STIVTo measure ROI on AI agents, pick one workflow, record how it performs today, run the agent on it with human review, and compare the same measures afterwards: time to complete, time spent reviewing, error rate, and what the workflow costs. ROI is the value of the hours and errors saved, minus the cost of the software and the review time it adds.
That sounds obvious, but most AI ROI figures you'll see skip the first step. Without a baseline, every number afterwards is a guess.
Record the baseline before anything changes
- Volume: how many times the workflow runs per week or month.
- Cycle time: from request to finished work, not just hands-on time.
- Hands-on time: hours people actually spend preparing the work.
- Error rate: how often work is corrected, returned, or causes a problem later.
- Cost: hours multiplied by a loaded hourly cost, plus any tools already used.
Two to four weeks of baseline is usually enough for a weekly or daily workflow. If nobody can say how long the work takes today, that discovery is itself useful — it's often where the first real savings come from.
Count review time honestly
With a human approval gate, people still spend time on the workflow: reading drafts, editing, approving. That time belongs in the 'after' number. A finance reconciliation that used to take six hours and now takes forty minutes of review saved about five hours, not six. Leaving review time out is the most common way AI ROI gets overstated.
Watch the edit rate, not just the speed
The share of agent outputs approved without changes is the best early signal of whether the workflow is working. A rising edit-free rate means the playbook is clear and the agent is following it; a flat or falling one means reviewers are quietly redoing the work, and the time saved is smaller than it looks. As covered in what a rejected approval looks like, every edit is also feedback that makes the next draft better.
Mistakes that make the numbers meaningless
- Measuring a demo workflow instead of one the team actually runs every week.
- Comparing against a bad week, or against an estimate nobody recorded.
- Counting hours 'freed' as savings without saying what those hours went to.
- Ignoring the cost of setup, connection, and playbook writing in the first months.
For STIV specifically, this piece on measuring a division covers which signals to read in the first 90 days. The method above applies to any AI agent, from any vendor: one workflow, a real baseline, honest review time, and a number you'd be comfortable defending to a CFO.