Case Study · Enterprise Technology

We put $1 billion in indirect spend under AI, and the system kept getting sharper.

A top-10 US technology company had a procurement review process that depended entirely on people reviewing purchase requests by hand. We replaced that with an AI system that learns from every decision it makes.

$1B+

Indirect spend under AI intelligence

6% → 3%

Avoidable spend reduction target

Every batch

Model accuracy improves with each decision cycle

Top-10 US Technology Company · 2023–2024 · Indirect Procurement · AI Implementation

The Situation

When you spend a billion dollars a year on indirect procurement, the review process matters a lot.

Indirect procurement covers everything from software licenses to external services to office supplies. It is high-volume and hard to control. For a company at this scale, that was costing real money.

When we came in, purchase requests were being reviewed manually. Different reviewers applied different standards. Different parts of the organization used different criteria. There was no shared model for what a good decision looked like, and no way to learn from decisions that had already been made. Every review started from scratch.

The outcome showed up in the numbers. Around 6% of indirect spend was categorized as avoidable, purchases that did not meet policy, were duplicates of existing contracts, or could have been negotiated better. The target was 3%. Closing that gap manually was not realistic at this volume.

Key insight

Most AI projects treat the model as the deliverable. We treat the learning loop as the deliverable. A model that gets deployed and stays fixed will decay. A model with a good feedback loop gets better every day it runs.

The Old Way vs The New Way

Human reviewers doing their best, versus a system that learns from every answer.

Manual review

Each reviewer applied their own judgment with no shared standard

High volume meant reviewers were stretched and decisions were rushed

No feedback loop, a poor decision stayed a poor decision

No pattern recognition across requests at scale

Avoidable spend at 6% with no clear path down

AI-driven review

One consistent model applied to every purchase request, every time

High volume is exactly where AI performs best, no fatigue, no shortcuts

Every batch of decisions feeds back to improve the model

Pattern recognition across millions of data points, continuously

Avoidable spend on a clear, measurable path toward 3%

How It Works

A system that starts smart and gets smarter with every decision it makes.

What made this system different from a standard rules-based approval workflow was the learning loop. Most procurement automation applies a fixed set of rules and stops there. Ours did not.

Request arrives

A purchase request enters the system. The AI reviews it against policy, past decisions, and category benchmarks.

Accept or escalate

Clear cases are processed automatically. Edge cases go to a human reviewer with a recommendation from the model.

Decisions are captured

When a human confirms or overrides a decision, that outcome becomes a training signal. The model notes what it got right and what it missed.

Batch learning cycle

After each batch of reviewed requests, the model updates. The next set of decisions comes from a stronger position than the last.

Autonomy expands

As accuracy improves, the system handles a larger share of decisions on its own. Human reviewers focus on genuinely complex cases.

📥

Purchase request received and classified

🤖

AI model makes an accept or escalate call

👤

Human reviewer handles edge cases

📊

Decision outcomes captured as training data

🔄

Model updates after every batch of decisions

📈

Accuracy improves and autonomy expands

The system running today is materially smarter than the one deployed on day one.

What We Learned

What makes a procurement AI system actually work in production.

The learning loop is the product

Humans make the system stronger, not weaker

Keeping people in the loop for edge cases is not a compromise. It is how the system learns. Every time a reviewer makes a decision, they are training the next version of the model. The goal is not to remove them, it is to focus their time on cases only they can handle.

Consistency at scale is a superpower

The biggest improvement over manual review was not speed. It was consistency. One standard applied to every request, regardless of volume or reviewer fatigue. That consistency is what made the avoidable spend number move.

Autonomy has to be earned, not assumed

We started with the AI making recommendations and people making decisions. Autonomy expanded as accuracy improved and trust was built. This is the only responsible way to deploy AI in a high-stakes financial workflow.