Step one: price the manual work honestly
Take one workflow. Count the hours per week actually spent on it — including the coordination, chasing, and rework nobody logs. Multiply by the fully-loaded hourly cost of the people doing it (salary plus benefits plus overhead, typically 1.25–1.4× base). Most teams underestimate this number badly, because the work is spread across people who each do 'a bit of it'.
Worked example: an ops team spends a combined 60 hours a week on order-document intake and reconciliation. At a conservative $45/hour fully loaded, that's $2,700/week — roughly $140K a year on work no one was hired to do.
Step two: add the costs that don't show up in payroll
Manual work carries a shadow cost that usually rivals the labor itself:
- Error cost — duplicate payments, missed discounts, compliance findings, refunds and goodwill credits from data entry mistakes
- Latency cost — deals that stall while paperwork sits in a queue; customers who churn waiting for answers
- Opportunity cost — what your best people would produce if they weren't doing robotic work
- Coverage cost — everything that arrives at night or on weekends and waits until Monday
Step three: subtract the true cost of the system
On the other side of the ledger: the one-time build investment, plus run costs (model usage, hosting, monitoring — usually modest relative to labor), plus the ongoing human-review time for exceptions. A credible vendor will give you all three numbers after mapping the workflow; be suspicious of anyone who quotes before looking.
In our worked example, a system that removes 70% of those 60 hours returns roughly $98K a year in labor alone — before error and latency savings. Against a build investment that's a fraction of that annual figure, payback lands within the first two to three quarters, and every year after is compounding return.
The two mistakes that wreck AI business cases
First: automating a low-volume, low-pain workflow because it looked easy. ROI lives where volume and pain are highest — start there. Second: comparing against software prices instead of labor costs. An agent system is not a SaaS seat; it's a worker that costs a fraction of one salary and never leaves.
If you want a second pair of eyes on the math, bring your workflow numbers to a scoping call — we'll run this exact formula with you and tell you honestly if the ROI isn't there.
Frequently asked questions
What is a good ROI for AI automation?
Well-scoped production systems routinely return 200–400% in the first year, because the labor and error costs they eliminate recur monthly while the build cost is paid once. If a projected business case shows less than ~100% first-year return, the workflow is usually wrong — pick one with more volume or more pain.
How quickly does AI automation pay for itself?
Typical payback is one to three quarters for a focused first system targeting a high-volume manual workflow. The payback clock is driven by the monthly manual cost being deleted, so higher-volume workflows pay back faster.
How do I estimate hours spent on a manual workflow?
Shadow the workflow for one week and count everything: the core task, the chasing, the corrections, the handoffs. Multiply observed hours by 1.2 to account for what you didn't see, then by fully-loaded hourly cost (roughly 1.25–1.4× base salary). Teams almost always undercount because the work is fragmented across people.