The pilot phase is over. The stakes are real. At Coveo’s Relevance 360, enterprise leaders from Deloitte, UKG, and Lexmark gathered around a question that’s getting harder to avoid: what does responsible AI adoption actually look like in production?
This wasn’t a conversation about whether AI belongs in the enterprise. That question is settled. The more urgent question is, what happens when generative and agentic experiences hit production and run into the realities of fragmented content, inconsistent answers, security requirements, rising customer expectations, and pressure to prove business value?
The organizations making real progress are the ones bridging the “wow-trust gap” of AI by treating trusted AI retrieval, governed knowledge, and disciplined deployment as the foundation.
“The wow-trust gap is the space between AI that looks impressive and AI that earns the right to be used in enterprise service because it is grounded, consistent, secure, and measurably valuable.”
The Trust Problem Starts Earlier Than Most Teams Think
Priscila Garcia, SVP of Global Account Management, Coveo, set the frame: AI without trusted search and retrieval becomes a liability. That isn’t because large language models are inherently flawed. It’s because they do not know your business unless they are grounded in the right enterprise content, retrieved securely, in real time.
For service leaders, the risk surfaces as inconsistent answers, rising case volumes, and a harder time proving ROI. For digital leaders, it unveils itself as weak discovery, low confidence in generative experiences, and a widening gap from modern search expectations. For workplace leaders, it shows up as lost productivity, frustration, and employees relying on institutional knowledge to get basic work done.
Trust is an operational requirement.
What “We Can’t Find Anything” Actually Costs
Anita Chu, CIO at Deloitte Canada, brought that point into sharp focus.
Deloitte was replacing an intranet platform that was more than a decade old. When Chu’s team asked employees what mattered most in the new experience, the answer was not design. It was findability.
To pressure-test that feedback, she ran the search herself and got back content from 1999: outdated policies, stale procedures, and information no one should still be using.
That moment clarified the problem. The issue was not cosmetic. It was productivity.
For a firm of more than 15,000 people, the cost of bad discovery compounds fast. Employees waste time searching across systems, rely on coworkers to point them in the right direction, and sometimes act on the wrong information altogether. That is not just frustrating. It slows the business down.
What Deloitte needed was a trusted, enterprise-grade foundation that could search across a complex environment, respect security requirements, scale across systems, and get to value quickly. Chu’s evaluation criteria cut straight to what actually matters: strong security, scalability, federation across multiple knowledge stores, and a partner that could support expansion over time without turning the project into a long, fragile rebuild.
That is a useful corrective for any workplace or IT leader under pressure to “do something with AI” right now. The first question is not what model to add. It is whether people can reliably find the right information in the first place.
“Relevance Before Generation” Is More Than a Tagline
If Deloitte’s story was about the cost of fragmented internal knowledge, UKG’s story was about what happens when generative experiences move into high-stakes service environments.
Alison Brotman leads customer experience intelligence at UKG and described a support environment where accuracy matters deeply. UKG’s customers are often dealing with HR and payroll issues, which means the questions are nuanced, unscripted, and often urgent. Traditional IVRs are poorly suited to that kind of interaction.
UKG paired its conversational voice platform with Coveo’s Passage Retrieval API so every answer would be grounded in ranked, permission-aware knowledge drawn from a content ecosystem of more than 60,000 pieces of content. Brotman made the logic explicit: in voice, you do not want a list of links. You want the right sentence. You want the exact right answer, delivered fast.
That architectural decision reflects one of the sharpest ideas from the event:
relevance before generation
Generation is the layer customers see. Relevance is the layer that determines whether the answer deserves their trust.
UKG also showed what disciplined deployment actually looks like. Before going live, the team ran proofs of concept with multiple vendors, had subject matter experts review hundreds of responses, piloted with 1,200 customers, and iterated for months before moving to full production.
Governance was foundational. As Brotman put it, in HR and payroll, AI cannot just look impressive. It has to be right.
Search Is Now a Strategic Infrastructure Decision
Herb Toews, Sr. Manager of Customer Service Enablement at Lexmark brought the digital experience angle into the conversation.
A year ago, Lexmark’s search experience depended on exact term matching — users only got results when they happened to use the right words. This experience was out of step with how people now seek information: conditioned by generative tools to expect natural-language prompts and concise summaries.
That gap matters more than many organizations realize.
Search is no longer a background website feature. It is one of the most important layers in the digital experience because it shapes how customers find products, support content, and answers that influence next steps and buying decisions. Lexmark’s team also recognized something many organizations learn too late: building and maintaining search internally carries hidden costs. Upfront affordability is one thing. Ongoing tuning, maintenance, and optimization are another.
Lexmark was pragmatic. The team first deployed on an authenticated field and partner portal, where high-frequency users would quickly notice if answers were wrong or hallucinated. That made the rollout measurable and the feedback loop credible.
The early signal was strong: Lexmark saw an approximately 90% reduction in zero-result queries.
What the Agentic Era Changes
Laurent Simoneau, Coveo’s CEO closed by reframing the trust gap: it is not fundamentally a model problem. It is an enterprise retrieval problem.
Enterprise content is fragmented, secured, and constantly changing. That means conversational and agentic systems are only as dependable as the retrieval layer beneath them. As more organizations adopt personal assistants and agentic platforms, the requirement for accuracy gets higher, not lower, especially when systems are expected to do more than answer questions and begin executing tasks.
Trust is currency in the agentic era because every experience now carries compounding consequences. A weak answer does not just disappoint a user. It erodes confidence in the whole system. A grounded answer does the opposite. It creates the conditions for self-service success, faster resolution, stronger employee productivity, and more credible digital experiences.
In the next phase of enterprise AI, that is the most important distinction of all.
Watch the full session and find an upcoming Relevance 360 event near you.