Start with the current process, not the proposed technology. The objective is to understand what the workflow costs today, what a better state changes, and which assumptions must be proven.
1. Establish the operational baseline.
Document volume, handling time, wait time, rework, error frequency, escalation, systems touched, and the people involved. Separate active labor from elapsed time: a task may require ten minutes of effort but delay a customer for two days.
Include the work around the work—copying information, searching for context, checking status, asking for approval, reconciling systems, and recovering from exceptions. These small steps often create more economic drag than the visible task.
2. Value more than hours saved.
Capacity
How much repeatable work can be absorbed without adding headcount?
Quality
What is the cost of incorrect data, missed steps, inconsistency, and rework?
Speed
What does faster response, scheduling, approval, or fulfillment change?
Visibility
What becomes possible when status and performance are measurable?
A simple first-pass model
Annual benefit = recovered labor capacity + avoided error cost + delay reduction + additional contribution enabled.
First-year cost = discovery + build + integration + change management + infrastructure + operating support.
First-year ROI = (annual benefit − first-year cost) ÷ first-year cost.
Keep each assumption visible. Use a conservative case, an expected case, and an upside case rather than hiding uncertainty inside one confident number.
3. Price exceptions and risk.
Automation rarely eliminates every manual step. Model the expected exception rate, the cost of human review, failure recovery, vendor limits, model usage, support, and future change. A system that handles 85% of a workflow reliably may be more valuable than one that claims 100% automation but creates invisible risk.
Where errors affect money, safety, compliance, access, or customer commitments, the design should include deterministic validation and explicit approval. The return model should reward reliability, not just touchless execution.
4. Define a production proof.
Choose one measurable slice with sufficient volume, a clear owner, accessible data, and an observable outcome. Establish the baseline before launch. Then compare throughput, quality, response time, exception behavior, and total operating cost after the system is in production.
The right question is not “Can AI do this?” It is “Can a controlled system improve this workflow enough to justify its cost and operational risk?”
Assess an automation opportunity