← Back
AI & product development

How to Calculate the ROI of AI Automation Before You Build It

Published on November 20, 2026 • Written by RM JDG team • Updated on November 20, 2026

The simplest honest formula for AI automation ROI is: annual value of the work it removes, minus the full annual cost of building, running and supervising it, divided by that cost. Most bad estimates get the first half wrong (they count hours saved that nobody actually reclaims) and the second half wrong (they forget review time, errors and maintenance).

This guide walks through how to estimate both halves with numbers you can defend, using a worked example, so you can decide whether a workflow is worth automating before you spend money on it.

Start with one workflow, not "AI"

ROI only makes sense for a specific piece of work. "Use AI to save time" cannot be measured. "Automatically triage and draft replies for inbound support emails" can.

Define the workflow in one sentence, then write down:

  • what triggers it and what the output is
  • who does it today and how often
  • what a wrong output costs
  • what "done" looks like

If you cannot describe it that precisely, the project is not ready for an ROI estimate. It is ready for a short discovery pass first. If the candidate is an agent that makes multi-step decisions, AI agent development costs are usually higher and less predictable than a fixed workflow, so price those separately.

Step 1: Measure the current cost of the work

Count what the work costs today, from real data where possible, not from guesses.

  1. Volume. How many times does this happen per month? Pull it from your ticketing system, inbox or CRM instead of asking people to estimate.
  2. Time per item. Sample ten to twenty real items and time them. People routinely overestimate or underestimate their own tasks.
  3. Loaded hourly cost. Use the fully loaded cost of the people doing the work (salary plus overhead), not just a base rate.
  4. Error and rework cost. Count the time spent fixing mistakes, plus any direct cost of errors (refunds, missed leads, delays).

Current monthly cost is roughly: volume × time per item × loaded hourly cost, plus rework.

Step 2: Estimate how much the automation really removes

This is where most ROI cases fall apart. AI rarely automates a task end to end. Usually it automates part of it and leaves a human reviewing the rest.

Split the volume into three buckets:

  • Fully automated. The system handles it and nobody touches it.
  • Assisted. The system drafts or prepares it, and a person reviews and sends it. This saves time per item but not all of it.
  • Escalated. The system cannot handle it and a person does it as before, now with a little extra triage overhead.

You will not know the exact split until you test, so use a range. A conservative case, an expected case and an optimistic case are enough. If you have no data at all, run the workflow by hand on a sample of real items with a model, and record how many outputs you would have been comfortable sending without edits. That sample is worth more than any benchmark.

A useful rule: only count time savings you can actually convert into value. Saving ten minutes across forty people does not necessarily free anyone to do something more valuable. Savings count when they reduce overtime or contractor spend, avoid a hire, shorten a revenue-critical turnaround, or let the same team handle more volume. Be explicit about which of these applies.

Step 3: Price the full cost of the automation

The cost side has more lines than people expect.

Build cost. Design, development, integrations, testing and deployment. If you use a visual tool, the build is cheaper but you still pay for design and testing. If you are unsure which route fits, see Zapier vs Make vs n8n vs custom code.

Run cost. Model tokens, hosting, platform fees, and any third-party APIs. Estimate tokens from real item sizes and the number of model calls per item. Our guide to reducing LLM API costs explains how much the model choice and prompt design move this number.

Human review cost. Every assisted item still has a person in the loop. Multiply the share of reviewed items by review time and loaded cost. This line is often larger than the token bill.

Evaluation and monitoring. Someone has to check that quality has not drifted after a prompt change, a model update or a change in the input data. Budget for an evaluation set and for ongoing monitoring. See how to test and evaluate LLM outputs before shipping.

Maintenance. Integrations break, APIs change and models are retired. A common budgeting approach is a yearly percentage of build cost; our guide to software maintenance costs after launch covers the usual ranges and what drives them.

Cost of errors. If the automation makes a mistake that reaches a customer, what does that cost? Put a number on it, even a rough one, and multiply by an expected error rate from your sample.

