Agentic AI in customer service: the edge is no longer the model, it's the data architecture

By Emerj (Emerj Artificial Intelligence Research) · July 1, 2026. It is worth clarifying up front something the article itself notes in fine print: it is content sponsored by Dialpad, a communications and AI platform for customer service. That does not invalidate the data it provides, but it does require reading the statements from its executives —Craig Walker, CEO of Dialpad, and Shezan Kazi, its head of AI transformation— for what they are: the view of the party selling the solution, not an independent verdict. The external figures, on the other hand, are substantial and deserve attention in their own right.

The facts: the U.S. Government Accountability Office (GAO) documents that federal agencies allocated nearly $4 billion to call center operations over a five-year period, with associated telecommunications infrastructure exceeding $30 billion. In a single program cycle, close to 10 million calls were recorded, with wait times often exceeding an hour. These are figures that portray a structurally overloaded system, not a one-off efficiency problem. At the same time, Stanford's Institute for Human-Centered AI estimates that generative AI reached around 53% population adoption in three years, a faster pace than the PC or the internet. But that same Stanford warns that even domain-specialized AI systems can hallucinate between 17% and 33% of the time, a figure that should temper any excessive enthusiasm about deploying these systems in regulated banking, health or insurance workflows.

The article's central argument, beyond the obvious commercial interest, has substance: most companies believe they know where the friction with their customers lies, and real conversation data systematically proves them wrong. The executives cite cases where managers expected one problem (flight changes, password resets) and analysis of six months of historical interactions revealed that the real volume came from entirely different matters. From there follows the thesis that the true bottleneck is not the model's capability —which, according to the text itself, 'has already arrived'— but the architecture: fragmented systems where the customer repeats their story every time the conversation switches channels, and legacy processes on which the AI merely observes instead of executing.

This connects directly with something we have already noted at Zendoric when analyzing AI's impact by sector: in banking, insurance and business administration, the first thing to disappear is repetitive administrative work, while roles involving judgment, compliance and customer relationships survive. Call centers are the textbook example: billions of dollars in infrastructure and staff dedicated to tasks that, according to the article's own sponsor, are mostly deterministic and automatable when redesigned as machine-executable steps —credential validation, eligibility checks, coverage rules. The piece also documents an interesting mechanism: AI does not replace the human agent, but acts as a filter (triage) that resolves routine matters and hands off ambiguous or sensitive ones with full context. It is a reasonable design, but the article provides no results data from real Dialpad customers beyond quotes from its own executives, which is a relevant limitation of this kind of sponsored content.

Our take: the 17-33% hallucination rate figure for domain-specific systems is, paradoxically, the most honest and most important figure in the article, and it comes from an independent source (Stanford), not the sponsor. It is a reminder that deploying agentic AI in regulated workflows —health, insurance, banking— remains an exercise in risk management, not simply 'plugging in the model.' In the short term, this means thousands of customer-service positions in call centers will continue to be reduced or transformed, with the social friction that entails, while companies invest in redesigning processes —a slow and costly task of organizational engineering, not a simple technological upgrade. But in the medium term, the underlying promise aligns with our thesis of abundance: if the systems manage to reliably handle the routine and free up human agents for nuance, empathy and ambiguous cases, the net result is less time lost in hour-long queues and more human capacity devoted to what truly requires judgment. The risk, as always, is that the promise of 'human augmentation' is used as a smokescreen for staff cuts without the investment in process redesign that the article itself identifies as a necessary condition for AI to work accurately.