AI automation ROI compares what a process costs today with what it will cost after automation, then sets the saving against the build cost. You need six inputs: volume, minutes per item, loaded hourly cost, error and rework cost, build and run costs, and the share of items a person still reviews. The rest is arithmetic.
The ROI formula
The core formula is simple. Work out the monthly cost of the process today, subtract the monthly cost after automation, and that is your monthly saving. Divide the build cost by the monthly saving to get the payback period. First-year ROI is the year's savings, minus the build cost, divided by the build cost.
Written out, the formulas are:
baseline monthly cost = (volume × minutes per item ÷ 60 × loaded hourly cost) + (volume × error rate × rework cost per error)automated monthly cost = review labour + remaining rework + run costs + maintenancemonthly saving = baseline monthly cost − automated monthly costpayback months = build cost ÷ monthly savingfirst-year ROI % = (monthly saving × 12 − build cost) ÷ build cost × 100
Notice that the automated cost is never zero. AI automation still needs people to review some items, fix some errors, and keep the system running. Leaving those out is the most common way ROI estimates go wrong.
The inputs you need
Every input should come from real data where possible, not a guess in a meeting. Pull volumes from your systems, time a few people doing the task, and ask finance for a loaded hourly cost. If you have a process map with volume and time columns, most of the work is already done.
| Input | What it means | Where to get it |
|---|---|---|
| Volume | Items per month | System reports, inbox counts |
| Minutes per item | Hands-on time, not waiting time | Time ten real items |
| Loaded hourly cost (CAD) | Wage plus benefits, payroll costs and overhead | Finance or payroll |
| Error rate | Share of items that need rework | Correction logs, team estimate |
| Rework cost per error | Time and cost to fix one mistake | Time a few fixes |
| Build cost | One-time design, build, test and launch | Vendor quote |
| Run costs | Monthly software, AI usage and hosting | Vendor pricing |
| Maintenance | Monthly support, fixes and updates | Vendor quote or internal estimate |
| Human review % | Share of items a person still checks | Pilot results |
| Review minutes | Time to check one flagged item | Pilot results |
The loaded hourly cost matters more than people expect. A salary divided by working hours understates the real cost. Ask finance to include employer contributions such as CPP and EI, benefits, and a share of overhead. Use the figure they give you rather than a rule of thumb.
A worked example, step by step
This example is hypothetical. The numbers are made up to show the method, not drawn from a client or a benchmark. Picture a company processing 400 supplier invoices a month by hand, at six minutes each, with a loaded hourly cost of 45 CAD and occasional rework from keying errors.
Step 1: baseline labour. 400 invoices × 6 minutes = 2,400 minutes, or 40 hours a month. 40 hours × 45 CAD = 1,800 CAD.
Step 2: baseline rework. Assume 5% of invoices need correction, so 20 a month. At 25 CAD per fix, rework costs 500 CAD. Baseline monthly cost is 1,800 + 500 = 2,300 CAD.
Step 3: review labour after automation. Assume AI extracts the data and flags 20% of invoices for a person to check. That is 80 invoices × 3 minutes = 240 minutes, or 4 hours. 4 × 45 = 180 CAD.
Step 4: remaining rework. Assume the error rate falls to 1%, so 4 invoices a month × 25 CAD = 100 CAD.
Step 5: run and maintenance. Assume 250 CAD a month for software, AI usage and hosting, plus 150 CAD for maintenance. Automated monthly cost is 180 + 100 + 250 + 150 = 680 CAD.
Step 6: monthly saving. 2,300 − 680 = 1,620 CAD a month.
Step 7: payback and ROI. Assume a build cost of 12,000 CAD. Payback is 12,000 ÷ 1,620 = about 7.4 months. First-year ROI is (1,620 × 12 − 12,000) ÷ 12,000 × 100 = (19,440 − 12,000) ÷ 12,000 × 100 = about 62%.
For this kind of process, see how we approach invoice processing automation.
Reading the payback period
Payback period is how many months of savings it takes to recover the build cost. It is often more useful than ROI percentage because it is easy to explain and easy to check later. Most teams want a payback they can see within a year, but the right threshold depends on your cash position and risk.
Run three versions of the calculation: cautious, expected and optimistic. Change the review percentage, the error rate and the build cost between them. If the cautious case still pays back in a reasonable time, the project is sound. If only the optimistic case works, the project is fragile.
Remember that saved hours only become saved dollars if the time is used elsewhere. Freed capacity often goes into growth, faster service or work that was being skipped. That is real value, but say so plainly rather than counting it as cash.
What people forget to include
The costs people forget are the ones that show up after launch. Maintenance when a supplier changes a form or a system updates its interface. Exception handling for items the automation cannot process. Change management to train the team and adjust roles. Leave these out and payback slips by months.
- Maintenance. Integrations break when connected systems change. AI prompts and rules need tuning as inputs drift. Budget for it every month.
- Exception handling. Some items will always fall out. Someone needs time, a queue and clear rules to deal with them.
- Change management. Training, updated procedures and a period of running old and new side by side all cost time.
- Measurement. You need logs and a simple report to prove the savings. Plan for it in the build.
- Accuracy testing. Measure AI accuracy on your own data during a pilot before you trust the review percentage.
Is the investment worth it?
An AI automation investment is worth it when the cautious case pays back within a period your business is comfortable with, the process is stable enough to automate, and someone will own it after launch. If volume is low or the process changes every month, a simpler fix or a better process often wins.
Start with a process map to get honest volume and time numbers. Then compare AI with rules-based tools in RPA vs AI automation. If the numbers work, our AI automation service covers design, build and support.