Most conversations about AI start with the technology: a model, an agent framework, a demo. That is the wrong starting point. The useful question is narrower: which specific piece of work, done today by a person, is worth automating — and why has it not been automated already?
A simple framework answers that question with four checks. A process only deserves investment once it clears all four.
1. The work is genuinely repetitive
If a task varies significantly every time it is performed, automation adds risk rather than removing it. Good candidates are tasks where the inputs and the decision logic are stable: extracting fields from a similar document type, routing a request based on a known set of categories, or preparing a report from the same handful of data sources every week.
2. The exception rate is known and manageable
No process is 100% standard. The relevant metric is not "is this process repetitive" but "how often does it deviate, and what happens when it does." A process with a 5% exception rate that routes to a human for review is a strong candidate. A process with a 40% exception rate usually means the process itself is not yet well-defined enough to automate — fix the process first.
If nobody can tell you the exception rate, that is itself useful information: it means the process has not been measured, which should happen before any automation work begins.
3. The data already exists in a usable form
Automation is often blocked less by the "AI" part and more by data access: a system without an API, a spreadsheet that is the real source of truth, or information that only exists in someone's inbox. Before scoping an automation, confirm the data is reachable — programmatically, reliably, and with clear ownership.
4. The business impact is worth the engineering cost
Time saved is not the only measure. Consider error reduction, faster response times for customers, and the cost of the current workaround (a spreadsheet macro, a shared email alias, a person doing double data entry). A process that saves four hours a week but generates the exact same value in error reduction and speed for a customer-facing team is often worth more than a process that saves twenty hours but never touches a customer.
What this means in practice
Ranking opportunities against these four checks produces a shortlist that is usually smaller and different from the first list a team writes on a whiteboard. It also produces a natural pilot: the item with the clearest exception rate, the most accessible data, and the highest combined impact-to-effort ratio.
The technology choice — a scripted workflow, a retrieval-augmented assistant, or a fully autonomous agent — comes after this analysis, not before it. Committing to a tool before this exercise is the most common reason automation projects stall: the tool works, but it was solving the wrong problem.