Enterprise AI strategies continue to return to the same difficult questions: where should workloads actually run, how should vendors be compensated for their results, and how can buyers determine actual AI capabilities from marketing labels? Episode 132 of ConstellationTV features all three.
On-premises or cloud? The debate continues
Host Holger Müller kicked off the episode with a big debate about whether AI needs to run in the cloud or can run on-premises, also joined by Michael Nie and Esteban Kolsky. Michael argued that AI follows its own gravity, pulling computing to where the inference demands and real-time decision-making actually occur, pointing to a distributed model that spans cloud, edge, and on-premises layers. Esteban countered with arguments about the gravity and elasticity of data, likening the economics of cloud to a historical shift away from companies generating their own electricity. Holger argued for the advantage from a cost and efficiency standpoint, pointing to Apple’s on-device AI push as evidence that companies don’t need to run everything through servers. The discussion landed on where these things usually take place. It depends on the tier, and the next few years will be about getting that placement right, rather than choosing sides.
Personalization is a means, not an outcome
Liz Miller took to Marketing Minute to share takeaways from Pegasystems’ annual event, Pegaworld. Her point is that personalization is a strategy, not a business outcome itself. The real goal is to build more lasting and profitable customer relationships, and Pega’s Customer Decision Hub has been quietly doing that job for years, long before the word “agency” became a modern parlance. Liz also referenced Pegasystems CEO Alan Trefler’s comments about intentionally testing and deploying agent AI without losing sight of why the work is important in the first place.
Outcome-based pricing: great on paper, difficult in practice
Larry Dignan covered the new focus on outcomes-based pricing, which Oracle, Pegasystems, HubSpot, and UiPath have all mentioned recently. His view is skeptical. Outcome-based pricing faces the same problems that revenue-sharing models always have. In other words, no one agrees on how to measure outcomes. When you have something to share, no one wants to give up their share, and it’s unclear who will bear the cost if the outcome takes longer or takes more tokens than expected. Lally expects this topic to feature heavily in the second-quarter earnings release, but his advice for now is to tread carefully.
5 ways to spot Agentic AI washing
Holger concluded the episode with a preview of a new best practices report on AI washing with agents. This extends the same skepticism that was applied to cloud washing a decade ago. Of the 18 questions included in the full report, he highlighted five: whether the functionality existed before generative AI became widespread in 2023, whether it’s just a chatbot with a new label, whether it actually runs in the cloud and can scale flexibly, whether it has access to third-party data as well as the vendor’s own, and whether the underlying functionality is extensible rather than fixed. In Holger’s view, vendors that don’t pass these tests are likely repackaging something that already existed.
