How to calculate AI ROI before you spend a penny

Published March 23, 2026

This is part of our AI for Small Business series.

McKinsey found that companies with rigorous AI ROI frameworks are 2.5x more likely to scale their AI investments successfully. Most AI projects get approved on vibes. Someone sees a demo, gets excited, and the budget gets signed off with a vague belief that “it’ll pay for itself.” Then six months later, nobody can actually prove whether it did. An AI ROI calculator isn’t a spreadsheet template from a consultancy’s website. It’s a way of thinking about the investment before you make it. Here’s the maths most consultants won’t show you.

Why most ROI conversations are useless

When you ask an AI vendor about ROI, they’ll show you case studies. “Company X saved 40% on operational costs.” Great. What were their operational costs? What did the system cost to build? How long until they broke even? Those details are always missing because the specific numbers would reveal that the ROI story is messier than the headline suggests.

The other problem: ROI projections from the people selling you the system are inherently biased. Of course they think it’ll pay for itself. They’re the ones getting paid to build it.

You need to run the numbers yourself. Before a single call with a vendor.

The simple AI ROI framework

Here’s how I model ROI for every project before we build anything. It’s not complex. It’s four numbers.

Step 1: Calculate the current cost

What does the process you’re automating cost you today? There are three components.

Labour cost

How many people spend how many hours per week on this process? Multiply hours by fully loaded hourly cost (salary + benefits + overhead, divided by working hours). If two people spend 15 hours each per week on invoice processing, and their fully loaded cost is $35/hour, that’s $1,050/week. $54,600/year.

Error cost

What do mistakes in this process cost you? Returned orders, rework, customer complaints, compliance issues. This is harder to quantify but often significant. If errors in order processing cause 2% of orders to need rework, and each rework costs $50, calculate the annual total.

Opportunity cost

What else could those people be doing? If your $80,000/year account manager is spending 40% of their time on admin, that’s $32,000/year of high-value capacity sitting idle. This is the number most people forget, and it’s often the biggest one.

Total current cost = Labour + Errors + Opportunity cost

Step 2: Calculate the AI system cost

What will the system actually cost? Three components here too.

Build cost

The one-time cost of design and implementation. For most SME use cases: $10,000-$30,000. (I’ve broken down the full pricing landscape for AI consulting if you want more detail on what things should cost.)

Running cost

Monthly infrastructure, API usage, hosting. For most systems: $200-$1,000/month. That’s $2,400-$12,000/year.

Maintenance cost

Retainer for ongoing support and optimisation. $2,000-$3,000/month if you want active management. $24,000-$36,000/year.

First-year total cost = Build + Running + Maintenance

Subsequent years = Running + Maintenance only.

Step 3: Calculate the savings

Be conservative here. Assume the AI handles 60-70% of the workload, not 100%. There’ll always be exceptions, edge cases, and supervision required.

Labour savings

If the AI handles 70% of a process that costs $54,600/year in labour, you save $38,220/year.

Error reduction

AI systems are more consistent than humans on repetitive tasks. Assume errors drop by 50-80%. Calculate the saving.

Opportunity gain

Your team now has 70% of their time back for that process. What’s that capacity worth when redirected to revenue-generating work?

Annual savings = Labour savings + Error reduction + Opportunity gain

Step 4: Calculate the payback period

Payback period = Build cost / (Annual savings - Annual running/maintenance costs)

Example using the numbers above: - Build cost: $20,000 - Annual savings: $38,220 (labour alone, before errors and opportunity) - Annual running cost: $6,000 - Annual maintenance: $30,000 - Net annual benefit: $38,220 - $36,000 = $2,220

Wait. That doesn’t look great, does it?

And that’s exactly why you need to run the numbers. If you only count labour savings, and your maintenance retainer is high relative to the saving, the ROI is thin. But add error reduction ($8,000/year) and opportunity cost ($32,000/year of redirected capacity), and the picture changes completely.

Revised net annual benefit: $38,220 + $8,000 + $32,000 - $36,000 = $42,220. Payback period: under 6 months.

The point: you have to count all three savings categories, not just the obvious one.

An AI ROI calculator you can use right now

Here’s a simple version you can run on the back of an envelope.

1. Pick the process you’re considering automating
2. Count the hours per week your team spends on it
3. Multiply by their fully loaded hourly rate
4. That gives you annual labour cost
5. Estimate error costs (even roughly)
6. Estimate what your team could do with the freed time
7. Add those three numbers: that’s your annual benefit ceiling
8. Multiply by 0.7 (assume AI handles 70%, not 100%)
9. Subtract estimated annual running and maintenance costs
10. Divide the build cost by that number

If the payback period is under 6 months, it’s a strong case. Under 12 months, it’s still good. Over 18 months, question whether this is the right use case to start with.

Where the maths gets dishonest

Vendors and consultants inflate ROI projections in predictable ways. Watch for these.

Assuming 100% automation

No AI system handles 100% of a process. There are always edge cases, exceptions, and situations requiring human judgment. Use 60-70% as your estimate.

Ignoring ongoing costs

The build cost is the number everyone focuses on. But running costs, API fees, and maintenance add up. A $15,000 build with a $3,000/month retainer costs $51,000 in year one. Make sure your ROI calculation includes the full picture.

Counting revenue attribution

“This AI system generated $500,000 in new revenue.” Did it? Or did it support a sales process that generated that revenue? Be honest about what the AI actually did versus what would have happened anyway.

Using pre-tax numbers

Cost savings come off operating expenses, which affects profit margin. Make sure you’re comparing like-for-like.

When the ROI doesn’t work

Sometimes you run the numbers and AI isn’t worth it. That’s a valid answer. If a process involves 5 hours of work per week and no meaningful error rate, the annual cost is maybe $10,000. Building an AI system that costs $15,000 plus ongoing maintenance doesn’t make financial sense. I’ve written an honest breakdown of when AI is and isn’t worth it for small business that covers this in more detail.

A good consultant tells you this upfront and steers you toward the use case where the numbers actually work. A bad consultant builds it anyway and bills you for the privilege.

The numbers that matter most

When I do ROI modelling for clients, the numbers that tip the balance are rarely what people expect. Labour savings are the obvious one, but opportunity cost is usually the biggest. Taking $80,000/year talent and freeing them from $20,000/year work is where the real return lives.

Run your own AI ROI calculator before talking to any vendor. Know your numbers. Know your payback threshold. And don’t let anyone sell you a system that doesn’t clear it.

Frequently asked questions

How do I calculate ROI on an AI project?

Four numbers. First, calculate what the process costs you today (labour + errors + opportunity cost). Second, calculate what the AI system costs (build + running + maintenance). Third, estimate savings at 60-70% automation (not 100%). Fourth, divide build cost by net annual savings. If payback is under 12 months, it’s a solid investment.

What ROI should I expect from AI?

Most well-chosen AI projects pay back in 3-12 months. Tier 1 use cases like support automation and document processing can break even in 6-8 weeks. The key variable is how expensive your current process is. A process costing $50,000+/year in labour is a strong candidate. A process costing $10,000/year probably isn’t worth automating with custom AI.

Why do AI vendors overstate ROI projections?

Three common tricks: assuming 100% automation (realistic is 60-70%), ignoring ongoing maintenance and API costs, and attributing revenue to the AI that would’ve happened anyway. Always run your own numbers with conservative estimates before accepting anyone’s projections. The vendor has a financial incentive to make the numbers look good.