AI & Automation
Mapping an approval workflow into automated steps
A practical guide for turning a familiar approval routine into rules a system can run, using a purchase request as the example and focusing on decisions that keep automation understandable and reviewable.
01
Step one: write the axioms before the steps
Start with the statements that must always hold. For a purchase request, examples might be: no spend proceeds without a budget owner's approval; requests above a set amount need a second approver; the requester cannot approve their own item. Write each as a single sentence with a named owner. Resist drawing boxes yet, because diagrams tend to preserve habits rather than principles. Then ask which rules are policy and which are merely convenience, since only policy should be enforced automatically. This short list becomes the reference that every later automated step must satisfy and every reviewer can check.
02
Step two: convert each rule into a trigger and action
Take every axiom and express it as a condition, a responsible party and an outcome. The amount rule becomes: when the total exceeds the threshold, add a finance approver and pause until both responses arrive. The self-approval rule becomes: if requester equals approver, reroute to the manager's delegate. Add timing as well, such as a reminder after two working days and escalation after four. Keep each step small so it can be tested alone. Note the information each step needs, like cost center or vendor, so the intake form collects it up front rather than prompting repeated follow-up questions.
03
Step three: decide exceptions and human checkpoints
No ruleset covers everything, so choose in advance where a person takes over. For the purchase example, urgent requests might skip the queue but require a post-approval review, while unusual vendors always go to a procurement specialist. Record every decision with who acted, when and under which rule, so the history explains itself. Run the workflow on a handful of past requests and compare results with what actually happened, adjusting rules where they disagree. Review the axioms periodically, because a stale rule automated at speed simply repeats an outdated decision faster.