Scaling AI-Driven Customer Service Without Losing Customer Trust

AI is cutting customer service costs — but it may be accelerating organizational risk even faster. The real executive problem isn’t whether to deploy AI; it’s whether the enterprise has built the readiness to withstand the moment it fails in front of a customer.

Stanford’s RegLab documented that general-purpose AI chatbots hallucinate between 58% and 82% of the time on legal queries — and that even purpose-built legal AI tools hallucinate in at least one in six benchmark cases. The Stanford HAI 2026 AI Index reports that hallucination rates across 26 leading models now range from 22% to 94%, and that documented AI incidents reached 362 in 2025, up from 233 the year before.

The CFPB warned that when customer service chatbots fail, they not only break customer trust but carry the potential to cause widespread harm — and that financial institutions risk active legal liability when poorly designed chatbot technology causes consumers to select the wrong product, misunderstand fees, or lose access to dispute handling. The FTC and three other federal agencies jointly committed in December 2023 to vigorously enforce existing law against AI tools that produce harmful outcomes for consumers.

A global survey of more than 1,000 consumers across six countries by COPC Inc., an independent customer operations standards body, found that satisfaction climbs above 90% when AI fully resolves a customer issue without further steps — but when AI fails to resolve the issue, a brand’s Net Promoter Score can plunge by as much as 70 points. The same research identified the handover from AI to a human agent as the most consistent point of failure across all markets studied — not the AI model itself, but the workflow design behind it.

Robert Rose, Senior Director of Customer Experience at Adobe, joined Emerj’s Matthew DeMello on the AI in Business Podcast to map the maturation curve of AI in customer service and to outline what enterprises must get right before they can safely scale these capabilities to the customer.

This article examines three critical insights from Adobe’s Robert Rose on how enterprises can deploy generative AI in customer service without outpacing their organizational readiness:

Trust threshold as a deployment map: Customer willingness to accept AI scales inversely with interaction risk, and deployment sequencing must reflect that reality.

Deterministic AI foundation as the prerequisite for generative personalization: Predictive AI maturity provides generative systems with the grounding to personalize accurately and defensibly.

Escalation design as the measure of service AI maturity: How AI transfers to a human agent determines whether a failing interaction recovers or compounds into a brand trust event.

Episode: Enhancing Customer Engagement with AI-Driven Solutions – with Robert Rose of Adobe

Guest: Robert Rose, Senior Director of Customer Experience at Adobe

Expertise: Customer Success, Technical Support Operations, Knowledge Management, Service Transformation

Brief Recognition: Robert Rose leads enterprise paid support for Adobe’s Creativity and Productivity Solutions business, where he created and scaled the company’s global paid support program for Digital Media solutions. Across a career spanning leadership roles at Adobe, NICE, EMC, and Symantec, Rose has led large-scale support and customer success transformations, including global technical support organizations, knowledge management systems, and customer experience initiatives. Beyond industry leadership, he served as Adjunct Faculty at Utah Valley University, where he received a Teacher of the Year award, and holds a Bachelor of Science in Business from the University of Phoenix.

Trust Threshold as a Deployment Map

Customers do not trust AI evenly. Their comfort expands and contracts based on the stakes of the interaction. People have been interacting with AI systems for years, often without realizing it, and in low-stakes situations, this creates very little friction. But the moment the consequences rise — a billing dispute, a financial decision, a healthcare concern — trust tightens, and tolerance narrows.

This creates a natural deployment map. Some interactions are ready for automation. Others require supervision. Some must remain human-led until trust is earned.

The sequencing is not determined by what the technology can do, but by what the customer is willing to let it do. Rose emphasizes that this curve is moving, but not on the enterprise’s schedule. Customers decide when trust expands, and deploying ahead of that curve invites frustration and churn.

Embedded in this trust question is a strategic decision most enterprises are quietly making right now — whether to disclose that a customer is talking to AI at all. Rose frames this not as an ethical question but as a brand and operational one:

“Companies are actually considering not saying that it’s a bot, and just saying we’re going to let you interact with this thing until we sense that it’s not working, and then we’ll send you to a human. We won’t tell you — but that’s really up to the company.”

