Cultivated AI · Mission Control · evidence
Real precedents for democratizing business intelligence with AI, and for showing up in AI answers. Sources are tagged by how much to trust them. The honest headline: the warehouse and the semantic layer are the project; the chat is the easy last mile.
On a real warehouse, plain text-to-SQL answered correctly about 33% of the time (GPT-4, 2023). Grounded in a governed semantic layer, the same questions hit 98 to 100%. Even with today's frontier models (90%+ raw), the semantic layer is still the line between "usually right" and "trustworthy." dbt Labs benchmark, public methodology, built on an academic benchmark.
Organizations that let non-technical people ask their data in plain language. The closest analog to Baker's Dozen is first.
Ask a question in plain English in Slack; Hermes generates SQL, runs it, and answers. V3 became a conversational "AI data analyst" on Snowflake, retrieving past validated queries as few-shot grounding, with an agentic loop and a confidence-plus-explanation layer.
For Baker's Dozen: same business, same geography, same warehouse choice many D2C brands make. The cheap trick to copy: reverse-generate questions from SQL your analysts already trust, then retrieve those as examples.
Internal natural-language-to-SQL for operators and analysts. A multi-agent pipeline (intent, table, column) with RAG over curated SQL, scoped by 12 business "workspaces," plus an evaluation harness scoring every answer.
For Baker's Dozen: the workspace pattern (scope the model to a domain before it writes SQL) and the eval harness are directly reusable. Note: the viral "140,000 hours saved" figure is a third-party extrapolation, not Uber's.
Text-to-SQL inside their open-source Querybook, evolved into an agent that also helps discover tables and reuse past queries. RAG table retrieval (OpenSearch) narrows the schema before generation.
For Baker's Dozen: proof that table and column retrieval (not a bigger model) moves accuracy, and that "first-shot acceptance" is a more honest success metric than raw benchmark scores.
A plain-English data assistant plus a SQL-dialect translation service, with a planned MCP bridge so one assistant can target Snowflake, Databricks, and Trino.
For Baker's Dozen: the multi-engine, MCP-mediated pattern is useful if you end up with more than one store (GobbleCube exports plus a warehouse).
Two adoption templates. Instacart's Ava (general assistant) reached 900+ weekly users by living where people already work. Wealthsimple's open-sourced gateway adds PII redaction, self-hosted models, and audit, used by over half the company.
For Baker's Dozen: adoption scales when the tool starts small and lives in Slack/Teams; and the gateway is the template for the PII, access, and audit layer that should wrap any leadership-facing assistant.
Every one of these depends on a governed semantic model for quality. The client's existing stack (Power BI) is first.
Copilot in Fabric is generally available: generate report pages from a prompt, write narrative summaries, answer in conversational language. Fabric "data agents" sit over your data and are consumable from Power BI, M365 Copilot, and MCP endpoints. Available from the F2 SKU since April 2025 (far below the old F64 floor).
For Baker's Dozen: the fastest on-ramp, but quality depends entirely on well-modeled, well-named semantic models. The "3-5x analyst productivity" claim is vendor-ecosystem, not verified. The warehouse and semantic-layer work is a prerequisite, not optional.
Conversational analytics over a lakehouse, embeddable into Teams, governed by Unity Catalog. The AA trained it on "golden questions," validated answers against existing Power BI dashboards, and used catalog RBAC.
For Baker's Dozen: The AA's recipe is directly copyable: golden questions, validation against trusted dashboards, governance from the catalog, delivered in the chat tool people already use.
Four ways to buy the same idea. Cortex Analyst answers plain-language questions against a semantic model written as YAML, inheriting RBAC. Looker + Gemini grounds answers in the LookML semantic layer. ThoughtSpot Spotter is retail/CPG-specific (inventory, sell-through, promotions). Tableau Pulse pushes metric digests and anomaly explanations without being asked.
For Baker's Dozen: ThoughtSpot's promotion-lift and sell-through use cases map almost one-to-one onto quick-commerce questions. Pulse's push model fits a leadership team that won't log in daily.
A public benchmark (built on an academic one) comparing raw text-to-SQL against semantic-layer-grounded querying, each question run 20 times across models.
For Baker's Dozen: this is the line to put in front of leadership. For questions inside the semantic model's scope, accuracy approaches 100% because SQL is generated from governed definitions, not inferred.
Every success above shares the same skeleton, in this order. The order is the lesson.
One governed store the GobbleCube and Power BI telemetry feed into. Without it, the AI is querying inconsistent silos.
Business definitions (revenue, active SKU, fill rate, incremental lift) encoded once, centrally. The highest-leverage component.
Ground the model in validated past queries and the right tables before it generates anything. Moves accuracy more than a bigger model.
Narrow the model to a business area before it answers (Uber's "workspaces"). Less ambiguity, fewer tokens.
Break a complex question into steps rather than one-shot generation.
RBAC, row-level security, masking, SSO, and audit logs, all from the warehouse layer.
Show the query or assumptions, attach a confidence score, validate against an already-trusted dashboard. This is what makes leadership act.
A golden-question test set scored continuously. You cannot improve what you do not measure.
Slack, Teams, or the existing BI tool. Adoption follows convenience.
When a shopper asks an AI "best protein chips in India," the answer is assembled, not ranked. The job shifts from ranking one page to being mentioned consistently across the sources AI trusts.
"GEO: Generative Engine Optimization" (Princeton, Georgia Tech, Allen Institute, IIT Delhi; KDD 2024) tested content tactics on a 10,000-query benchmark. Optimizations lifted source visibility by up to 40%. Best low-effort moves: quotations (~28%), statistics (~26%), fluency (~25%), citing sources (~25%). Keyword stuffing did almost nothing.
Outside the Princeton paper, most AEO numbers come from vendors and aggregators, not peer-reviewed work. Use the academic study for credibility and the tactics as a checklist, not a guarantee. AI shopping referrals are small today (around 1% of web traffic) but growing fast and converting well, so it's worth starting now, cheaply.
Every claim above is cited in full in bakers-dozen/research/bi-democratization-case-studies.md, with source URLs and credibility tags.