As part of Coveo’s Relevance 360 event in Atlanta, a group of AI practitioners, digital commerce leaders, and customer experience professionals gathered at the Georgia Aquarium for a candid afternoon panel discussion: “From POC to Production: Thrive with AI, Generative and Agentic Experiences.”
This wasn’t a room full of people trying to decide whether to adopt AI. These were teams already in production, already managing the messy realities of data quality, change management, and customer expectations — now trying to figure out how to push further and faster.
Two practitioners took the stage to share hard-won lessons and a clear-eyed view of where the AI journey actually leads once the pilot phase is over.
The “Good Enough” Trap
The panel opened with a question that landed differently depending on who was in the room: at what point did you realize your search wasn’t working — and that it was costing you?
For Andy Patel, VP of Marketing, E-Com & Product Development at Hercules Sealing Products, there was no single dramatic failure. It was a pattern. High-intent customers — people who had already found the product, who knew the part number, who were ready to buy — were still abandoning carts. Still calling in. Still leaving. His team had transformed the business from 15% online to 70% online over five years, yet the search experience wasn’t keeping pace.
The pattern Patel described has a name in the data. Coveo’s Website Search Readiness Crisis report, drawn from 213 B2B decision-makers, found that 78% of organizations rate their website search as “good” — yet 80% expend moderate-to-high manual effort just to keep it there. What organizations call “good” is, in practice, expensive mediocrity sustained through constant tuning.
What makes this especially dangerous is the perception gap across organizational levels. The report found that managers — who operate search day-to-day — were most likely to rate it positively. Directors, who actually implement improvements and manage maintenance resources, see the picture more clearly: 55% of those rating their search as “bad” are directors, compared to just 5% from the C-suite. Executives may believe search is adequate. The people responsible for building on top of it know otherwise.
Patel put it plainly: when his team saw high-intent customers still abandoning and still calling in, they recognized it wasn’t a UX issue — it was a revenue leakage issue. That realization is what triggered the search for a better solution, and ultimately led to their Coveo implementation.
Relevant reading: The Website Search Readiness Crisis: When “Good Enough” Search Meets AI-Era Expectations
What “Revenue Leakage” Actually Looks Like
Industrial distribution is an unforgiving environment for poor search. When a job site goes down, every hour of downtime carries a cost — which means the customers coming to Hercules Sealing Products’ platform are arriving with urgency, not patience. They need the right part, the right specification, the right fit. A search result that returns the wrong thing, or nothing, doesn’t just frustrate them. It sends them somewhere else.
Hercules had moved aggressively to digital, but its product data hadn’t kept pace. Much of the catalog dates back to the 1960s, and the engineers who built those parts didn’t document them with SEO or structured ecommerce data in mind. With only five developers on staff running a massively scaled web store, the team couldn’t solve the data problem through brute force.
The broader industry data reflects this challenge. The report found that only 32% of organizations measure content discoverability — whether users can actually find what they’re looking for — despite this being the core function search is supposed to serve. Meanwhile, 62% track engagement metrics like pageviews, and 49% track conversion rates, without diagnosing whether poor search is the root cause of underperformance.
For Patel’s team, Coveo became the connective tissue between their SEO investment and their conversion funnel — particularly in their retail marketplace, where new customers arrive via Google with no prior relationship with the brand.
That segment is now up 50% year over year, a trajectory that coincided directly with the implementation.
Building for AI on a Foundation That Can Hold It
Matt Lacy, Senior Principal Knowledge Engineer at Blackbaud — a SaaS company serving the nonprofit sector — came to the same conclusions from a different direction. His challenge wasn’t part-number search. It was AI-powered support deflection: getting customers to find accurate answers on their own rather than escalating to a live agent.
Lacy had followed KCS (Knowledge-Centered Service) best practices for over a decade, but had drifted from disciplined execution. When Blackbaud committed seriously to AI deflection, the lesson was immediate: good outcomes require good data, and good data requires intentional governance.
His team’s response was structural:
- Built a dedicated knowledge management team focused specifically on content strategy, not just maintenance
- Tasked those managers with regularly reviewing AI-generated answers — refining content to match brand voice, not just factual accuracy
- Used Coveo’s knowledge help functionality to make this accessible to non-technical staff — including how answers are chunked and surfaced
- Established a formal AI governance committee to ensure changes are documented, approved, and aligned with brand standards before reaching customers
The governance work is slow and, as Lacy acknowledged, sometimes frustrating. But it’s also what makes AI-powered deflection actually work at scale rather than just in a demo.
