The Stacked Data Podcast

The episode contains several genuinely non-obvious operational ideas - treating the data team as a deliberate non-embedder, the two-layer semantic model for AI, and the auto-PR feedback loop for wrong numbers - but these are diluted by extended career-advice segments, generic self-serve commentary, and sponsor content that pads the runtime.

“if I say that the data team are responsible for building reports and dashboards for the commercial team, and the commercial team is growing very fast and I'm slow or hiring or something goes wrong or whatever else, I become a point of failure for the commercial team”

“we've now had to go through a process of rewriting the entire semantic layer, um, for AI. Um, but we've found that with cursor, we can create a new topic in 10 minutes”

The deliberate-restraint framing - data teams should explicitly refuse ownership of dashboards and hand accountability to business users - is a contrarian and coherent position rarely argued this directly; the 'two semantic layers' concept (warehouse layer plus AI skill-file layer) is a fresh structural idea, though some surrounding points about self-serve and stakeholder respect are familiar.

“I don't actually think that comes from decision making... from my experience where companies really succeed or fail is their ability to actually operate and execute well”

“I think my most controversial view is I think that a lot of data people, particularly semi junior data people, to be honest, I think they don't have enough respect for their stakeholders”

Ed Mansi is a genuine practitioner who was the first data hire at a credible, high-growth AI company and previously held go-to-market analytics roles at GoCardless and Brandwatch; his cross-functional commercial background gives his operational views real grounding, though director level at a ~Series C company is solid rather than exceptional seniority.

“I was the first data hire there when I joined two years ago and uh, I'm responsible for the data platform as a whole, um, across product and commercial”

“before that I was um, managing kind of go to market analytics at GoCardless and then before that I kind of cut my teeth at uh, variety of different roles at brandwatch where I was for sort of eight years”

The episode has several concrete data points - 10x Snowflake spend growth in 12 months, a July 2024 go-live that was months ahead of a January 2025 plan, 150% quota attainment attributed to an AI workflow, and a 5-minute PR turnaround - but many of the operational arguments remain at the level of illustrative anecdote rather than being backed by systematic evidence or broader benchmarks.

“in a 12 month period we'd seen something like a 10x growth in the Snowflake spend”

“I joined in late April two years ago. So 2024 and so I imagined that we'd be going live with something in January, February 2025. We actually ran the ballpack out of Omni in July 24th”

The host asks a few legitimately probing questions (calling out whether the self-serve model only works because Synthesia hires unusually technical people, pressing on blockers) but the pre-existing relationship, the fact that Omni is simultaneously the episode sponsor and the guest's primary tool, and a pattern of validating responses before following up all constrain genuine challenge and allow several key claims to go unexamined.

“One thing I think that uh, if I was another data leader, uh, listening to this, that I might question is Synthesia is arguably one of the most exciting companies globally, hires very analytical, tech focused people across the business”

“I think that's an excellent um, way of framing it and yes, really good tangible both there”