Anthropic credits Duvo’s AI agents with over €2.8 million in annualized retail savings

Retailer Rohlik Group achieved more than €2.8 million in annualized savings over three months using Duvo’s Claude-powered operations agents, according to an Anthropic case study. The account describes more than faster paperwork: agents took on price monitoring and supplier follow-ups that staff had not been able to run systematically across the business.

The distinction is central to Duvo’s pitch. Its customers’ operations teams prioritize major suppliers and urgent problems, leaving smaller tasks unattended. Duvo calls that backlog “abandoned work”: potentially valuable processes that fall between disconnected systems or exceed the hours available to employees.

The results are company claims published by Duvo’s AI supplier, not independently verified findings. Annualized savings express a yearly rate; the headline figure does not mean Rohlik saved that amount in cash during the three-month period.

Faster negotiations versus previously neglected work

Rohlik tracks thousands of products whose prices are linked to commodities across five markets. Before Duvo, that monitoring happened quarterly at best, and sometimes not at all. The agents now monitor continuously, prepare cases for negotiating prices and initiate contact with suppliers, according to the case study.

Anthropic attributes €1.45 million in annualized savings in the first week to price monitoring and automated outreach covering more than 120 products and 15 suppliers. That example concerns work Rohlik had not been able to run systematically, rather than simply completing an established process more quickly.

Other reported gains fit the more familiar efficiency argument. Approximately 80% of Rohlik’s annual supplier-negotiation process was automated, from preparation through contract generation and updating its business software. Negotiations finished one month sooner. Inbound delivery confirmations also rose from 52% to 90%, with every supplier covered daily.

The real process lives in people’s heads: which supplier needs chasing, which portal field actually matters, which exception to escalate and which to just fix

Marek Paris, Duvo co-founder and CPTO

One job across several business systems

Duvo builds agents for procurement, supply chains and retail category management using the Claude Agent SDK, a developer toolkit for building agents on Claude Code. Its agents span enterprise resource planning software—the systems used to manage business operations—supplier portals, spreadsheets, email and phone calls.

The case study describes agents operating software through its existing interfaces, rather than depending on dedicated software connections that may not exist. A single illustrative session can include several steps:

1. Log into a supplier portal, retrieve delivery status for 50 purchase orders and cross-check the information against SAP.
2. Identify discrepancies, check contract terms and decide whether to escalate a problem or correct it automatically.
3. Send follow-up emails and record the outcome without handing the job to another agent.

Duvo’s engineering head, Ondrej Romancov, says earlier agent handovers lost critical context. The current design gives one agent each job and access to the tools needed to finish it, keeping context together throughout the operation.

Human approval inside a single-provider stack

The agents do not have unrestricted authority. High-risk actions require human approval, and the system retains those decisions for future runs. This is how Duvo says it captures operational judgment that previously sat with individual employees.

Duvo chose a single-provider architecture after evaluating multiple model providers. It defaults to Sonnet 4.6, while browser-use agents often use Opus 4.6. Anthropic says every Duvo API call runs under its Zero Data Retention mode.

The company puts its average time from first customer conversation to production deployment with measured savings at eight weeks. That timeline covers the customer rollout, not just switching on the underlying agent toolkit.