How to Calculate the ROI of AI Automation

The ROI of AI automation is the value of the time and errors it saves, minus what it costs to set up and run, divided by those costs. The formula is simple; the difficulty is in the inputs. A credible calculation counts the time people still spend reviewing AI output, includes all running costs, uses only the share of cases the AI can really handle, and is honest about whether saved time turns into money. This guide walks through each input, then shows a complete worked example.

To follow along with your own numbers, open the free AI ROI calculator. It uses the same method.

The core formula

For a first-year view:

ROI = (annual benefit − total first-year cost) ÷ total first-year cost

Where:

• Annual benefit = hours saved per year × fully loaded hourly cost (+ other quantified benefits)
• Total first-year cost = one-time setup cost + 12 months of running costs

Two companion figures make the result easier to judge:

• Net monthly benefit = monthly benefit − monthly running costs
• Payback period = setup cost ÷ net monthly benefit

If net monthly benefit is zero or negative, there is no payback, regardless of how impressive the technology looks.

Input 1: Volume

How often does the task occur per month? Use data where you can: ticket counts, number of invoices, emails received, documents processed. Volume is the multiplier for everything else, and a task that happens twenty times a month rarely justifies a custom build.

Input 2: Time per task, before and after

Measure how long the task takes today. Even a rough sample of ten cases timed with a stopwatch beats an estimate from memory.

Then estimate the time with AI, and this is where most calculations go wrong. The time with AI is not the seconds the model needs to respond. It is the total human time per case:

• preparing or triggering the request,
• reading and checking the output,
• correcting mistakes,
• handling the cases where the output is unusable.

A pilot is the best way to measure this. Our guide to AI adoption for small business describes how to set one up with a baseline.

Input 3: Share of cases handled with AI

Rarely will AI handle 100% of cases. Some are unusual, sensitive or need information the system does not have. If 75% of cases go through the AI-assisted route and 25% stay manual, the saving applies only to the 75%.

Input 4: Fully loaded hourly cost

The hourly cost of staff time is more than the hourly wage. A fully loaded figure includes gross salary, employer contributions, benefits and a share of overheads such as office space and equipment, divided by productive hours (not all paid hours). Finance teams often have a standard figure; if not, estimate it once and reuse it consistently.

Input 5: Running costs

Monthly costs of keeping the automation going:

Cost item	Examples
Model usage	API tokens or per-seat subscriptions
Software	Automation platforms, vector databases, hosting
Maintenance	Updating prompts, fixing integrations, handling model changes
Monitoring and quality checks	Sampling outputs, reviewing errors
Support	Answering staff questions, handling exceptions

API costs depend on volume and model choice. Estimate them with the LLM API cost calculator, and read LLM API pricing explained if the token model is new to you.

Input 6: Setup costs

One-time costs to get started:

• Process analysis and design
• Building and integrating the workflow
• Writing and testing prompts
• Data preparation
• Security, privacy and legal review
• Training the team
• Change management time during the transition

Internal staff time counts, even if no invoice is paid.

Worked example

A company wants to automate the first draft of replies to standard customer enquiries. All numbers are assumptions for illustration.

Input	Value
Enquiries per month	300
Minutes per reply today	15
Minutes per reply with AI, including review	5
Share handled with AI	80%
Fully loaded cost per hour	45
Running costs per month	150
Setup cost	4,000

Hours saved per month: 300 × 0.8 × (15 − 5) ÷ 60 = 40 hours

Value per month: 40 × 45 = 1,800

Net monthly benefit: 1,800 − 150 = 1,650

Payback period: 4,000 ÷ 1,650 ≈ 2.4 months

First-year cost: 4,000 + 12 × 150 = 5,800

First-year benefit: 12 × 1,800 = 21,600

First-year ROI: (21,600 − 5,800) ÷ 5,800 ≈ 272%

These are the defaults in the AI ROI calculator, so you can reproduce the result and then replace each value with your own.

Sensitivity: which assumption matters most?

A single ROI figure hides how fragile it is. Change one input at a time and see what happens. With the example above:

Change	Hours saved	Net per month	Payback
Base case	40	1,650	2.4 months
Review takes 9 minutes instead of 5	24	930	4.3 months
Only 50% of cases handled with AI	25	975	4.1 months
Volume is 100 per month, not 300	13.3	450	8.9 months
Setup cost doubles to 8,000	40	1,650	4.8 months

In this example, the business case survives every single change, but volume and review time move the result most. That tells you what to measure carefully in the pilot. In your case, the most sensitive input may be different.

From time saved to money saved

This is the point most ROI calculations skip. Saved hours only turn into financial value if one of these happens:

• Paid hours fall: less overtime, fewer temporary staff, or not filling a vacancy.
• Capacity is redeployed: the team handles more volume without new hires, or spends time on work that produces revenue or prevents losses.
• Speed creates value: faster responses reduce churn or win more deals, and you can measure that.

If none of these apply, the automation still creates capacity, which can matter for workload and quality of work, but it should not be reported as cash savings. Separate "hard" savings from "capacity" benefits in your business case so decision-makers can judge each.

Benefits beyond time

Some benefits are real but harder to quantify:

• Consistency: every reply follows the same standards.
• Fewer errors: if the AI-assisted process catches mistakes humans miss. Measure this rather than assume it.
• Faster turnaround: customers wait less.
• Employee experience: less repetitive work.

Include them qualitatively, or quantify them only where you have data. Avoid inflating the case with optimistic numbers that nobody can verify later.

Costs and risks that reduce ROI

• Errors that slip through. If a wrong answer reaches a customer, the cost of fixing it, and possibly losing the customer, counts against the benefit. Good review design matters; see human-in-the-loop AI.
• Model and price changes. Providers update models and prices. Budget maintenance time for re-testing prompts.
• Usage growth. Successful automations get used more, raising running costs along with benefits.
• Compliance effort. Some use cases trigger obligations under data protection law or the EU AI Act, which add cost.
• Adoption risk. If staff do not trust or use the tool, the saving never materialises.

Common mistakes

• Counting model response time as the new task time. The human review time is the real figure.
• Assuming 100% coverage. Use a realistic share.
• Using hourly wage instead of fully loaded cost. Or the reverse: using a high fully loaded cost for time that will not actually be freed.
• Leaving out internal setup time. Staff hours spent building and testing are a cost.
• No baseline. Without measuring the before, you cannot prove the after.
• One-off calculation. Recalculate after the pilot and again after six months with real data.

A simple template for your business case

1. Task description: what is automated and what stays manual.
2. Baseline: volume, time per task, error rate.
3. Pilot results: time with AI including review, share handled, quality observations.
4. Costs: setup and monthly running costs, with assumptions listed.
5. Results: hours saved, net monthly benefit, payback, first-year ROI.
6. Sensitivity: what happens if the two most uncertain inputs are worse than expected.
7. Hard savings versus capacity: how saved time will be used.
8. Risks and controls: review process, data handling, compliance checks.
9. Decision and next review date.

If you are still looking for the right task to evaluate, start with which business tasks to automate with AI. When you have a candidate, put the numbers into the AI ROI calculator and test the assumptions before anything is built.