06-reference

ae roundup bi second unbundling

Sat May 02 2026 20:00:00 GMT-0400 (Eastern Daylight Time) ·reference ·source: Analytics Engineering Roundup ·by Tristan Handy (dbt Labs CEO)
bi-unbundlingpresentation-layeragent-deployersemantic-layeridentity-moatmac-positioning

“BI’s Second Unbundling” — @Tristan Handy

Why this is in the vault

Tristan is calling the second unbundling of BI in real time, from the seat of dbt Labs (i.e. the canonical “transformation layer won” vendor). His framing — agent-as-frontend collapses visualization/controls/connectivity into commodity React + MCP + Arrow + Vercel, leaving identity as the persistent moat — directly maps onto MAC + Client Reporting positioning territory. This is the clearest articulation we have of WHO is supposed to eat the BI presentation layer next, and where the defensible wedge lives.

The core argument

  1. BI was a bundle. MicroStrategy / Cognos era: visualization + ingestion + transform + compute + cache + semantics + identity + hosting all in one product. The BI tool WAS the data stack.
  2. First unbundling, 2015–2022. Modern data stack pulled out compute (Big 5 warehouses), ingestion (Fivetran), transformation (dbt). BI was left with visualization, interactive controls, semantic definitions, identity/access, and hosting. Tristan calls this the “Copernican Revolution of BI” — the universe stopped rotating around the BI tool.
  3. Second unbundling, happening now. Driven by analysts shifting left into agentic coding environments (Claude Code, etc.). Building a chart in Claude is now a common modality; Anthropic’s live artifacts launch made it 15-min-conversation viable.
  4. Two early-but-right precedents: Evidence.dev and Hashboard. Both built “BI as code” (SQL + markdown, version-controlled, deploys like a web app) and got modest traction. Tristan’s read: hand-coding YAML wasn’t native to the analyst workflow then. Coding-agent-as-frontend changes that — generating dashboard YAML via an agent feels more natural than clicking through a GUI tab when you already live in Claude Code.
  5. Reconstituted BI as frontend engineering:
    • Visualization → React charting libraries
    • Interactive controls → React components
    • Semantic definitions → MCP servers from infra vendors (dbt’s MCP server, Snowflake Semantic Views)
    • Database connectivity → ADBC / Arrow Flight
    • Hosting → Vercel / Cloudflare
  6. The gap: the components exist but no integrated, analyst-purpose-built solution stitches them. Tristan’s MVP sketch: dashboard format spec + agent skills for writing it + YAML files alongside dbt project + lightweight renderer. Hugo/Jekyll-for-dashboards.
  7. The hard part — identity is the persistent moat. Jupyter-notebook-sharing problem analogy: notebooks were brilliant for analysis, nightmare for distribution. BI tools solved that gap with permissions, RLS, SSO. dbt punted on this at the transform layer (DB credentials = identity), but that shortcut doesn’t work at the presentation layer because most BI consumers don’t have warehouse accounts. In the agent-front-end world, someone needs to be the identity provider — and current BI tools aren’t necessarily positioned to keep that role.

Closing line: “Software engineering started eating data a decade+ ago; it just took a little while to get to the presentation layer.”

Mapping against Ray Data Co

This is the most important mapping in today’s batch.

MAC + Client Reporting positioning — direct hit.

The canonical RDCO positioning territory (01-projects/positioning/STRATEGY, 01-projects/positioning/2026-04-23-brand-architecture-umbrella-and-bets) puts MAC (Marketing Analytics Consulting) and Client Reporting at the center of the consulting bet — the wedge is “we run reporting agents for marketing-analytics teams who can’t or won’t operate them themselves.” Tristan’s frame validates the timing AND sharpens the technical bet:

Cross-bet thread: Service-as-a-Software, again.

Pairs cleanly with 2026-05-03-heyrico-service-as-a-software-shift (services market collapsing into agent-operated software) and 2026-05-03-yc-build-company-with-ai-from-ground-up (build agent-native from the ground up). Tristan is describing the same shift in BI specifically. The agent-deployer pattern shows up across services, vertical software, and now analytics — these aren’t independent observations, they’re the same wave.

Per-token economics tie-in.

2026-05-02-moonshots-ep252-google-anthropic-gpt55-cloud (per-token costs collapsing, models commoditizing) explains why “build a dashboard in a 15-minute Claude Code conversation” is even economically viable. The compute side of agent-driven analytics is no longer the binding constraint — which removes the last excuse for BI vendors to keep the workflow inside their walls.

Stratechery commodity-layer thread.

2026-04-30-stratechery-amazon-earnings-trainium-commodity argues the substrate (compute, models) is commoditizing and value moves up to integration layers. Tristan is making the corollary observation at the BI layer: the visualization/connectivity components are commoditizing into React + MCP + Arrow, and value moves to (a) integration into analyst workflow and (b) identity. RDCO’s bet should be on the integration-and-identity wrapper, not on building a better chart library.

Vendors named (positional context):

Decision threshold check:

Does this article say agent-deployer pattern is right for analytics specifically? Yes — implicitly and strongly. Tristan stops short of saying “you should run an agency that operates these for clients” because he’s a vendor not a consultant, but the gap he describes (“integrated, analyst-purpose-built solution missing”) is the exact wedge MAC + Client Reporting was designed for. The article does NOT change RDCO’s direction, but it RAISES the priority of MAC + Client Reporting prosecution within the bet portfolio because it confirms (a) timing is now, (b) the incumbent BI vendors are vulnerable, (c) the technical stack to build on top of is largely OSS.

All quotes paraphrased or ≤15 words. Source URL preserved for canonical reference. Tristan Handy / Analytics Engineering Roundup retains all rights.