Direct answer
Measure the cost and quality of the whole workflow, not the percentage of steps touched by AI. The useful metrics are successful outcomes, operator minutes per case, exception rate, rework, time to recovery, and business leakage prevented. A workflow can be “90% automated” and still create more work than it removes.
Automation percentage rewards the happy path
AI automation demos usually show one clean input moving through one clean path. Operations are made of duplicates, missing fields, conflicting records, expired credentials, rate limits, policy exceptions, and people who reply in unexpected ways.
Across Zenveus presales research, automation appeared in 49 of 115 conversations. Reliability and data consistency were explicit desired outcomes in only 12, even though every automation eventually depends on them. That gap explains why many teams underestimate production work: they price the visible path and discover the exception system later.
Use a workflow ledger
A better ROI model treats every case as a small ledger. Record:
- whether the intended outcome completed;
- how many AI and API attempts were made;
- whether a person reviewed or corrected it;
- how long the person spent;
- whether the case was delayed or abandoned;
- what financial or operational value was created or protected.
This produces metrics that leadership can act on.
Outcome completion rate shows whether the workflow actually finishes. Operator minutes per completed case captures hidden human effort. Exception rate shows how often the normal path breaks. Mean time to recovery reveals whether failure is cheap or disruptive. Rework rate captures plausible-looking outputs that create downstream correction.
Human review is not automatically failure
Some workflows should retain review. A person approving a high-value renewal, clinical note, financial action, or customer commitment may be an intentional control—not inefficiency.
The question is whether review is focused. If an operator must inspect every output because the system offers no confidence, evidence, or prioritization, AI has moved the work rather than removed it. If the system routes only uncertain or high-risk cases, human attention becomes more valuable.
Evidence from renewal operations
In an insurance renewal workflow covering more than 4,000 policies, the important result was not that AI drafted messages. The system validated policy data, ranked renewals by urgency, logged actions, and alerted operators when execution failed. The reported lapse rate moved from 11.4% to 4.1%, protecting roughly $680,000 in annual premium.
That outcome came from prioritization and control around AI, not from maximizing the number of generated emails.
A second workflow monitored contractual deadlines every two hours, deduplicated warning, escalation, and breach alerts, and maintained an execution trail. The breach rate fell from roughly 14% to below 5.5%. Again, the valuable unit was a prevented breach, not an automated task.
A practical ROI formula
For a defined period, calculate:
Net workflow value = value created or leakage prevented − model/API cost − operator cost − recovery cost − maintenance cost.
Then compare it with the previous workflow. Include the cost of exceptions, not just average API spend.
For an early pilot, four measurements are enough:
- Baseline operator minutes per case.
- Successful outcomes without rework.
- Exceptions requiring intervention.
- Time and cost to recover failed cases.
If the pilot cannot measure those, it is not ready to make an ROI claim.
What should an automation dashboard show?
Show volume, completed outcomes, exception categories, operator minutes, retries, recovery time, model and tool cost, value created, and unresolved cases. Separate “AI produced an output” from “the business outcome completed.”
That distinction prevents impressive activity metrics from hiding operational debt.
Related Zenveus service: AI and Automation