Skip to content
Social Rocket

AI Agents in Practice: Crocs Keeps Humans in the Loop

Crocs is automating shopping and support with Rivet, but deliberately keeps humans in the process. The Dreamforce case shows why data quality, escalation and control remain important for AI agents.

From Social Rocket editorial team

This text was created with AI assistance and reviewed editorially before publication. All sources are linked below.

Illustrative image: working at a laptop on an online shop for fourth-quarter marketing budgets
Photo: Shoper / Unsplash

Crocs is expanding shopping and customer service with AI agents, but does not want its systems to run without human oversight. At Dreamforce 2026, Feliz Papich, SVP Digital Technology, Experience & Insights at Crocs, stressed that a "human in the loop" is still needed while companies learn the capabilities and limits of automation.

Rivet connects shopping and support

Crocs has introduced Rivet, an assistant based on Agentforce. It is designed to bring product discovery, order status, returns and customer questions together in one shared conversation. Crocs plans to later extend the system to HEYDUDE and other markets.

The approach fits Crocs' broader agentic commerce strategy. Salesforce already named Crocs in March 2026 as one of the retailers connecting product catalogues to ChatGPT via the Agentic Commerce Protocol.

The data foundation matters more than the interface

At Dreamforce, Papich emphasised according to Adweek that the real complexity often lies beneath the visible chat interface. Product data, customer history, inventory and service processes must be reliably combined so that an agent can give dependable answers.

For companies, this results in a clear checklist: current data, clearly defined actions, fixed escalation points to employees and traceable logs.

Human handover is not a step backwards

In customer contact, full automation can cause additional damage if an agent gives incorrect product information or mishandles an edge case. A targeted handover to employees is therefore more a sign of quality than a failure of automation.

The same principle applies to AI focus groups for content and to the Claude workflow as AI Creative Director: AI takes on work, but rules, data and approvals remain decisive.

Specialised agents are also being deployed more widely. DeepIntent automates parts of pharma media planning, while Meta is opening WhatsApp Business tools to MCP-compatible AI agents.

What marketers should take away

The next phase of AI automation will be decided less by whether an agent can do something and more by whether companies give it the right task, data foundation and decision scope. Crocs' example shows a pragmatic path: automate, observe real usage, improve data and keep humans in the process wherever context or risk requires it.

Sources