How to rank automation opportunities before you build

Start with the operating constraint

Do not begin with a model or a tool. Write down the manual workflow, the person who owns it, the systems it touches, the volume, the failure modes, and the controls that cannot be removed.

A useful inventory is specific enough to expose whether the problem is classification, data movement, exception handling, approval latency, or missing source data.

Assess five factors separately

The worked example uses a qualitative screen, not a weighted numerical score. Time to value is left unestimated; resolve that gap before committing to either candidate pilot.

• Impact: the value of solving the delay, error, or capacity constraint.
• Feasibility: whether the systems expose stable APIs, events, files, or other usable interfaces.
• Risk: the cost of a wrong action, including privacy, compliance, and reversibility.
• Data readiness: whether representative, permissioned inputs and expected outputs exist.
• Time to value: how quickly a bounded pilot can produce a decision rather than a broad promise.

Reject the wrong first pilot

High impact does not compensate for low readiness and irreversible risk. In the demonstration, autonomous exception resolution is rejected as a first project. Document intake is a candidate because it combines high impact, high feasibility, and a clear human-review path. Its delivery effort and time to value still need discovery.

Write acceptance criteria before architecture

• Every item produces either a valid structured record or a routed exception.
• Every automated action records its source, model or rule version, decision, approval state, and final disposition.
• High-risk classifications require human approval; irreversible actions are not autonomous.
• The pilot can be disabled without data loss and has a documented manual fallback.

Make the first delivery decision-ready

The useful deliverable is not a slide that says automation is possible. It is a ranked inventory, a reference architecture, a pilot charter, named owners, measurable gates, and an explicit list of what should not be automated yet.