-
One in four customer interactions are affected by repeated explanations, manual escalations, incomplete context and resolution delays, per aTalkdesk CX surveyreleased on Tuesday. The key culprits include disjointed knowledge management systems with incomplete data.
-
Nearly 30% of a human agent’s time each day is spent on “swivel-chairing” among systems or on data re-entry across multiple, disconnected tools versus directly assisting customers. Most organizations have not consolidated their data and information into a single repository that human and AI agents need to address customer issues.
-
Moreover, 94 percent of companies lack AI-assisted knowledge management, Talkdesk found. “Getting companies to a global [retrieval-augmented generation] system is much more challenging for them than initially thought,” Pedro Andrade, Talkdesk’s vice president of AI, told No Jitter. “It’s not a technical challenge. It’s a challenge of getting the data in. It needs to be labeled [and] the graphs need to be fed with metadata, so the system knows where to find the right answers in an accurate way.”
Fragmented data is one of the foundational challenges organizations must surmount before building an agentic AI system. Only 15% of the respondents had gone through the process of getting individual departments to bring their knowledge, documents, rules, PDFs, etc., into an enterprise knowledge layer which the agentic system can access.
“That is why 81% of organizations have less than 10 use cases implemented,” Andrade said. “They need to focus their energy not on building the agentic system, not on orchestrating them, but actually resolving the most foundational thing — accessing the data — [so that] the agent has enough to be autonomous, meaning it [can] reason over policies and rules, and then access the tools to get the job done.”
Talkdesk defined orchestration as autonomous cross-departmental execution — i.e., the orchestrator agent knows what to do and what is required to solve a given issue. This involves delegating tasks to multiple, specialized sub-agents associated with different back-end systems like billing or procurement, for example. However, this requires the organization to build direct integrations to those systems or use MCP to knit them together. And while the cross-organization data can be used to develop the context for each individual customer interaction, that data must be made AI-ready.
“This is all about the metadata you need to put on your knowledge. The knowledge is still the same, but the metadata will help AI retrieve the right answer accurately,” Andrade said. “If you don’t do this, two things can happen: you don’t have enough data, and the data is ‘okay-ish’ in terms of semantic space, so it should work, but you can’t ensure that the quality of the outcome is going to be great. So, if it works without metadata, you just got lucky.”