Robert Rose, Senior Director of Customer Experience at Adobe

The implication for senior leaders is direct. The disclosure decision is no longer a default — it is a choice with consequences in both directions. Disclosing AI sets expectations and protects the brand when the system fails. Not disclosing raises the stakes of every failure because the customer feels misled, not just underserved. Executives need a deliberate position on this, not an inherited one from the vendor implementation.

The deployment segmentation that follows from Rose’s framework:

Low-risk transactional interactions: AI can operate autonomously now

Medium-risk interactions: AI with human oversight is the current standard

High-risk regulated interactions: human-led with AI in an assist role until trust is established

Deterministic AI Foundation as the Prerequisite for Generative Personalization

Rose draws a distinction most enterprise AI roadmaps miss: predictive and generative AI are not interchangeable. They are sequential. Predictive systems follow rules and patterns; generative systems produce responses. Without a reliable deterministic layer beneath them, those responses have no grounding.

The governance shift this creates is significant. In the era of predictive AI, companies programmed the bot — if this question comes in, this is how you answer it. Generative AI removes that constraint entirely. The system decides what to say based on the data it has access to. That is where the capability expands — and where the risk enters.

He explains the failure mode directly:

“They make stuff up. Now, they don’t really make it up — they just found it somewhere. And they misinterpreted it, and they recommended it, and it wasn’t in the right context.”

Robert Rose, Senior Director of Customer Experience at Adobe

Where generative AI is already delivering value — even under human supervision — is in response personalization. The Adobe executive describes a capability that is available today and deployable now with the right oversight model:

“It can personalize a response based on the profile, the recent interactions — is this customer traditionally angry, or are they angry now? If so, I’m going to put some calming words in there. Generative AI can do all that today.”

Robert Rose, Senior Director of Customer Experience at Adobe

That capability — real-time sentiment-aware response generation — is the near-term value case for generative AI in customer service. But Rose is explicit that it requires adult supervision. His recommended path for organizations that want to capture that value without the liability:

“The best thing for these companies to do probably is to experiment internally and to let the human eyes look at it and say, ‘ Hey, that’s really good, and then utilize what you get and temper it before delivering to the customer.

Robert Rose, Senior Director of Customer Experience at Adobe ​

The sequencing checklist Rose implies for senior leaders:

Predictive AI is producing outputs that the organization can defend

Customer data is structured and stable enough for generative models to draw from accurately.

Internal generative outputs are reviewed before any customer contact.

Legal and risk teams are inside the deployment process, not handed a finished system.

When legal pushes back, Rose treats it as a signal that the deterministic foundation is not yet strong enough to support generative personalization at scale. That resistance is diagnostic, not obstructive.

​

Escalation Design as the Measure of Service AI Maturity

Rose reframes one of the most misread signals in customer service AI. When customers repeat “representative, representative,” the failure is not the model. It is the workflow. The system either failed to detect frustration early enough or handed off the customer in a way that forced them to start over. Most organizations respond by improving the model. Rose argues that the model is rarely the problem. The real test is what happens next.

What makes early escalation possible is emotion detection — a capability Rose identifies as actively improving and underutilized. AI systems are increasingly able to sense customer frustration before the customer explicitly asks for a human. That signal, acted on early, is what separates a recovered interaction from a compounding one. Most enterprises have not yet built the workflow triggers to act on it.

The escalation design questions every CX leader must be able to answer:

Does the system detect frustration early enough to prevent escalation before the customer demands it?

How completely is the customer’s context — history, sentiment, prior inputs — transferred to the agent?

Does the agent begin informed, or does the customer repeat themselves?

Are human agents trained to receive transferred context, or to restart the conversation?

That last point is where most implementations fail silently. The technology can transfer context. The agent’s behavior often does not use it. That is a training and change management problem, not a technology one — and it sits entirely within the enterprise’s control.

COPC Inc. found that when AI fails to resolve an issue, Net Promoter Score can drop by as much as 70 points — and identified the bot-to-human transfer as the most consistent failure point across every market studied. The model was not the variable. The handoff was.

Rose’s rule of thumb: measure AI maturity not by what the model achieves when it works, but by how cleanly the system recovers when it does not