Step 4: A worked example (illustrative numbers)

The figures below are made up to show the arithmetic. Replace them with your own.

Suppose a company handles 2,000 inbound support emails a month. Each takes about 6 minutes to read, look up and reply to, at a loaded cost of 30 per hour.

  • Current cost: 2,000 × 0.1 hours × 30 = 6,000 per month, or 72,000 per year.

After testing on a sample, assume the system drafts a reply for every email, and:

  • 30% are good enough to send after a 1-minute glance
  • 50% need a 3-minute edit
  • 20% are escalated and still take the full 6 minutes, plus 1 minute of wasted triage

Time after automation per 2,000 emails:

  • 600 × 1 minute = 600 minutes
  • 1,000 × 3 minutes = 3,000 minutes
  • 400 × 7 minutes = 2,800 minutes
  • Total: 6,400 minutes, about 107 hours, which is about 3,200 per month

Gross saving is about 2,800 per month, or 33,600 per year.

Now subtract costs. Assume 600 per month for model tokens, hosting and platform fees (7,200 per year), 20,000 for the build, and 4,000 per year for maintenance and evaluation upkeep.

  • Year one cost: 20,000 + 7,200 + 4,000 = 31,200
  • Year one net: 33,600 − 31,200 = 2,400
  • Year two cost: 7,200 + 4,000 = 11,200, net: 33,600 − 11,200 = 22,400

The point of the example is the shape of the result: a modest first year and a stronger second year. It also shows how sensitive the answer is to the review share. If the "good enough to send" bucket were 10% instead of 30%, the case would shrink fast. That sensitivity is the thing to test with real samples before you commit.

Step 5: Calculate payback and set a threshold

Two numbers are more useful than a single ROI percentage:

  • Payback period: build cost divided by monthly net saving (monthly saving minus monthly run and upkeep cost). In the example above the monthly net saving is about 2,800 minus 600 for running costs minus roughly 330 for upkeep, or about 1,870, so payback is 20,000 divided by 1,870, a little under 11 months.
  • Downside case: what happens if the automation delivers half of the expected saving? If the project still breaks even within a reasonable window, it is low risk. If it only works in the optimistic case, it is a bet, not an investment. In the example, half the saving is 1,400 per month, which leaves only about 470 after running and upkeep costs. Payback then stretches to more than three years, a sign that the case depends on hitting the expected review share.

Decide your threshold before you look at the numbers. Many teams use a payback target of under a year for operational automation, but the right figure depends on your cost of capital and how reliable the savings are.

Benefits that are real but harder to count

Some value does not fit neatly into hours saved:

  • Faster response times that affect conversion or retention
  • Consistency, such as every lead being followed up within minutes
  • Capacity, handling a spike without hiring
  • Better data, because the system tags and structures everything it touches

Treat these as upside, not as the foundation of the case. If the project only pays off when you include soft benefits, ask whether it is the right first project.

Common ROI mistakes

  • Counting 100% of time as saved. Review and escalation remain.
  • Using vendor or case-study numbers instead of your own sample. Your data, volume and quality bar are different.
  • Forgetting the cost of wrong answers. One bad automated reply can cost more than a month of savings.
  • Pricing the build but not the run and upkeep. The ongoing costs are what turn a good year one into a bad year three.
  • Starting with the most complex workflow. Pick the one with clear volume, clear inputs and a tolerable cost of error.

What to do next

Pick one workflow and run the five steps with real data. If the downside case still works, build a small pilot, measure the real review share and error rate for a few weeks, and update the numbers before scaling up. If you want a second opinion on an estimate, or help scoping a pilot, our team builds production AI automations and can review your numbers and tell you honestly whether the project is worth doing. For a broader view of typical budgets, see how much AI automation costs a small business.

Services

Not sure where to start? Tell me what you want the product to do.

Related work

    How to Calculate the ROI of AI Automation Before You Build It | RM JDG