Start with a bounded buying decision
Pick a workflow where the inputs, policy, and approval owner are clear: an RFQ comparison, a repeat purchase, or a tail spend request. Write down what the system may recommend and what it may execute. A purchasing decision needs more than a plausible answer.
Separate interpretation from calculation
Use a language model to read specifications and supplier responses. Calculate totals, currency conversions, and policy thresholds in ordinary code. Keep the original quote beside each extracted value. Missing freight or an ambiguous unit price should stop the comparison, rather than become a guessed number.
Measure the work that remains
Track buyer corrections, exceptions, time to an approved recommendation, and total cost per completed request. Start in recommendation mode. Expand execution permissions only after the system handles your difficult cases reliably.