Rohlik's Availability Problem
Rohlik is one of Europe's fastest-growing online grocers. The business depends on one simple number: in-stock availability. If a customer orders ten items and two are out of stock, they don't come back.
At the start of 2026, Rohlik's availability sat at 78%. That's not catastrophic for a large-format grocer, but it was well below the target the ops team had been chasing for two years. Every point of availability maps to measurable repeat-order lift.
The team had tried better forecasting, better supplier contracts, tighter reordering rules. The root problem was different: too much decision-making across the ops stack that humans simply didn't have bandwidth to handle in time.
The Duvo Deployment
Rohlik partnered with Duvo, an AI-agent specialist, to rebuild the decision layer on Claude. The agents sit between Rohlik's ERP, POS, and supplier systems and make four kinds of decisions continuously:
- Reorder calls: Which SKU, which supplier, how many, ship-by when — pushed as pending decisions for a human to approve or override.
- Substitution calls: When an item is about to go out of stock, the agent proposes the best substitute to offer the customer (and in which orders to substitute vs. backorder).
- Price calls: Short-dated / clearance pricing recommendations for fresh stock that will otherwise get scrapped.
- Flag calls: "Something weird happened in this warehouse overnight" — the agent surfaces anomalies the humans should look at first.
What Happened
Within two weeks of the agents going live, availability climbed from 78% to 93%. That number alone drove a lift in repeat-order rate that, compounded across Rohlik's volume, translated to €2.8M+ in annualized margin within three months of the deployment.
The non-obvious result: the ops team didn't shrink. Instead, 40%+ of team capacity was freed from decision overhead — and redirected to relationships with suppliers, new warehouse launches, and category expansion. The agents became the throughput floor; humans became the strategy layer.
Full case study: Anthropic customer page on Rohlik.
The SMB Translation
You probably don't run a European online grocer. But you probably do run a business with a "decision layer" that's bottlenecked on human attention. A few shapes that match:
- Restaurant: Daily prep quantities for 80+ SKUs based on weather + reservations + last week's leftovers.
- Home service: Which of next week's jobs should get the A-team vs. the B-team based on scope, revenue, and client history.
- E-commerce: Which campaigns to scale, which to kill, which ad creatives to rotate in for the weekend.
- Professional services: Which invoices to chase, in what order, with what tone.
In each case, the same mechanism works: put a Claude agent in front of the decisions, let it propose, let a human approve or override. The agent doesn't replace the human — it clears the queue so the human has time to handle the three decisions that actually matter.
What It Takes to Ship
Rohlik's rollout was three months end-to-end. For an SMB version of the same pattern, you're looking at a 2-4 week sprint: pick the one decision flow that eats the most of your day, wire a Claude agent to the relevant data sources, and put a simple approval queue in Slack or email. That's what the Ascero AI Weekly Report and Follow-Up Machine Packs ship as fixed-fee projects.