On-Prem's Comeback, Qualtrics' New Playbook, and Why AI Needs Context, Not Just Data

Enterprise AI is outgrowing the model and the chatbot. That's the thread running through Episode 139 of ConstellationTV, where the conversation keeps landing on the same question from different angles: once AI has to actually operate inside the business, across infrastructure, customer experience, and decision-making, what does it need to work reliably?

The Big Debate: Is On-Prem AI Actually Making a Comeback?

Mike Ni moderated a debate between Holger Mueller and Esteban Kolsky on one of infrastructure's oldest questions, reopened by AI: does it run on-prem or in the cloud?

Mueller pointed to strong recent hardware numbers from Dell and HP as evidence of a real on-prem renaissance, but argued the pull toward public cloud is still structural. His analogy: running AI on-prem today is like a company generating its own electricity a century ago. It's possible, but the economics of elasticity eventually win. He introduced a concept he's calling "GPU gravity," the idea that faster GPUs will pull workloads toward them the same way data gravity pulls more data together, because in agent-to-agent interactions, the agent running on faster infrastructure simply makes better decisions, faster, more often.

Kolsky pushed back with an inference-economics argument: for steady, predictable, high-volume AI workloads, owning the infrastructure can be cheaper than the public cloud, and mainframes are seeing renewed relevance because of their CPU management strength. Ni added the practical wrinkle, that CIOs can't easily justify large capital expenditure to a board when workloads are still in experimentation mode, and that latency from data routing is a real cost most people underweight.

The debate didn't resolve into a single answer, and it wasn't supposed to. All three converged on hybrid as the near-term reality, but disagreed on the tilt: Mueller sees the long-term center of gravity moving to public cloud, Kolsky sees a genuine opening for on-prem economics, and Ni framed the real strategic question as workload routing, balancing quality, latency, data access, policy, and cost, rather than picking one mode and committing to it.

Qualtrics' New Chapter: From Measuring Experience to Simulating It

Liz Miller, Mike Ni and R "Ray" Wang sat down with the Qualtrics team in Salt Lake City following the company's recent launch event, and the shift they described is meaningful. Experience management has historically meant deciding on an experience, pushing it out, and measuring the response after the fact. Qualtrics' new direction moves that earlier: simulate the experience against synthetic panels, predict the outcome, and only then act, rather than running hundreds of live experiments against real customers.

Wang framed the shift around three ideas: simulate, trust, and outcome. The simulation piece lets teams test far more granular audience segments than traditional A/B testing ever allowed, down to sub-sub-personas rather than broad 50/30/20/10 splits. But the more interesting shift is philosophical. Miller pointed out that CMOs have spent years assuming they control the customer experience, when the customer has always been the one actually managing it. Qualtrics' new positioning directly acknowledges this, aiming to deliver "tailored" experiences that adjust in real time based on context, time, location, journey stage, even physiological signals, rather than a single predetermined path.

The two also unpacked "ontology" in plainer terms: it's the structured relationship between the signals a system tracks and what those signals actually mean for the experience being delivered. That structure, built on Qualtrics' roughly 20 years of experience data, is what makes simulation trustworthy rather than just another guess.

Splunk's Reinvention: Making Agents Reliable, Compliant, and Cost-Aware

Splunk used its recent conference to lay out a broader reimagining of itself as it deepens integration with parent company Cisco. Two announcements stood out to Constellation analyst Chirag Mehta. The new Observability Studio focuses on tokenomics, directly addressing a problem enterprises are already facing with agents in production: keeping them reliable, compliant with policy, and within budget. And Splunk is extending into security operations centers, where AI agents can now handle detection, investigation, response, and governance tasks directly.

The Cisco relationship is providing Splunk with meaningful distribution, with existing Cisco customers now able to discover Splunk's capabilities through products such as Cisco Command Center and AI Canvas. The framing throughout: Splunk's core strength was always turning data into answers, and that strength now extends to monitoring the behavior, security, and economics of AI agents themselves, not just the systems around them.

New ShortList: Semantic and Context Management

Mike Ni closed the episode by introducing Constellation's newest ShortList, covering semantic and context management, and made the case for why it deserves to be its own category rather than a subset of data management.

His framing: most executives can't cleanly define "enterprise context," but they know immediately when it's missing. A dashboard showing a 12% pipeline shrink isn't actionable on its own; a human operator fills the gap with judgment and experience. AI doesn't have that judgment, which turns missing context into an architectural problem the moment agents start taking action rather than just generating insight.

Ni drew a clear line between data management's traditional questions (what data do we have, can we trust it, what does it mean) and the newer layer semantic and context management adds: what matters right now, and what is an agent allowed or expected to do about it. He evaluated the category across nine capabilities, from semantic management and AI grounding to policy-aware context and explainability, and spotlighted Actian as a vendor worth watching.

Actian's approach centers on a federated knowledge graph that connects metadata, lineage, definitions, and policy across systems without requiring organizations to physically move their underlying data, paired with a knowledge graph that captures relationships and dependencies in a form AI systems can actually traverse, not just search. Ni also highlighted Actian's move toward "activated" context: machine-readable data contracts and support for standards like MCP that let governed context reach analytics tools, applications, and AI agents directly at runtime.

His closing question for AI leaders is the one worth sitting with: are your metadata investments just documenting the business, or are they becoming something your AI can actually use?

The Throughline

Every segment in this episode circles back to the same tension: enterprise AI's hard problems have moved past the model itself and into the infrastructure, context, and operating choices that surround it. Where workloads run, how experience decisions get made, whether agents stay reliable and cost-aware, and whether AI has the context it needs to act rather than just respond — these are the questions defining the next phase of enterprise AI adoption.