Golden Data

Issue #8 · June 28, 2026

The dashboard wrote itself.

Paul Brown Writes Golden Data from Ogden, Utah 6 min read

Every Sunday I look at the demo a vendor wrote that the user is no longer needed to see. This fortnight four BI vendors shipped the same one. An agent walks up to the semantic model and starts writing.

For a decade the BI tool was the surface the analyst clicked. The dashboard was the artifact the analyst shipped. The semantic model was the contract that backed both. The vendor sold the surface, the analyst built the artifact, the model held the truth. This fortnight four BI vendors quietly shipped the agent that authors the model and the dashboard and the embed. The BI tool stopped being a query surface. The BI tool became a write-back.

The next quarterly review will have a row for who authored each dashboard. The row will not always say a person.

What’s actually shipping this week

1. Microsoft open-sourced Fabric Skills for Claude and GitHub Copilot through the June Fabric Updates Blog. Open-source toolkit on GitHub at microsoft/skills-for-fabric, shipped alongside the Power BI June 2026 Feature Summary. The skills let an external agent — Claude, Copilot, the Fabric CLI — author Power BI reports, query semantic models, and validate Fabric data assets without a human in the UI. Power BI Report Authoring Agent Skills produce a publishable report from a description or screenshot. Fabric Data Agents are now publishable as declarative agents inside M365 Copilot. The BI tool turned its UI into an API the external agent calls.

2. Databricks shipped Genie One GA at Data + AI Summit on June 16. Agentic coworker for business users across structured and unstructured, analytical and operational data, inside or outside Databricks. Unity Catalog Metrics shipped the same day with multi-fact relationships, parameterized metrics, and agentic UI-driven authoring — the agent drafts metric definitions, the human approves. Unity AI Gateway governs the runtime. Catalog Federation pulls every connected catalog into one substrate the agent grounds on. The metric became a governed object the agent writes against.

3. ThoughtSpot integrated Spotter with Snowflake Cortex AI and Semantic Views at Snowflake Summit 26 on June 2. Spotter’s four agents — SpotterViz writes dashboards, SpotterModel writes the semantic layer, SpotterCode writes the embed, Spotter 3 is the conversational core — now query Cortex Analyst and Cortex Agents directly. The customer’s BI surface and the customer’s warehouse stopped being two services. The agent walks between them with one set of credentials.

4. Sigma Agents went Public Beta on June 12. Native agentic AI inside dashboards and apps, with MCP tool support, warehouse agent integrations, and a runtime registry the agent reads to pick the right tool. The agent answers the dashboard, but it also opens the workbook and edits the calculations the dashboard reads from. The dashboard stopped being the artifact the analyst published. The dashboard became the surface the agent rewrites.

What I’d ship in your app this week

The model picked up an author. Two small tools your team can ship in two weeks so the external agent in the customer’s stack can read, write, and account for what it changed inside your product.

Feature one: the Skills package your product publishes. Open-source a small toolkit on GitHub — five to ten declarative skills mapped to the writes your product supports (create dashboard, define metric, edit query, publish report). Generate it from the OpenAPI spec your docs already render. The external agent in the customer’s stack — Claude, Copilot, in-house copilots — does real work in your product without a human navigating the UI.

  • Shape. One GitHub repo with N declarative skill manifests plus a thin MCP-server entrypoint. Generated at build time from your existing OpenAPI spec.
  • Data shape. Per skill: name, description, input schema, auth scope, write target, idempotency key.
  • System shape. Skill manifests plus a thin MCP wrapper. No new model call on the production path.
  • Latency budget. None on the production path. Skill discovery under 200ms, edge-cached.
  • Cost ceiling. Build-time only — under $5 per month in CI.
  • Eval. 40 skill invocations per month reviewed by a senior engineer on "would an external agent pick this skill from the description alone, and did the write land where it was supposed to."
  • Instrumentation. Log skill_name, calling_agent, calling_user, and outcome on every invocation. The most-called skill is the next pricing-page rewrite.
  • Two weeks in. Three external agents — Claude Managed Agents, Copilot, one customer-internal copilot — write into your product without a human onboarding. If they don't, the skill descriptions read like docs, not search results — rewrite them as answers to a task query.

Feature two: the agent-authored review queue, on the customer’s admin surface. When an external agent authors a metric, dashboard, or query through the Skills package above, route the draft to a human-review queue in the customer’s existing admin UI. One row per drafted artifact, with the prompt that produced it, the calling agent, and a one-click approve or reject. The customer’s analytics lead sees what the agent wanted to ship before it lands in production.

  • Shape. One table — agent_drafts — plus an admin-UI view. Approve writes to production; reject discards.
  • Data shape. Per (draft_id, tenant, agent, calling_user, prompt_fragment, artifact_type, proposed_diff, status, decided_at).
  • System shape. Synchronous when the skill is called; the write lands in a draft row, not production. No new agent loop.
  • Latency budget. Under 500ms on the skill call. The review is async on the human side.
  • Cost ceiling. Storage only — under $30 per month per 5,000-seat tenant.
  • Eval. 40 drafts per quarter sampled by a senior CSM on "would the customer's analytics lead recognize the proposed change in their workflow."
  • Instrumentation. Track approve / reject / time-to-decision. A low approve-rate is a skill description problem; a high time-to-decision is a queue UX problem.
  • Two weeks in. Three customer admins have approved at least one agent-authored artifact through the queue. If they haven't, the queue is empty — work backward through the Skills package to find which skills the customer's agents are not calling.

Both ship in two weeks with the team you have. Both turn an external agent in the customer’s stack from a chat surface into a write surface in your product, with a paper trail at the door. The BI vendors who shipped agentic authoring this fortnight made the dashboard the agent’s artifact, not the analyst’s. The vendors who don’t will be the read-only tool the customer’s agent walks past on its way to write somewhere else.

Sources

Send me an email and we will talk. If something here landed close to what you're working on, the door is open. No calendar funnel, no pitch deck — I read every note that comes in.

Doing the work rather than deciding what to build? Crafting is the column for that chair.

Email me Read more Signal

← All Signal issues · Drafted with Claude · Edited by Paul Brown