Procurement Automation: What Production Deployments Show

Yes — and this category's numbers arrive in the currency procurement respects. The deepest documented result is Duvo's negotiation agents at Rohlik Group: €1.45 million in annualized savings identified in the first week, €2.8 million-plus by three months, with roughly 80% of the supplier negotiation process automated, product availability up from 78% to 93%, and annual supplier negotiations shortened by a month — eight weeks from first conversation to production with measured savings. The orchestration side compounds it: Cribl reports 15% average contract savings with benchmarking cut from three-to-four hours (or days) to under a minute; Bandwidth cut NDA turnaround 60% with zero added headcount; and Ramp's AI-powered intake and three-way match has caught more than $50 million in overbilling errors while making purchasing 8.5x faster.

One detail worth its own sentence for calibration: OpenAI runs its procurement on Zip's agent platform — 1,400 hours saved annually across ten-plus deployed agents, with two-thirds of the company engaging with procurement through it. When the frontier lab buys rather than builds its own procurement AI, that's a data point about where the leverage actually sits: in the workflow, the contract data and the policy engine — not the model.

Grounding — and this record preserves the failure with rare candour. One documented team's initial ChatGPT API integration failed because foundational models lacked domain knowledge and hallucinated fake suppliers; the subsequent agent-based attempt remained unpredictable and nearly impossible to debug, with agents randomly fabricating supplier answers even after domain adaptation and guardrails. In most functions a hallucination wastes time; in procurement it invents a counterparty. The systems that work — Duvo's negotiation agents included — succeed precisely because every claim traces to real contracts, catalogues and supplier records, with humans approving before commitment. Treat "where does this answer come from" as the first demo question for any procurement agent, and walk if the answer is vague.

The second failure is integration fragmentation, and it wears a specific documented face: a contract repository wired to the procurement platform so loosely that only final signed contracts synced, leaving stakeholders blind to versioning and status while email and Slack became the informal system of record under pressure. Procurement's control value lives entirely in before — before commitment, before signature, before payment — and every gap between systems is a place spend escapes into afterwards. The wins above all rest on one intake front door and repositories that stay in sync while documents are still moving.

Buy the orchestration spine — intake, policy checks, approval routing and contract visibility are exactly what the platform lane (Zip, Coupa, Ramp's procurement suite in this record) has productised, and the documented outcomes above ride on it. The OpenAI-buys-Zip data point is the argument compressed: even organisations with unlimited model access buy the workflow layer, because the moat is the policy engine and the integrations, not the intelligence.

The agent layer above the spine is where the interesting decision now lives, and the record documents a genuine third path: specialist agent builders deploying on frontier models into your procurement domain. Duvo's Claude-based agents at Rohlik are the proof case — and their before-state explains why the approach exists at all: traditional automation could not handle enterprise stack heterogeneity, with no clean APIs, years-long IT backlogs, and exception handling that required judgment. Agents that can read messy systems and exercise bounded judgment are precisely the tool for terrain integration projects die on — eight weeks to measured savings is the documented speed of that path done well.

Decision inputs: start with the spine if requests still live in email — nothing else works without intake. Add negotiation and benchmarking agents where supplier volume justifies them, grounded or not at all. And hold every layer to the same bar: checks before commitment, humans approving spend, and answers that trace to your data.