Procurement · Production

Duvo builds AI procurement agents on Claude, delivering €2.8M+ in annualized savings for Rohlik Group

The problem

Enterprise operations teams know exactly which processes need to run but cannot execute the full volume — procurement, supply chain, and category management work spans heterogeneous systems with no clean APIs, creating 'abandoned work' worth millions in aggregate that never gets touched.

First attempt

Traditional automation could not handle enterprise stack heterogeneity — no clean APIs, years-long IT backlogs, and exception handling that required judgment made prior approaches unworkable. Before the Agent SDK, critical context disappeared between agent handovers.

Workflow diagram · grounded in source

Abandoned work identified

trigger

“operations teams that know exactly what needs to happen but can't get to all of it”

Agent logs into systems and extracts data

ai_action

“log into a supplier portal, extract delivery status for 50 purchase orders, cross-reference against SAP”

Cross-reference and discrepancy detection

ai_action

“cross-reference against SAP, identify discrepancies, check contract terms”

Escalate or auto-correct decision

routing

“decide whether to escalate or auto-correct”

Human approval for high-risk actions

human_review

“High-risk actions require human approval”

Emails sent and outcome logged

output

“send follow-up emails, and log the outcome—all in one session”

Decision persistence and judgment accumulation

feedback_loop

“When a human responds, the agent persists that decision for future runs. Over time, the system accumulates the operational judgment that used to exist only in people's heads”

Reported outcome

Rohlik Group achieved €1.45M in annualized savings in the first week from continuous price monitoring across 120+ SKUs and 15+ suppliers, growing to €2.8M+ across three months.
Promotional setup dropped 65-70%, supplier onboarding chasing fell 50-70%, product availability rose from 78% to 93%, inbound delivery confirmations jumped from 52% to 90%, and annual supplier negotiations shortened by one month with approximately 80% of the process automated. Across enterprise deployments, Duvo frees up 40% of team capacity on average.

Reported metrics

Annualized savings — first week€1.45M

Annualized savings — three months€2.8M+

Promotional setup time reduction65-70%

Supplier onboarding chasing reduction50-70%

Show all 12 reported metrics

annualized savings — first week€1.45M

annualized savings — three months€2.8M+

promotional setup time reduction65-70%

supplier onboarding chasing reduction50-70%

product availability improvement78% to 93%

supplier negotiation process automatedapproximately 80%

inbound delivery confirmations52% to 90%

team capacity freed40%

annual supplier negotiations shortenedone month

SKUs monitored continuously120+

suppliers monitored continuously15+

time from first conversation to production with measured savingseight weeks

Reported stack

ClaudeAgent SDKMCPcomputer useSonnet 4.6Opus 4.6SAP

Source

https://www.anthropic.com/customers/duvo

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Rohlik Group achieved €1.45M in annualized savings in the first week from continuous price monitoring across 120+ SKUs and 15+ suppliers, growing to €2.8M+ across three months.

What tools did this team use?

Claude, Agent SDK, MCP, computer use, Sonnet 4.6, Opus 4.6, SAP.

What results were reported?

Annualized savings — first week: €1.45M; Annualized savings — three months: €2.8M+; Promotional setup time reduction: 65-70%; Supplier onboarding chasing reduction: 50-70% (source-reported, not independently verified).

What failed first in this deployment?

Traditional automation could not handle enterprise stack heterogeneity — no clean APIs, years-long IT backlogs, and exception handling that required judgment made prior approaches unworkable.

How is this procurement AI workflow structured?

Abandoned work identified → Agent logs into systems and extracts data → Cross-reference and discrepancy detection → Escalate or auto-correct decision → Human approval for high-risk actions → Emails sent and outcome logged → Decision persistence and judgment accumulation.