€2.8M+ in annualised savings captured in three months, from processes that had never been systematically run
78% → 93% product availability within two weeks, no APIs, no IT project
Eight weeks from first conversation to production deployment with measured savings
At Duvo, we build AI agents that run procurement, supply chain, and category management processes end to end for multi-billion-dollar retail and CPG companies. Our agents work across every system involved (ERPs, supplier portals, spreadsheets, email, even phone calls), and every action is auditable, governed, and human-approved where it matters.
Duvo is 100% built on Claude. Every API call we make runs under Anthropic's Zero Data Retention mode.
"In the first week, Duvo captured €1.45 million in savings our team never had time to go after. That was one process. We're now running dozens."
Olin Novák, CRO, Rohlik Group
The challenge: Millions in value nobody has time to capture
Retail and CPG operations teams know exactly what needs to happen. They just can't be everywhere at once.
Buyers manage ordering across SAP, supplier portals, and email. Category managers track commodity prices in spreadsheets they built themselves. Procurement coordinators chase confirmations by copying status between systems. These teams cover the top 20 suppliers and the most urgent issues. The long tail of hundreds of smaller actions worth millions in aggregate never gets touched.
"No company does daily competitive intelligence across tens of thousands of SKUs," said Marek Paris, our co-founder and CPTO. "The top 20 suppliers get a quarterly review. The rest get nothing. Not because it isn't valuable, but because there aren't enough hours."
The blocker isn't a missing dashboard. It's that the work spans systems that don't connect (ERPs, portals, email, spreadsheets, phone calls) and the real process lives in people's heads. Traditional automation stalls because there are no clean APIs, the IT backlog is years long, and every exception requires judgment. That's the gap we set out to close.
Why Claude: Reasoning depth meets enterprise trust
After evaluating multiple model providers, we committed to a single-provider architecture on Claude.
"What set Claude apart was its performance on messy, ambiguous tasks," Paris explained. "Parsing a supplier email that half-confirms a delivery while raising a pricing dispute. Navigating a SAP GUI screen with dozens of fields. Making judgment calls on exceptions without hallucinating business rules."
Two factors cemented the decision for us:
- Reasoning depth for multi-step work. A single agent run might log into a supplier portal, extract delivery status for 50 purchase orders, cross-reference against SAP, identify discrepancies, check contract terms, decide whether to escalate or auto-correct, send follow-up emails, and log the outcome, all in one session.
- Enterprise-grade trust through a purpose-built ecosystem. Anthropic didn't just build a powerful model. They pioneered the modern agent stack: the Agent SDK, MCP, Computer Use, and the tooling around them. We chose to go all-in on this ecosystem because it delivers enterprise-level control, reliability, and security by design, from Zero Data Retention on every API call to structured human-in-the-loop workflows and auditable tool orchestration. "Other providers offer a model. Anthropic offers the infrastructure to run agents in production with the governance enterprises require," Paris said.
Building on the Agent SDK, MCP, and Computer Use
We adopted the Claude Agent SDK before it was officially named, building production workflows while the SDK was still in early development. Our collaboration with Anthropic has been deeply technical. We've shared feedback on the SDK roadmap, contributed to patterns around hooks and structured behaviours, and pushed the boundaries of production agent systems.
- MCP lets us connect browser automation, ERP systems, data warehouses, email, and file storage through a single protocol
- Computer Use means our agents can operate supplier portals and legacy systems with no API, navigating the same screens a human would
- The Agent SDK gives us tool orchestration, sub-agent coordination, and human-in-the-loop workflows out of the box
- A generic tool system means any action, not just pre-defined approval steps, can be made to require human sign-off, letting operations teams place guardrails exactly where they need them
"Before the SDK, critical context disappeared between agent handovers," Paris said. "Now we run one capable agent per job with access to all the tools it needs. Context stays intact across the entire operation."
High-risk actions require human approval. 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.
Results: Rohlik Group
Rohlik Group, Europe's leading online grocer with €1.5B+ in revenue, is our anchor deployment.
- €1.45M in annualised savings in the first week, from continuous commodity price monitoring and automated supplier outreach across 120+ SKUs and 15+ suppliers. Not from a process that was slow, but from one the company had never been able to run.
- Annual negotiations shortened by one month, ~80% automated from preparation through contract generation to ERP write-back. 60+ FTEs shifted from data entry to strategic negotiation.
- Product availability: 78% → 93% in two weeks, by reconciling demand signals, stock levels, and supplier lead times scattered across systems. No API integrations, no IT project.
- Inbound deliveries confirmed: 52% → 90%, every supplier, every day, not just the top 20.
- Promotion setup time cut 65 to 70%, issues caught before launch instead of after.
- 100% of daily variance reporting automated, month-end close accelerated by ~10%.
- Margin reviews delivered 48 hours before trading week, 120 hours per week saved across commercial teams.
Total: €2.8M+ in annualised savings. Zero API integrations required.
Why this partnership matters
"We didn't want a model vendor," Paris said. "We wanted an AI partner that ships fast, listens to what production agents actually need, and takes enterprise trust as seriously as we do. That's Anthropic."
We're scaling across European retail and CPG, with early traction in banking and manufacturing. We guarantee time-to-value: eight weeks from first conversation to production deployment with measured savings. If the savings aren't there, full refund.
Every deployment deepens our understanding of how enterprises actually operate. That operational intelligence compounds, and the work that would make the biggest difference is the work nobody could get to before. Now someone can.
Read the full story on Anthropic's site: claude.com/customers/duvo
Frequently Asked Questions
Q: How does Duvo automate enterprise operations without API integrations?
Duvo agents use Computer Use to navigate legacy systems through the same screens a human would. They log into supplier portals, ERP transactions, and email clients directly. No custom API integrations, no IT project, no waiting for the integration backlog.
Q: What is Anthropic's Zero Data Retention mode and why does Duvo use it?
Zero Data Retention means Anthropic does not store prompts or outputs from API calls beyond transient processing. Duvo runs every API call under this mode to meet enterprise security and compliance requirements. Sensitive operational data — supplier contracts, pricing, inventory levels — never persists on Anthropic's infrastructure.
Q: How long does a Duvo deployment take from start to production?
Eight weeks from first conversation to production deployment with measured savings. Duvo maps the real process first using Duvo Clarity, then deploys agents against the actual workflow. If savings aren't delivered, the deployment comes with a full refund guarantee.
Q: What results did Duvo achieve at Rohlik Group?
Duvo captured €2.8M+ in annualised savings in three months. Product availability improved from 78% to 93% in two weeks. Inbound delivery confirmations went from 52% to 90%. Annual negotiations were shortened by one month with 80% automation. These results are documented in the case study published on Anthropic's customer stories page.
Q: How does Duvo compare to RPA for retail operations?
RPA relies on rigid scripts that break when interfaces change or exceptions occur. Duvo agents reason through ambiguity — parsing partial supplier confirmations, navigating changing portal layouts, making judgment calls on exceptions. High-risk actions require human approval, and every decision is persisted for future runs. The system compounds operational judgment over time rather than repeating the same brittle script.
Q: What systems can Duvo agents operate across?
Duvo agents work across SAP, Oracle, supplier portals, carrier sites, WMS platforms, Excel, email, and phone calls. MCP connects all tool types through a single protocol. Computer Use handles systems with no API. The agent orchestration layer coordinates work that spans multiple systems in a single session.