AI Automation5 min read
How to Measure AI Automation ROI: Baselines, Formulas and a Worksheet
Most AI projects cannot prove they worked because nobody recorded the before. A practical method for baselines, cost per task, hard and soft benefits and an honest payback calculation.
Contents
The most common reason an AI project is judged a disappointment is not that it failed but that nobody could show it worked. Six months in, the process feels faster, and the finance team asks for numbers that do not exist. This guide shows how to build a defensible ROI case before you write a line of automation, and how to keep it honest afterwards.
Step 1: capture the baseline
Before any build, measure the current process for two to four weeks. If you must estimate, sample real items and time them; do not rely on memory.
| Measure | How to capture | Example |
|---|---|---|
| Monthly volume | System counts | 800 invoices |
| Handling time per item | Timed sample of 20 to 30 items | 6 minutes |
| Error or rework rate | Audit sample; reopened or corrected items | 4% |
| Cycle time | Timestamps from start to done | 2.5 days |
| Fully loaded hourly cost | Salary plus overhead, from finance | $35 |
| People involved | Roles and share of time | 2 clerks, 40% of time |
Step 2: compute cost per completed task
Tool prices are the visible part. Include everything that scales with the work:
- Model and platform usage per completed task. Divide monthly spend by completed items, not by runs, since failures and retries cost money too. See the metering differences in n8n vs Zapier vs Make.
- Review time at the loaded hourly cost, including approvals and sampling.
- Rework and corrections on items the system got wrong.
- Maintenance, monitoring and prompt or rule updates, allocated per month.
- Build cost, amortized over 12 to 24 months.
Step 3: separate hard from soft benefits
| Hard benefits (bank these) | Soft benefits (report, do not bank) |
|---|---|
| Hours removed that reduce overtime, contractors or avoided hires | Employee satisfaction from less drudgery |
| Errors avoided, with a known cost per error | Better customer experience, unless tied to measured retention |
| Faster cash collection or faster lead response with measured conversion effect | Improved brand perception |
| Volume handled without added headcount | Learning and capability building |
Step 4: the calculation
Worked example using illustrative numbers: baseline cost per invoice is 6 minutes at $35 per hour, or $3.50. After automation, net handling time including review is 2 minutes ($1.17) plus $0.19 of model and platform cost per invoice, for $1.36 in all. The saving is $2.14 per invoice. At 800 invoices a month that is about $1,712, less $200 of fixed monitoring and maintenance, for a net benefit of about $1,512 a month. With a $9,000 build, payback is roughly six months. If volume were half, payback would double, which is why volume is the first factor in process selection.
Step 5: use a control, not a feeling
- Before and after on the same process, ideally with the same period length and seasonality.
- Parallel run or shadow mode, where the automation processes live items but a person still does the work, so you can compare outputs.
- A/B or staggered rollout where one team or region uses it and another does not.
Leading and lagging indicators
Financial payback takes months, so track leading indicators from week one: accuracy on the test set, share of items handled without intervention, reviewer edit rate, cost per task and time to first response. Lagging indicators, such as cost per case, customer retention and headcount avoided, confirm the case later. If the leading indicators are flat at day 60, the 90-day roadmap's Gate 2 is where you adjust or stop.
Common ROI mistakes
- Counting gross time saved and ignoring review and rework.
- Assuming savings are realized. An hour saved from a salaried person only counts if the capacity is used or a cost is removed.
- Forgetting variable cost at scale, especially for agents whose usage per task varies.
- Measuring activity, not outcomes: emails sent, tickets touched, rather than meetings held or issues resolved.
- Not re-measuring. Costs and volumes change, and so does the model.
The stakes are real. Gartner has forecast that over 40% of agentic AI projects will be canceled by the end of 2027 over cost, unclear value and risk controls, and the MIT NANDA report found most pilots showed no measurable profit-and-loss impact. A baseline and a cost-per-task model answer two of those three directly; the governance checklist addresses the third. See also why agentic AI projects fail and the AI automation guide.
Sources and further reading
Frequently asked questions
How do you calculate ROI for AI automation?
Net benefit divided by total cost. Net benefit is the annual value of hours saved, errors avoided and revenue impact, minus ongoing costs. Total cost includes build, platform and model usage, review and maintenance. Use a baseline measured before launch and conservative assumptions.
What metrics should I track for an AI workflow?
Volume handled, time per item, cost per completed task, accuracy or error rate, share needing human intervention, cycle time, customer or employee satisfaction and rework. Pick three or four that map to the business case and track them weekly.
Why do so many AI pilots fail to show ROI?
Commonly there is no baseline, benefits are vague, costs such as review time are ignored, or the process chosen has low volume. MIT's Project NANDA 2025 report found most pilots delivered no measurable profit-and-loss impact, and Gartner expects over 40 percent of agentic AI projects to be canceled by 2027 over costs, unclear value or weak controls.
How long should payback take?
For a well-chosen workflow on an existing platform, payback within three to nine months is a reasonable target. Longer paybacks are acceptable for strategic builds, but any project should show early leading indicators within the first 60 to 90 days.
How do I count time saved fairly?
Count net time: minutes saved per item minus minutes spent reviewing and correcting, multiplied by volume. Only count savings that are realized, for example reduced overtime, avoided hires or redeployed capacity with a defined use, not theoretical hours.
Work with Navelo Software
Published by Navelo Software
An independent software and AI studio building AI systems, web platforms, mobile apps and SaaS for clients worldwide.
