Enterprise CX and IT leaders face an ongoing architectural challenge when deploying AI agents. Customer service workflows require real-time access to information such as operational data, product documentation, policies, and legal guidelines. But connecting AI agents to disparate repositories often creates operational friction or information silos. Metrigy’s data shows that relying on a hybrid knowledge model is the top method for customer service and support use cases.
When evaluating primary relationships between AI agents and knowledge repositories, the 759 organizations studied globally in our<a href="https://www.metrigy.com/product/ai-technology-foundation-strategy-2026-27/” rel=”nofollow noopener” target=”_blank”>AI Technology Foundation & Strategy: 2026-27research report four structural choices:
-
Hybrid systems account for 47.5% of deployments. In this model, companies use CX platform-native tools for common questions and federate to external documentation for complex technical or legal queries.
-
Specialized third-party platforms — i.e., best-of-breed systems — comprise 24.8% of deployments, pushing data directly into the CX platform for AI agent use
-
18.8% of organizations use platform-native systems, relying entirely on built-in knowledge tools from the CX vendor
-
A small portion, 7.9%, use federated or mesh architectures, allowing AI agents to retrieve knowledge directly from original sources without a central CX knowledge base
Additionally, the study shows that the hybrid architecture results in positive business outcomes. Within the success group, determined based on measured changes to business metrics from AI’s use, 56.9% of organizations utilize a hybrid knowledge architecture, compared to 36.8% of the non-success group.
A hybrid model balances immediate retrieval speed with access to deep technical context. Frontline customer service inquiries often involve simple transactional questions. CX platform-native tools handle these routine requests efficiently. When a customer inquiry requires specialized technical steps or compliance verification, the hybrid framework federates the query to external documentation without duplicating data into the primary contact center repository.
System integration plays a critical role in overall operational visibility. Across all surveyed organizations, nearly 75% report that their systems make it easy to find and connect information from different parts of the company. The importance becomes especially clear when evaluating causes of AI project failure. Data fragmentation is the leading culprit for 45.3%.
Connecting AI agents to knowledge repositories requires optimizing unstructured content, such as manuals, transcripts, and policy documents. Companies employ several techniques to prepare unstructured data for AI ingestion:
-
48.0% of organizations automate content optimization using a third-party knowledge platform
-
46.5% rely on built-in AI tools within their CX platform for auto-indexing
-
41.7% process unstructured content through custom middleware before feeding it to AI engines
While third-party knowledge platforms and native CX tools remain widely used overall, custom middleware demonstrates a wide performance gap between organizations. Among the success group, 53.6% utilize custom middleware for data processing, compared to 28.8% of the non-success group. Custom middleware allows IT teams to clean, structure, and enrich data prior to retrieval, enhancing accuracy for complex customer interactions.
To establish an effective knowledge architecture for AI agents, CX and IT leaders can follow several practical steps:
-
Evaluate current knowledge query types to separate high-volume routine inquiries from specialized technical requests.
-
Implement native CX tools for basic knowledge retrieval while setting up federated links to authoritative external databases.
-
Invest in custom processing middleware to structure unstructured enterprise files before vector indexing.
-
Establish cross-functional data governance to ensure federated sources maintain proper access controls and current information.
