Expert roundtable: What actually breaks enterprise AI initiatives

Every enterprise AI pilot starts the same way. A vendor demo, a promising proof of concept, then a rollout that quietly stalls.

Leadership wants to know why. The instinct is to blame the model: wrong tool, wrong vendor, needs a bigger budget for a better one.

That’s rarely the actual problem.

During a CMSWire roundtable in May 2026, “From content sprawl to competitive advantage: A knowledge operations roadmap”, practitioners from Eli Lilly, AMD and Brightspot compared notes on what actually breaks AI initiatives inside large, regulated enterprises.

Three companies, three different industries. Same diagnosis.

Will Chu, SVP of Client Engagement at Brightspot, said his team sees the pattern “very early in an engagement,” the moment a customer says their AI initiative isn’t delivering.

Ask Chu what actually breaks these engagements and he doesn’t hedge: “A common challenge folks have is really just duplication of content, and finding the true source — then establishing the frameworks to have that clean room, so to speak.”

The content lives in five different systems. Half of it’s duplicated. Nobody owns whether any given version is current. That’s the actual bottleneck, and it shows up long before anyone gets to evaluate a model.

AMD’s Doug Brook made the operations case from his own seat: scattered content was already costing his team before AI entered the picture, in duplicated effort, inconsistent documentation and slower localization. Eli Lilly’s Nicki Usiondek raised the regulated-industry version of the same problem: when content quality is inconsistent, you can’t trust what your systems surface, model or no model.

The working definition that came out of the panel: findable, governed, consistently structured and living in one place. That’s what “AI-ready” content actually means: a foundation that holds, not a bigger model.

Doug Brook added that when content is consolidated, it is easier to avoid redundancy across your library and, while MCP does make it easier to connect to multiple platforms, you also have to factor in maintaining one MCP versus across several instances.

When leadership hears “we need to consolidate our content,” it tends to get treated like a vendor swap rather than examining their content foundation.

Same functionality, different logo, new contract to negotiate.

That framing undersells what’s actually happening. Brightspot’s own research on architecture failures found a similar root cause behind failed content management system (CMS) migrations. The organization behind the tooling matters more than the tooling itself.

Consolidation done right replaces myriad ungoverned sources of truth with one governed one. Which CMS wins is beside the point.

The number Chu points to from the panel: an average 30% drop in operating costs in the first year after consolidation, driven by less duplication; faster search and retrieval; lower governance overhead.

That’s real, board-level budget.

And it shows up before it ever hits a dashboard. “You’ll start to anecdotally hear from folks: ‘This is a lot easier than it was in the past, it’s much more streamlined, I can produce more content, I can get to other things in my day.’ Establishing that traceability and tracking is what helps you measure and see the results that come in,” Chu said.

The pattern isn’t unique to content teams. In Brightspot’s earlier reporting on developer knowledge management, Chris Westerhold, Sr. Director of Developer Experience and Platforms at HTEC Group, said one enterprise client was losing about $40 million a year to ineffective documentation, with developers burning 60% of their time hunting for answers instead of building. Ungoverned knowledge costs the same way no matter which team is stuck searching for it.

The most common question Chu says he hears from prospective customers: we know we need to consolidate, where do we even start?

His answer: start with a content inventory. Find out what you actually have, where it lives and who’s using it, before you commit to moving any of it.

That inventory takes longer than the org chart suggests. Chu said: “It’s not like you have a catalog or database somewhere that’s always up to date. There’s a lot of interviews and discussions; it’s very much networking. You’re talking to one person who says, ‘Oh, maybe it’s crowd — village knowledge,’ and then you have to talk to a bunch of other people to really understand that. So it takes more time than you think.”

Budget for that. It’s discovery work, not data entry.

That answer also explains a symptom a lot of content ops leaders will recognize. Official documentation exists, and teams still work around it anyway. Usually that’s a trust and findability problem. Teams can’t find the current version fast enough, or they don’t trust that it’s actually current. Fix findability and governance first, and the workaround habit tends to disappear on its own.

Once the foundation holds, the AI conversation changes.

Chu’s framing from the panel: agentic AI adds the most value on repetitive, structured tasks, surfacing content, recommending updates, routing a piece for review.

The assistant’s real job is to narrow the list of judgment calls before a person makes the final one. We have written before about where agentic AI adds real value for content teams, and the pattern holds here too.

The rollout should be phased, too. “It’s not about boiling the ocean — it’s finding folks who are increasingly proficient with AI and its use in their day-to-day operations, helping those folks become your early adopters, and finding critical use cases that are beneficial across your value chain,” Chu said.

Same logic applies once you’re past the pilot. “It’s not about applying AI everywhere — it’s about looking at your critical path, what’s necessary, and streamlining those areas,” Chu said. “Maybe it’s not a big bang all at once across your workflow — maybe it’s at certain stages and steps, and you learn as you go.”

Brook’s point about risk lands in the same place: automation makes mistakes at scale when the review step gets bolted onto the end of a workflow instead of built into it.

That’s also the case for rethinking how you measure content ROI once AI is doing some of the work. The math changes when a person is reviewing ten AI-drafted updates instead of writing ten from scratch.

The enterprises getting real value from AI right now are deliberate about it: which tasks go to the model, which stay with a person and why.

That’s a leadership decision. It decides where budget goes, who owns governance and what “AI-ready” actually means for your organization.

Chu ties that discipline directly to leadership, not IT: “Getting champions within the organization matters, both your peers and colleagues, but also from the leadership team. When it’s coming from leadership, it’s built into the OKRs and the program initiatives for the year, so you’re not really convincing them — there’s a broad understanding that this maturity needs to happen.”

And the metric that actually lands with executives isn’t the one you’d expect. “It’s not so much about headcount reduction — it’s repurposing people along your value chain so they can develop new product ideas, create more helpful content for customers, or keep content continually fresh,” Chu said.

If you already know you need to consolidate and you’re not sure where to start, an audit of your existing content and experience is a lower-risk first move than a full replatform.

One more thing Chu is clear about: this doesn’t stay fixed. “Even if you think you’ve really matured and achieved it, it’s going to keep evolving — your content isn’t static,” he said. “You’re going to have new throughput, new flows, new topics, new content that you’re going to continually put through the frameworks and processes you’ve established.” The foundation is a practice you keep, not a project you finish.

The model was never going to fix your content problem for you. Fix it yourself, and the model finally has something worth working with.