Most businesses that come to us have already decided they want AI automation. What they haven't done is work out whether the numbers actually stack up. And that gap, between wanting automation and knowing it'll pay off, is where a lot of money gets wasted.
This post walks through how to calculate AI automation ROI properly, using a real client build as the worked example. Not a hypothetical. Actual hours, actual costs, actual payback period.
By the end, you'll have a framework you can apply to any process you're considering automating, before you commission anything.
Why Most Businesses Guess at ROI Instead of Calculating It
Here's what usually happens. Someone in the business reads an article about AI saving companies thousands of hours a year. They see a stat like "automation reduces admin costs by 40%". They decide they want some of that.
Then they either dive straight into a build without doing the maths, or they ask a vendor who tells them exactly what they want to hear.
Neither gives you a reliable number.
The honest reason most businesses guess at ROI is that calculating it properly requires knowing exactly what the current process costs. And most businesses don't track that. They know tasks happen, they know they take time, but they've never sat down and put a number on it.
That's the starting point. Everything else flows from there.
The Only Numbers That Actually Matter
Three numbers determine whether an automation build makes financial sense. You don't need a fancy spreadsheet. You need these three, and you need them to be honest.
Current cost of the manual process (time × rate)
This is the baseline. Take the time spent on the process each week, multiply by the fully-loaded hourly cost of the person doing it, then scale to a year.
So if someone spends 10 hours a week on a process, and their total cost to the business is £25 per hour (salary plus employer's NI, pension, and overhead), that's £250 a week. Over 48 working weeks, that's £12,000 a year tied up in one process.
A few things to get right here. Use the fully-loaded cost, not just the salary. Use realistic hours — not what you think the process should take, but what it actually takes, including interruptions, error correction, and the time spent chasing information.
And if the process touches multiple people, add up all of them.
Build cost and ongoing maintenance
This is the investment side. A bespoke AI automation build will typically run from £3,000 to £15,000+ depending on complexity. Ongoing maintenance, covering updates, monitoring, and handling edge cases that come up, usually adds 10-20% of the build cost annually.
Be realistic about both. Vendors who tell you there's no ongoing cost are not being straight with you. Any system that touches real business data needs monitoring and occasional adjustment. That has a cost.
Payback period and break-even point
To calculate ROI on AI automation, the core formula is: annual saving from automation minus annual cost of automation, divided by total automation cost, expressed as a percentage.
Or more practically: divide the total build cost by the monthly saving to get your payback period in months. If a build costs £8,000 and saves £1,000 a month, you break even in 8 months. Everything after that is pure return.
That's the number worth focusing on. Most businesses can tolerate a 6-12 month payback period for a process that then runs indefinitely.
A Worked Example: What AMPL Calculated for a Client Build
Theory is one thing. Here's a real one.
The process: email-to-quote automation for a parts supplier
One of our clients, a parts supplier, was receiving between 30 and 60 customer enquiries a day by email. Each one needed to be read, interpreted, looked up in the product catalogue, priced, and a quote sent back.
Their team was doing this manually. It took around 8 minutes per enquiry on average. On a 40-enquiry day, that's over 5 hours of work — just for quoting, before anything else got done.
We built an email-to-quote automation using Claude. It reads the inbound email, identifies the parts being requested, pulls the relevant pricing from the catalogue, generates a formatted quote, and sends it without a human in the loop for standard enquiries. Unusual or high-value requests get flagged for human review.
Hours saved per week and what that translated to annually
The automation handled around 80% of enquiries without human involvement. On a 40-enquiry day, that's 32 quotes handled automatically, saving roughly 4.3 hours of staff time daily.
Across a 5-day week, that's approximately 21 hours saved. At a fully-loaded staff cost of £22 per hour, that's £462 per week, or just under £22,200 per year.
The remaining 20% of enquiries still needed a human. But because the system flagged them clearly and pre-populated what it could, even those took less time than before.
Build cost vs 12-month saving
The build cost for this project was £7,500. Ongoing maintenance is budgeted at £1,200 per year, covering updates when the catalogue changes, monitoring, and the occasional edge case that needs handling.
Total first-year cost: £8,700. First-year saving: £22,200. Net first-year return: £13,500, or roughly 155% ROI.
Payback period: just under 5 months.
From month 6 onwards, the business was saving more than £1,800 a month in staff time. That time got redirected to customer relationships and handling the more complex enquiries that genuinely needed a person.
Where ROI Calculations Go Wrong
The numbers above are real. But they're also the clean version. Here's where calculations get optimistic in ways that cause problems later.
Underestimating implementation time
A build doesn't go from zero to fully running overnight. There's scoping, build, testing, refinement, and a period where the team learns the new workflow. During that time, the old process often still needs to run in parallel.