This reflects a gap the research captures sharply. The report found that 75% of organizations are running on platforms never architected to support AI workloads — and only 29% say grounding AI in their actual content is “critical,” even as 41% cite hallucinations as a top concern. Organizations fear the symptom but deprioritize the solution.
Lacy’s message to the room was direct: allowing the model to monitor the model isn’t optional if you want deflection to hold up. Someone has to own the quality of what the AI produces — and that ownership needs to be formalized, not assumed.
What’s Overhyped — and What Still Deserves Attention
One of the sharpest exchanges of the session came when moderator Priscila Garcia asked the panel to separate signal from noise: what in AI is genuinely overhyped right now, and what’s the unglamorous work that still matters most?
The boring work that’s actually essential
Both panelists landed on the same answer independently, and it was the same answer the research points to: data hygiene and governance aren’t exciting, but everything else is built on top of them.
Patel was direct about it: merchandisers at Hercules have to do, in his words, “a really, really strong job” of enriching product data — SEO tags, H1s, title tags, descriptions — so that customers searching for parts can actually find them. None of that work is glamorous. All of it is necessary. The team uses content gap data surfaced by Coveo to prioritize which products need enrichment most — feeding search signals back into product development.
Lacy framed governance similarly. Defining business outcomes before deploying AI, ensuring brand consistency across tools, building a formal approval process — these are the conditions that let AI scale without eroding trust. Without them, speed becomes a liability.
Relevant reading: 6 Data Cleaning Challenges Blocking Enterprise AI (& How Coveo Can Help)
What’s actually overhyped
- AI image and video generation: interesting in concept, but not yet delivering clear ROI in technical commerce contexts, according to Patel
- Chatbots in highly technical B2B environments: promising trajectory, but Patel noted he hasn’t yet seen it done well in markets like his
- Deploying AI because leadership demands it: Lacy pushed back on the pattern of mandating AI adoption without defining the problem, the outcome, or the ROI first — including the possibility that ROI might be time savings rather than cost reduction
The research validates the last point in particular. The report found that 50% of organizations are “exploring” agentic AI — but only 22% have made it a strategic priority and are actively implementing. That 28-point gap between exploration and execution isn’t hesitation for its own sake. It reflects organizations recognizing, as Lacy put it, that you have to peel back the onion to understand what you’re actually solving for before committing to a full rollout.
Lacy’s team validated this through practice: they ran a POC on a subset of their top products for a couple of months, brought back the results, and used that data to make the case for broader deployment. The discipline of proving it before scaling it is what made the broader rollout defensible.
Relevant reading: Data Retrieval: The Agentic AI Competitive Advantage
What Comes Next
Blackbaud: Personalization at scale
For Lacy, the next frontier is personalization. Blackbaud holds rich data about how customers use their platform — which features they rely on, how often, what their organizational goals are. The opportunity is to move from answering questions to proactively surfacing insights: not just resolving the problem a customer raises, but connecting that problem to a goal they’ve already shared with Blackbaud.
The example he gave: a customer comes in with a support issue, but Blackbaud knows they’ve recently moved into a leadership role and that one of their stated goals is growing recurring giving. Rather than simply resolving the ticket, the system could say — here’s how to solve your problem, and here’s how a slight adjustment to your approach could also move the needle on the outcome you’ve told us matters. That’s the difference between support and partnership.
Hercules Sealing Products: A full technology transformation
Patel described a simultaneous rollout that would be ambitious for any organization — ERP, search, CRM, and chat implementations all happening in the same period, alongside active acquisition discussions. The goal is to build infrastructure capable of supporting another doubling of the business, with technology that can scale alongside it rather than constrain it.
The timing is intentional. Right now, with major platforms going in together, there’s an opportunity to get the integrations right from the start rather than layering new capabilities onto a fragmented foundation after the fact. That’s a lesson the research makes clear applies well beyond Hercules: the organizations that fix their retrieval infrastructure now are the ones positioned to make AI actually deliver on its promise.
The Bigger Picture
The conversation at the Georgia Aquarium was specific — two companies, two industries, two sets of very concrete problems. But the pattern it described is broad. Coveo’s research found that 72% of organizations across industries are exploring conversational AI, while three-quarters of them operate on platforms never designed for AI workloads. Ambition is not the constraint. Infrastructure is.
What Patel and Lacy demonstrated is what it looks like to move past that constraint. Not by waiting for the perfect moment or the perfect platform, but by doing the foundational work — enriching the data, building the governance, validating through POCs — that makes everything else possible. The organizations that will thrive with AI are the ones that started treating search as infrastructure, not an afterthought.