How Do You Calculate the ROI of AI in a Business?

Last reviewed: 28 October 2026 · Chris Paton

Why most AI ROI calculations are wrong before they start

The single most common mistake I see is treating AI as a drop-in replacement for a process that was already broken. You automate the broken thing, it breaks faster, and the numbers look terrible. I started in lean manufacturing, and the rule there is simple: remove the waste first, then automate what's left. That order matters. It changes what you're measuring and usually halves the apparent cost of the AI itself.

The second mistake is measuring inputs instead of outputs. 'We're using Copilot' is not an ROI figure. 'Our engineers reviewed and shipped 40% more pull requests per week' is the beginning of one. The metric has to connect to something the business cares about: capacity, revenue, cost per unit of output, or customer outcome.

Start with one question: what would you have paid a person to do this, and how long would it have taken? That's your baseline. If you can't answer it, you're not ready to measure ROI, and you should probably start with an AI readiness assessment before spending another pound on tooling.

What goes into the cost side of the calculation?

Most teams undercount costs because they only log the API bill. The real cost stack looks like this:

• Tooling: API costs, platform subscriptions, cloud compute
• Build time: engineering hours to design, prompt, integrate and test the system
• Integration: connecting the AI to your data sources, CRM, CMS or ops stack
• Human review: the time your team spends checking, correcting or approving AI outputs
• Monitoring and maintenance: the system doesn't run itself; prompts drift, APIs change, edge cases appear
• Opportunity cost: the engineering capacity you used to build this instead of something else

For a rough number, take your best engineer's loaded hourly rate, count the hours spent across all of those categories, and add the direct tooling spend. That's your denominator.

The AI ROI Calculation

What goes into the value side?

Value comes from four places. Not every AI system delivers all four, and that's fine. Be honest about which ones apply.

Time saved. The most common and most measurable. Track the time a human spent on the task before the system existed, then track how long the same task takes now, including review time. Multiply the delta by the loaded hourly rate of the person who would have done it. This is where I'd point you to the AI ROI calculator to run the numbers quickly.

Capacity added. Some tasks become possible that weren't before, not because a human got faster but because the volume was never achievable manually. The TrustedVillas sync pipeline handles 63,000+ properties as at 14 October 2026. No human team could keep that data current at that scale. The ROI question there isn't time saved per task; it's what the business can now offer that it couldn't before.

Revenue added. Harder to attribute cleanly but worth tracking. If an AI-driven content pipeline lifts organic traffic and you can trace conversions back to those pages, that delta belongs in the value column. Use your analytics to isolate it. Don't claim a revenue figure you can't trace.

Cost removed. Headcount that didn't need to grow, agency retainers that ended, tools that were retired. These are real savings, but I'd be careful about attributing headcount avoidance as ROI without being clear that the position was genuinely unnecessary, not just unfilled.

A worked example: how I run this across Toro's own businesses

I don't write about AI ROI in the abstract. Toro's scheduled agents run 24/7 as at 14 October 2026 across Campercation, TrustedVillas and Verqos. That's the lab. Here's what the measurement practice looks like in reality.

For the Verqos content engine, the value side of the calculation starts with time. The engine saves hundreds of hours per week as at 14 October 2026. I measure that by tracking how long equivalent content took to produce before the pipeline existed, then comparing it to the current human review and editing time. The cost side includes the API spend, the engineering time I put into building and maintaining the prompts, and the editor hours that still go into each piece. The gap between those two numbers is the ROI.

For Campercation, the live map holds 25,000+ stopovers as at 14 October 2026. Keeping that data accurate would be a data-entry job at meaningful scale. Agents handle the updates. The value isn't time saved on a task I was doing manually; it's a product capability that wouldn't otherwise exist. In cases like this, I think about ROI differently: what would it cost to hire for this function, or what would I lose in product quality if I dropped it?

The honest version of these calculations takes a quarter to stabilise. The first few weeks of a new system are noisy: setup costs are lumped in, the team is still learning the tool, and the baseline numbers are often estimated rather than measured. Don't present week-three figures to a board.

If you want to apply the same structure to your own stack, the Agentic Playbook walks through how we set up these measurement loops before building anything.

How do you track AI ROI over time?

ROI is not a one-time calculation. Costs change as API pricing shifts, as your team spends more time on maintenance, or as the underlying process changes. Value changes too: a content pipeline that drove strong traffic growth in month three might plateau by month nine.

I keep a simple monthly agent log for each system. It records:

• Tasks completed by the agent that month
• Human review hours logged against it
• Direct tooling cost
• Any incidents, failures or prompt changes needed
• The business metric it connects to (traffic, listings synced, content published)

That log is the audit trail. If a system's value drifts down or its cost drifts up, you see it before it becomes a problem. And if someone asks you whether the AI is worth the money, you have an answer that isn't a guess.

What's the minimum viable measurement setup?

You don't need a data warehouse. You need three things before you start building:

1. A baseline. How long does the task take now, and who does it? Write it down before you build anything.
2. A cost tracker. A spreadsheet works. Log tooling spend and engineering hours weekly.
3. A business metric. One number the business already cares about that the AI should move. Not 'AI usage'. A real number.

If you haven't got a baseline, stop. Build the baseline first. An AI readiness assessment is often the fastest way to establish it, because it forces you to document the current process before you touch a single API key.

The companies I see struggling with AI ROI aren't struggling because the technology doesn't work. They're struggling because they skipped this step and then built something they can't measure. The technology is the easy part. The measurement discipline is harder, and it's what determines whether AI pays back or just costs money.

If you want a quick estimate of what an AI system could return for your specific business, run your numbers through the calculator and I'll give you an honest view of what's realistic.

Frequently asked questions

What is the formula for calculating AI ROI?

AI ROI = (Value created, Total cost of AI) ÷ Total cost of AI × 100. Value created includes time saved (measured in hours multiplied by loaded hourly rate), revenue added and capacity gained. Total cost includes tooling, build time, integration, human review, monitoring and ongoing maintenance. Measure over a full quarter minimum for a reliable figure.

How do you measure time saved by AI?

Establish a baseline before you build: log how long the task currently takes and who does it. After the AI system goes live, track the time still spent on the task, including any human review of AI outputs. Multiply the time delta by the loaded hourly rate of the person who would have done the work. Do not estimate; measure it.

Why is it hard to prove AI ROI?

Most teams undercount costs by ignoring build time, maintenance and human review hours, and overcount value by attributing results they can't isolate. Attribution is the core problem: if organic traffic grows at the same time you launch an AI content pipeline, you need analytics that can separate the two. ROI is also a moving target as API costs and team habits change over time.

How long does it take to see ROI from an AI system?

In my experience, the first three to four weeks of a new system are distorted by setup costs, learning curves and initial prompt failures. A reliable ROI picture takes a full quarter. Simple automations replacing a single manual task can show payback sooner; integrated pipelines that touch multiple systems take longer to stabilise and measure accurately.

Should you automate a process before or after measuring it?

Always measure and clean the process first. Automating a broken or wasteful process makes it fail faster and makes the ROI look worse than it should. Remove unnecessary steps, clarify who owns each step and establish your baseline metrics before you write a single line of automation. This is the lean-before-agents principle: fix the process, then automate what remains.

What business metrics should AI ROI connect to?

Pick one metric the business already tracks: cost per unit of output, time to complete a task, volume of work processed, or revenue from a specific channel. 'AI usage' is not a business metric. The metric must be something leadership cares about independently of AI. If you can't name it before you build, you're not ready to measure ROI.