Spend under intelligence
A top 10 US technology company reviewed every purchase request by hand. We replaced that with a system that learns from every decision it makes.
Top 10 US technology company · 2023–2024
A billion dollars a year, reviewed inconsistently.
Indirect procurement covers software licences, external services and supplies: high volume and hard to control. Different reviewers applied different standards, different parts of the organisation used different criteria, and nothing learned from decisions already made. Around 6% of spend was avoidable.
A system that starts smart and gets smarter.
Each request is reviewed against policy, past decisions and category benchmarks. Clear cases are processed; edge cases go to a human with a recommendation. Every confirmation or override becomes a training signal, and after each batch the model updates.
Thresholds rise as accuracy does.
As measured accuracy improved, procurement leadership raised the thresholds for autonomous handling. Reviewers focus on the genuinely complex. The system running today is materially smarter than the one deployed on day one.
Four things worth carrying forward.
The learning loop is the product
A model that stays fixed decays. One with a good feedback loop improves every day it runs.
Humans make the system stronger
Keeping people on edge cases is how the system learns.
Consistency at scale is a superpower
One standard applied to every request is what moved the avoidable-spend number.
Autonomy has to be earned
Recommendations first, decisions by people, autonomy expanding with accuracy.