Avis Budget Group: AI Transformation in Global Procurement

Context

A CEO-sponsored AI transformation initiative at a NASDAQ-listed mobility company with ~25,000 employees and a 600,000+ vehicle global fleet. I was embedded with Global Procurement on the AI Solutions and Enablement team, reporting to the Office of the CEO, owning AI use cases end to end: discovery, scoping, build, and adoption.

Problem

Procurement runs on high-volume, judgment-heavy workflows: ~100,000 emergency roadside tows billed per year with identified supplier over-mileage, commodity supply shocks that surface too late to act on, and analyst work drowning in RFP responses and spend reports. The challenge: apply AI where it moves real money, with the governance an enterprise demands, and get a non-technical team to actually adopt it.

Approach

Deterministic-first architecture with a bounded LLM judgment layer: free deterministic checks run on 100% of events, paid routing and LLM inference are reserved for the flagged, dollar-ranked subset, and a human approves every dispute. Adoption was treated as its own workstream, not an afterthought: live enablement workshops built around the jagged-frontier model of which tasks to hand to AI and which to keep human.

Frameworks

Deterministic-First AI Design (bounded LLM judgment)AWS Serverless Architecture (Lambda, Step Functions, EventBridge)Amazon Bedrock (schema-validated JSON, fail-to-human routing)AI Governance (NIST AI RMF, OWASP LLM Top 10)Jagged-Frontier Task Allocation (adoption design)

Implementation

  • •Designed an end-to-end AWS serverless pipeline auditing ~100,000 annual roadside-tow invoices, targeting recovery of $1M+ of $2.5M in identified supplier over-mileage; containerized for deployment on AWS EKS
  • •Built a capability-aware "should-have-distance" engine (BallTree haversine nearest-neighbor over repair-capability-filtered facilities) with road-distance verification via Amazon Location Service
  • •Built a commodities market-intelligence agent on AWS Bedrock: early warning on supply and price shocks, mapped to the company's own supplier exposure, with a layered anti-hallucination system (span-level citations, deterministic citation checks, separate-model entailment verification)
  • •Added a governed human-in-the-loop feedback loop that stores adjudicated reviewer decisions as precedent (Postgres/pgvector) while the LLM stays frozen
  • •Authored a 15-page system design and a 28-page build specification; produced a spec-driven agentic build package for AWS Kiro and Claude Code
  • •Designed and led live AI enablement workshops for the US and UK Procurement teams with live demos on RFP triage, contract summarization, and supplier-risk briefings

Outcomes

  • ✓Audit system targeting $1M+ recovery of identified supplier over-mileage (projected by business owners; delivery in progress)
  • ✓Two live AI enablement workshops delivered to US and UK Procurement teams, including executive leadership
  • ✓Market-intelligence agent validated against a real 2026 base-oil supply squeeze with a hand-built sample flag
  • ✓Per-cycle cloud cost held to tens of dollars against six-figure recovery via the tiered cost funnel

Learnings

  • →Enterprise AI wins come from bounding the LLM, not unleashing it: code decides the flow, the model fills one narrow judgment slot
  • →Adoption is a designed outcome: vocabulary, mental models, and live demos matter as much as the system itself
  • →Dollar-ranked funnels keep AI economical: spend inference only where the money is
  • →Impact figures stay labeled as projections until the results land; credibility compounds