"The scarce resource is consensus" — @Ian Macomber
Why this is in the vault
Ramp's head of data, Ian Macomber, returns to the Analytics Engineering Roundup to argue that the post-AI data team has two distinct jobs — enabling everyone to use data, and maintaining a single trusted version of reality — and that the second job is now harder and more valuable than the first. The "consensus as scarce resource" framing is one of the cleaner articulations of why AI doesn't make data teams obsolete; it reorients what the hard work actually is.
⚠️ Sponsorship
Sponsored by dbt Labs (self-promo): AER is dbt Labs' community publication, and the footer reads "This newsletter is sponsored by dbt Labs." A mid-episode dbt Summit 2026 promo block also appears (Sept 15-18, The Cosmopolitan, Las Vegas). Weight any implicit advocacy for dbt-centric tooling (abstraction layers, semantic layer) as coming from the publisher's own stack.
The core argument
Macomber's central claim: as building data artifacts gets faster and cheaper through AI, analysis is no longer the scarce resource — company-wide consensus on a single version of reality is. "A 7-out-of-10 dashboard is dangerous precisely because bad data fails quietly, where bad design fails loudly." The post-AI data team has two jobs:
Enable every employee to use data accurately, powerfully, and independently. Ramp built Ramp Research — an internal tool that functions as an API layer over their data lake — scaling self-serve 50x. Finance analysts can build stateful apps on it; it handles evals, permissions, and telemetry. Over the past year Ramp made "enormous progress on job one, partly at the expense of job two."
Build and champion the singular reality the company operates on. This is the harder, less visible job. Manufacturing consensus for both humans and agents alike is the next frontier.
Three tactical takeaways from the episode:
- Route agents through an abstraction layer, not raw data. Wiring agents directly to sources (e.g., Gong) burned tokens and returned noisy results; pulling data through the lake first and summarizing cut payload size roughly 20x.
- Optimize for your fastest movers, not the median employee. Macomber's biggest do-over: the team was "too empathetic." Back the most resourceful people early, make them champions, let the rest follow.
- Retire the token leaderboard once everyone is on board. The "token maxing era" needs a deliberate endpoint; the transition signal is when usage shifts toward "value on intelligence" rather than raw volume.
Mapping against Ray Data Co
The consensus job is Ben's phData DSA role, named plainly. Clients can spin up AI-driven analytics faster than ever, but the actual delivery is getting stakeholders to trust and operate on a single version of reality — not just deploying the tool. Macomber's "two jobs" framework maps directly to the solution-selling surface: selling job one (self-serve enablement) is the easy close; winning job two (organizational consensus infrastructure) is the durable moat and the harder close. That's exactly where RDCO's positioning lives in phData engagements.
The COO agent parallel: the agent automates execution, but organizational consensus — across phData clients, across internal stakeholders — is irreducibly human-intensive work. "Manufacturing consensus for humans and agents alike" is a near-direct description of the gap the agent doesn't close. Worth using in client conversations to explain why the data team doesn't shrink; it reorients.
The "route agents through an abstraction layer" point is pitch-ready for phData engagements where clients want to wire AI agents straight to raw sources. The 20x payload reduction stat from Ramp Research is a concrete anchor.
Related
- [[2026-01-11-ae-roundup-ai-agents-data-lake]]
- [[2026-05-31-analytics-engineering-roundup-small-long-running-agent]]
- [[2026-05-10-ae-roundup-computers-talk-to-us-now]]