Overslaan naar belangrijkste inhoud

Why Zendesk built a unified retrieval platform for AI

Learn how Zendesk's Unified Retrieval Platform provides a shared foundation for AI knowledge, ensuring consistent answers across AI agents and human support.


Daria Bogusz photo

Daria Bogusz

Senior Product Manager at Zendesk

Laatst gewijzigd 19 november 2026

Blog Assets_2026_Bridge Campaign_H2 - 6

Two support experiences can have access to the same company knowledge and still provide conflicting guidance. A customer asking whether they can cancel a subscription after renewal might hear one version of the policy from an AI agent, then receive different guidance when the conversation reaches human support. The difference may come from which sources each experience retrieves, before either one responds.

As Zendesk expanded AI across customer service, we made an architectural choice: invest in a shared retrieval platform rather than building a separate retrieval system for every experience.

The result is Zendesk's Unified Retrieval Platform, designed to make finding relevant knowledge a core capability of our AI platform rather than something every product must build and improve independently.

What is unified retrieval?

Unified retrieval is an approach in which multiple applications use common infrastructure to find and rank relevant knowledge.

Instead of maintaining separate retrieval systems for each AI-powered experience, a unified retrieval platform provides a consistent way to access information across supported sources. Retrieval happens before an application generates an answer, recommends a next step, or presents search results.

Why Zendesk made retrieval a platform capability

Building retrieval into an individual AI product can help a team move quickly. But as more experiences emerge, independent implementations introduce complexity.

An AI agent might rely on one approach to finding knowledge, while an agent-assistance tool uses another. Each may rank information differently, require separate work to support new sources, or need its own improvements to retrieval quality.

For organizations, that fragmentation can surface as inconsistent information across customer interactions. For engineering teams, it means solving similar retrieval challenges repeatedly.

The Unified Retrieval Platform provides a common retrieval API that multiple AI experiences can use to access relevant information. Improvements to retrieval can be developed centrally and made available to the applications that depend on it, rather than implemented separately within every product.

Shared infrastructure doesn't require every experience to retrieve information in the same way. Each integrated experience can select supported knowledge sources, filters, and retrieval methods, subject to applicable access permissions. The available options depend on the integration.

This is the architectural advantage: a common foundation that can be improved centrally, with the flexibility to support different products and use cases as Zendesk's AI platform evolves.

iagram showing how Zendesk's Unified Retrieval Platform retrieves relevant knowledge from internal and external sources for Help Center, AI agents, and Copilot.

One retrieval foundation across the Zendesk experience 

Zendesk's AI experiences serve different purposes, even when they're supporting the same customer issue. A question about canceling a subscription might begin in self-service, continue into an AI agent conversation, and eventually reach a human support agent.

In the Help Center, Quick answers can provide an answer grounded in relevant policy information. An AI agent can use retrieved knowledge to inform its response and next steps. If human assistance is needed, Copilot can bring supporting information into the agent workspace.

Those experiences may need different sources or operate under different permissions. Where they're integrated with the Unified Retrieval Platform, they can use the same retrieval infrastructure without being limited to identical requests or retrieval behavior.

Diagram illustrating how Zendesk Copilot's Auto assist uses Help Center content and connected external knowledge to generate agent reply suggestions.

Inside unified retrieval: From question to relevant knowledge

Behind each of these experiences is a request for relevant knowledge. In our example, the cancellation policy might live in a help center article, subscription documentation, or a connected knowledge source.

1. The application requests knowledge

When the customer asks a question, the AI experience sends a retrieval request.

Along with the question, the request can specify relevant conditions such as brand, language, content source, or other filters. Applicable access controls help determine which information is available to the requesting experience.

For example, a customer-facing experience may be limited to public policy information, while an agent-assistance tool may also be authorized to retrieve internal guidance.

2. The platform retrieves from supported knowledge sources

The platform can retrieve relevant information from supported Help Center content and ingested external knowledge, including content brought in through Knowledge connectors, web crawlers, and other supported ingestion methods. Which sources are available depends on the integration and applicable access controls.

This is particularly valuable for organizations whose knowledge extends beyond a traditional help center into documentation, business applications, and other systems.

3. Retrieval methods find and rank relevant information

Finding the right policy isn't always as simple as matching keywords.

A customer might ask, “I was charged for another year. Can I undo it?” while the relevant article is titled “Subscription cancellations and refunds after renewal.”

Keyword search helps identify information containing specific terms. Semantic retrieval can find content based on similarity in meaning, even when the words differ.

The platform supports keyword, vector-based semantic, and hybrid retrieval. Hybrid combines keyword and semantic results to rank relevant information and is the default approach. Integrated experiences can select another supported retrieval method when their needs differ.

Diagram comparing how self-service, AI agents, and Copilot use shared knowledge retrieval to address a subscription cancellation question.

4. The retrieved knowledge is returned to the application

The platform returns relevant information to the experience that requested it.

An AI agent may use that information as context when generating a response. Copilot may use it to support a human agent. Help Center Quick answers may use it to generate an answer grounded in the retrieved sources.

Retrieval determines which supporting information is returned. The consuming experience determines how that information is used or presented.

What unified retrieval changes

For the cancellation example, two experiences integrated with the Unified Retrieval Platform can retrieve the same supporting policy sources when their queries, retrieval methods, source selections, filters, and access conditions match. Each experience may then use that information differently, but both draw on the same retrieval foundation.

That doesn't mean the platform automatically corrects outdated or conflicting knowledge. Organizations still need to maintain accurate information, and different applications may have different permissions.

What changes is the infrastructure behind those experiences. Retrieval improvements can be made at the platform level and benefit multiple applications, while each experience retains the flexibility to use that knowledge according to its needs.

Building AI on a common knowledge foundation

As AI becomes part of more customer service interactions, finding the right knowledge becomes increasingly important. Zendesk’s architectural choice is to make retrieval a shared capability while preserving flexibility for different experiences. That creates a common place to improve how relevant knowledge is found, while each integrated product determines how to use it.

Frequently asked questions

Daria Bogusz photo

Daria Bogusz

Senior Product Manager at Zendesk

Daria Bogusz is a senior product leader with more than 12 years of experience building technology products. At Zendesk, she has worked on AI agents, semantic search, and retrieval, helping shape how customers find answers and resolve issues.