The growing disconnect between AI investment and frontline trust is becoming more evident across customer service organizations, despite rapid adoption and daily use.
UJET’s latest report reveals that agents remain wary of AI’s accuracy, context, and real‑world usefulness, as 93% do not fully trust AI outputs at face value. despite it now being embedded in nearly every customer interaction.
As a result, AI has been deployed faster than the data foundations, workflows, and system architectures required to make it truly reliable at the frontline.
Speaking with CX Today, Vasili Triant, CEO of UJET, argues that agents don’t distrust AI because they resist change, but because poorly designed, data‑fragmented systems haven’t earned their trust.
“The 93% verification rate reflects a system design problem, not a behavior problem,” he explained.
“When AI is layered onto fragmented data environments, the outputs reflect poor AI implementation.”
Why Accuracy Breaks Down at the Frontline
In enterprise environments, AI tools are frequently layered onto fragmented data sources, where customer records, interaction histories, and real-time signals are split across multiple systems.
As a result, this creates an incomplete operational picture, likely leading to structurally inaccurate outputs, particularly when models generate responses without synchronized context.
From here, the architecture issue can become more pronounced when AI systems lack access to real-time customer context, resulting in models that produce partially outdated or inconsistent responses that don’t align with the present situation.
That gap increases the likelihood of hallucination and misalignment between recommendations and customer needs, as 15% of agents agree that real-time AI recommendations are unreliable or inaccurate, as well as 54% saying AI is helpful but lacks sufficient context and depth.
“When AI does not have access to the latest real time data on the customer’s latest touchpoint, the risks for AI hallucinating responses skyrockets,” Triant continued.
“Agents have seen enough incorrect answers and context-free suggestions to know that blind trust creates risk for the customer experience.”
As a result, many organizations interpret this as a behavioral issue requiring more user training or compliance when the underlying constraint is architectural friction.
When verification requires excessive time or effort, users may typically default to skepticism because the cost of error is high in live customer interactions.
“The goal shouldn’t be eliminating verification,” explained Triant.
“It should be making it effortless. When an agent can validate an AI recommendation in two seconds instead of twenty, quality stays high and efficiency finally becomes real.”
How Early Assumptions Shaped Today’s Friction
Secondly, the agent-AI trust gap is also shaped by how the technology was initially positioned and deployed inside enterprise systems, with the common early industry narrative focusing on automation, deflection, and cost reduction.
From here, that framing encouraged organizations to treat AI as a replacement layer placed on top of existing operations, meaning that AI tools were introduced into environments that were already operationally fragmented.
Instead of redesigning the underlying workflow, the result was limited improvement in day-to-day effectiveness, even when the models themselves performed well in isolation.
In fact, around 78% of agents reported that their AI tools are not transformative, while 81% are required to manage more than four tools during a single interaction, and nearly 20% handle seven or more tools at once.
Despite these challenging conditions, 93% agreed they could still do their job without AI, implying that the tools are not yet integrated into a coherent workflow that changes core execution.
However, success metrics are still frequently defined in terms of reduced staffing rather than improved workflow performance, and do not address the structural inefficiencies agents face.
“For years, the industry framed AI’s value around deflection and headcount reduction,” he explained.
“That pushed investment toward the wrong outcomes and set unrealistic expectations for ROI.”
With AI originally frequently layered onto workflows that were never designed for real-time intelligence or cross-system coordination, many agents today still move between disconnected tools, re-entering or reconciling information across systems that lack a unified context.
Even accurate AI outputs lose value when they cannot be applied directly within the flow of work.
“Enterprises also share responsibility. AI is often deployed on workflows that were designed decades ago,” Triant continued.
“When agents are still juggling five or six disconnected systems, even good AI struggles to deliver value.”
This enables enterprises to shift the ROI discussion toward a meaningful measure of AI performance, evaluating how effectively it reduces friction for existing teams for fewer context switches, faster resolution paths, and more consistent decision support during live interactions.
The Emotional Cost of Failed Self‑Service
As a result, the breakdown in trust around AI-powered self-service is largely a result of how these systems are structured, as many organizations focused on containment rather than resolution.
When many implementations were designed to keep customers within automated flows for as long as possible, this technique optimized for deflection rates over successful outcomes, creating a structural gap.
When a tool is capable of answering questions but struggles to execute changes, it pushes unresolved cases forward with increasing frustration attached to them, with roughly 65% of customers reporting frustration when having to repeat information after moving from AI to a human agent.
In fact, 14% of agents say they now handle more emotionally charged interactions as a direct result of failed self-service, as AI is already present in more than three-quarters of customer interactions for 75% of agents.
“Most self‑service is designed around containment – keeping customers away from humans,” said Triant.
“But customers aren’t trying to avoid agents. They’re trying to solve a problem, and they feel that disconnect immediately.”
Moreover, when self-service systems cannot take meaningful action, they tend to pass unresolved issues downstream, often after the customer has already repeated steps or provided information multiple times.
“When self-service can only answer questions and can’t execute actions, it escalates the most emotionally charged cases,” he explained.
“Customers arrive at agents frustrated and forced to start over.”
However, a more effective architecture treats self-service as part of a continuous, stateful journey where context is preserved across transitions so agents receive the full history of what has already been attempted.
For the customer, this reduces repetition friction and allows the human interaction to begin at the point where automation left off.
“Self-service should be the first chapter of a continuous journey, not a dead end,” emphasized Triant.
“If escalation happens, the agent should inherit full context so the experience feels like a continuation, not a restart.”
Ultimately, the disconnect between AI adoption and frontline trust signals that enterprise architecture has been incapable of keeping up with CX ambition.
As AI becomes embedded across customer journeys, its customer impact will be dependant on how effectively it is grounded in real‑time data, integrated into coherent workflows, and aligned with how agents work.
If an AI is able to reduce cognitive load, preserve context, and remove operational friction, the customer trust will naturally follow.