For most builds, budget for 4-8 weeks before you see the full saving. Factor that into your payback calculation. Your effective break-even is usually 4-6 weeks later than the raw numbers suggest.
Ignoring edge cases and error-handling overhead
No automation handles 100% of cases perfectly. There will always be unusual inputs, edge cases, and the occasional thing the system doesn't know what to do with.
The question is how much overhead that creates. A well-built system flags exceptions cleanly and routes them to the right person. A poorly-built one creates more work than it saves because staff are constantly firefighting failures.
In your ROI calculation, be conservative about the automation rate. We used 80% above because that's what we measured. If you're estimating without real data, 70% is a safer assumption for a first build on a complex process.
Forgetting the cost of doing nothing
This one cuts the other way. Most ROI calculations focus on the cost of the build. They don't account for what it costs to not build it.
If your quoting process takes 5 hours a day and you're getting slower as enquiry volume grows, you'll either need to hire someone, lose enquiries, or both. That has a cost too. The decision isn't "build vs free". It's "build vs the current trajectory".
To be honest, this is the framing that changes the most minds when we're working through a brief with a client.
A Simple Framework You Can Use Before You Commission Anything
Four steps. Run through them for any process you're considering.
Step 1 — Time the process honestly. Track it for a week. Not how long it should take — how long it actually takes, including interruptions and rework. Multiply by the number of people involved.
Step 2 — Cost it out. Fully-loaded hourly rate times hours per week times 48 working weeks. That's your annual baseline.
Step 3 — Get a real build quote. Not a vendor estimate pulled from thin air. An actual scoped quote based on your specific process. Add 15% for ongoing maintenance.
Step 4 — Calculate payback. Divide total first-year cost by the monthly saving. If it's under 12 months, the build almost certainly makes sense. If it's 12-18 months, it depends on how stable the process is. Over 18 months, you're either looking at a complex build that needs phasing, or this might not be the right process to start with.
That's it. No automation ROI calculator required, though you can build one in a spreadsheet in about 10 minutes if you want to run scenarios.
When the ROI Maths Works — and When It Doesn't
The numbers work well when the process is high-volume, relatively consistent, and currently eating significant staff time. Quoting, invoice processing, onboarding, reporting, data extraction — these are processes where automation compounds quickly because they happen constantly.
The numbers work less well when the process is low-volume, highly variable, or relies on judgement that's difficult to codify. If something happens 3 times a month and each instance is completely different, automation is probably not the right tool. Or at least not the starting point.
The other thing that kills ROI is starting with the wrong process. We see this fairly often. A business automates something peripheral because it seems easy, then wonders why the impact isn't there. The question to ask first is: what process, if it ran itself, would actually change how the business operates? Start there.
And honestly, if you're not sure which process to start with, that's exactly what an audit is for. We look at the full operation, identify where the time and cost is concentrated, and tell you what the ROI looks like before you commit to anything.
FAQ
How do you calculate ROI on AI automation?
The formula is: (annual saving from automation minus annual automation cost) divided by total automation cost, multiplied by 100. Annual saving is the hours saved times the fully-loaded hourly staff rate. Annual automation cost is maintenance and licensing. Total automation cost is the build fee plus first-year maintenance. A payback period under 12 months is generally a strong case for proceeding.
Is AI automation worth it for small businesses?
It depends on the volume and cost of the manual process, not the size of the business. We've built automation for businesses with under 10 staff where the ROI was over 200% in year one, because the manual process was genuinely consuming most of a person's working week. The filter is operational drag, not headcount.
What's a realistic payback period for an AI automation build?
For well-scoped builds targeting high-volume processes, 4-9 months is typical. Builds that take longer to pay back usually involve either a lower-volume process or a complex implementation. If a vendor quotes you a payback period under 3 months on a significant build, push them to show the working. The numbers may not be honest.
What ongoing costs should I include in an AI automation ROI calculation?
Include maintenance (typically 10-20% of build cost annually), any API or platform costs, and a realistic allowance for staff time handling exceptions the system flags. It's also worth accounting for the occasional update when your underlying process changes. And processes do change, so assume at least one per year.
How do I know which process to automate first?
Start with the process that has the highest manual time cost and the most consistent inputs. High volume, relatively predictable, time-consuming — that's the profile. A structured audit of your operations is the most reliable way to identify it, because the highest-cost process is often not the most visible one from the inside.
Can I use a generic automation ROI calculator for AI builds?
Generic calculators are fine for a rough sense check, but they won't account for your actual process complexity, automation rate, or maintenance requirements. The numbers in a generic calculator are usually optimistic. Use one as a starting point, then build your own model with your actual costs. The four-step framework above gives you everything you need.
