, 6 and 12 months. Replace every assumption with the actual number. Then extend, fix or stop.

If the actual numbers fall short, check three things before you give up: adoption (are people working around the system?), exceptions (are too many cases falling back to a person because the rules or the knowledge base are too narrow?), and scope (did the build miss the real bottleneck?). Each has a different fix, and the weekly data will show which one you have.

Our results page sets out how Worcia defines and measures outcomes, including the rule that revenue we can't attribute lead by lead isn't counted.

How do you spot an inflated ROI claim?

Be wary of any projection that does one of these:

• Counts revenue instead of gross profit.
• Treats every freed hour as cash.
• Leaves out maintenance, usage or your own staff time.
• Uses another company's average instead of your baseline.
• Calculates payback without the one-off build cost.
• Assumes the best-case resolution rate from day one.
• Has no downside scenario.

If you'd like a second pair of eyes on your numbers, see how Maestro scopes automation against a measured baseline and agreed targets.