Book A Demo

Case study • Helix Sleep

When chat volume scaled past 50,000 conversations in five months, Helix needed automation that could act, not just answer

  • 72%

    Chat containment

  • 14,640

    Autonomous policy execution

  • 4.19★

    CSAT

Challenge

The team was running Zendesk automations that could answer FAQ-type questions: policy information, product comparisons, basic order status language. The capability ceiling showed up quickly with action-oriented requests. Updating a Shopify order, applying a current sales discount, routing a damaged-item claim through the right resolution path: these required a live agent to click through multiple systems. With inbound volume climbing and major sale events creating predictable spikes, the automation gap was a staffing problem with no clean headcount solution.

Helix needed automation that could close the loop on customer requests: look up the order, run the logic, take the action, and write the result back, without a human in the middle for every ticket. And not only that, but they needed a solution that could differentiate and scale across all of their brands.

Moving from rule-based automation to Kodif's generative approach changed what our AI could actually do for customers. Instead of building rigid flows, we write policies in natural language, and instead of telling customers it can't help, the AI answers generatively from our own content with just some tuning of guidelines. We went from roughly 35–40% containment to 71.8%, which has been a tremendous help during busy sale periods. We're excited to keep building out policies and automations so our agents stay free to tackle the harder issues.

Zachary Gentry

Sr. Director of CX

Solution

The Kodif deployment started with the highest-volume, most-repeatable flows: order status and damaged/wrong-item resolution. The goal was to clear the action-shaped work consuming agent capacity on tasks with predictable outcomes, and to give the CX team a policy layer they could extend without rebuilding every time the support mix shifted.

Where Is My Order, with real-time order data

WISMO is the highest-volume policy in the deployment. The AI Agent pulls live order data from Shopify, calculates business days in transit using carrier routing logic, and gives the customer a specific answer tied to their order, not a generic “allow 5-7 business days” response. For orders with more complex status (partial fulfillment, warehouse hold, or split shipments where the frame, mattress, and foundation arrive on separate timelines), the agent routes to a human with order context already attached.

Wrong and damaged item resolution

This is the second most-used action policy. For a high-ticket product like a mattress, wrong or damaged delivery is a high-stakes interaction. The policy runs through item verification, captures damage context, and routes to the correct resolution path based on the item type and damage category. CSAT on this policy sits at 4.01 stars.

Unsubscribe and shipping address updates

Lower-volume but operationally important: the unsubscribe policy (4.33 stars) and shipping address update flow (4.54 stars) close gaps that would otherwise require agent access to Shopify. The address update policy runs a pre-fulfillment check before acting, so changes only execute when the window is still open.

FAQ coverage across product, shipping, and policy

Alongside action policies, the AI handled 17,206 FAQ interactions across the period: product comparisons, sleep quiz guidance, trial period terms, in-home delivery service details, and warranty questions. The top knowledge sources by reference count are mattress product pages, the In-Home Setup FAQ, and the mattress comparison guide, reflecting the pre-purchase and early-post-purchase shape of Helix’s inbound traffic.

Results

72%

Chat containment

65%

Tickets resolved end-to-end

4.19

CSAT

Across 31,821 eligible conversations from January through May 2026, Kodif contained 22,838 (71.8%) without a customer-requested escalation, and fully resolved 20,663 (64.9%) without any human involvement. Total chat volume over the period reached 50,209 conversations, with 31,846 actively engaged sessions.

The AI executed 14,640 action-type policies autonomously (WISMO updates, damage resolutions, subscription changes, address corrections) while handling 17,206 informational FAQ responses in parallel.

CSAT across 4,098 rated conversations averages 4.19 stars. Of those rated sessions, 78.5% landed at 4 or 5 stars, with 71.0% at 5 stars specifically. The highest-rated action policies are shipping address updates (4.54 stars) and the unsubscribe flow (4.33 stars), both cases where the customer got a definitive answer and a completed action in a single interaction.

KODIF has now successfully implemented automations across ten 3Z Brands.

What's Next

Helix is extending the deployment in several directions:

  1. 1

    WISMO response specificity

    Improving order status detail for split-shipment scenarios, where mattress, foundation, and frame arrive on different carrier timelines

  2. 2

    Sales and discount policies

    Automating the logic for applying current promotional pricing (one of the higher-volume agent workflows during sale events)

  3. 3

    Damaged item policy depth

    Expanding resolution paths for complex damage scenarios beyond current routing logic

  4. 4

    KB gap coverage

    Closing the knowledge gaps identified by the 1,203 KB Miss smart-tag events, each a measurable drag on satisfaction scores

  5. 5

    Multi-brand expansion

    The 3Z portfolio includes multiple brands with shared tooling; Kodif coverage across Brooklyn Bedding and other brands is on the roadmap

The takeaway for DTC brands with high-ticket products

The automation challenge for a brand like Helix is not FAQ deflection. It is action completion. Customers buying high quality mattresses expect the same clarity on their order that they would get from a knowledgeable human agent. The lever is not answering faster; it is answering completely and taking the action the customer came in to get done. A policy layer that can look up the order, run the logic, and write the result back changes the economics of the support team, not by removing people, but by reserving them for the problems that actually need one.