Cultivated AI · Part Two · Mission Control · facilitator plan
A small, mixed room: leadership plus the people who own the data. This is a build, not a lecture. By the end there is a real pipeline plan, real questions answered, and a roadmap. This is the internal run-sheet.
Inventory the sources, choose the warehouse, design the daily dump.
The semantic layer, golden questions, and a trust surface.
Query in plain language, then work one high-impact problem end to end.
The protein-chips exercise, an AEO playbook, and the roadmap.
Scattered data can't be asked. Get every source into one governed place on a daily automated dump, then ask the first real question and see what's still missing.
Leave with a complete data inventory, a decision on the central warehouse, a daily-dump plan, and a precise list of what's blocking trustworthy answers.
The single nervous system Discuss
Four dashboards are four places to look, not one place to ask. The goal of the whole workshop: ask the business a question, get an answer you can act on.
Data inventory Lab
List every source: Power BI telemetry, GobbleCube (city, store, and platform analytics), production, and supply chain. For each, capture the grain, the owner, how often it refreshes, and who can see it. This map is the foundation.
The central warehouse Teach Decide
What a single source of truth is and the realistic options. Because Baker's Dozen already runs Power BI, Microsoft Fabric / OneLake is the lowest-friction path; Snowflake and BigQuery are the common D2C alternatives. Map current state to target.
Break
Short reset.
The daily dump Lab
Design a scheduled daily refresh that lands every source in the warehouse: GobbleCube exports, Power BI semantic models, production and supply-chain files. Sketch the pipeline and name an owner for each job, so it runs without anyone remembering to press a button.
First query, warts and all Lab
Point an AI layer (Power BI Copilot or a warehouse copilot) at the data and ask one real question. Watch it half-work. Write down exactly what it got wrong, missing definitions, ambiguous joins, unclear metric names. That gap list is Session 2.
Close: the gap list Discuss
Agree the top blockers and who chases each before next session.
A data inventory, a warehouse decision, a daily-dump pipeline sketch with owners, and a gap list.
The highest-leverage session. Define your metrics once so the AI stops guessing, seed the questions leadership actually asks, and build the trust surface that makes answers safe to act on.
Leave with a first metric dictionary, a golden-questions list, a low-cost way to seed examples, and a routine for validating answers.
Why answers come back confidently wrong Teach
The proof point: on a real warehouse, plain text-to-SQL is right roughly a third of the time; grounded in a semantic layer it approaches near-perfect on in-scope questions. The danger is not an error message, it's a wrong number that looks right. This is the slide that justifies the whole project.
Your metric dictionary Lab
Define, once and centrally: revenue, active SKU, fill rate, incremental lift, city contribution, wastage. Resolve the quiet disagreements (when marketing's "customer" is not finance's "customer", no AI can be right for both until you decide).
Golden questions Lab
Write the ten to twenty questions leadership genuinely asks ("which SKUs slipped on Blinkit in Pune last week, and why"). These double as the test set you'll score the system against.
Break
Short reset.
The Swiggy trick Teach Lab
Swiggy (same business, Indian quick-commerce) went from 54% to 93% accuracy by retrieving past validated queries as examples. The low-cost move: reverse-generate plain-language questions from queries your analysts already trust, then feed those back as grounding. Plan how to seed yours.
The trust surface Lab
Make every answer show its working: the generated query or assumptions, a confidence signal, and a check against an existing trusted Power BI dashboard. Leadership acts on answers they can sanity-check, and not before.
Governance, inherited Teach
Don't rebuild security. Inherit role-based access, row-level rules, and audit logs from the warehouse, and wrap PII handling around it. Decide who can see what before anyone asks the first sensitive question.
Close
Confirm owners for the dictionary and the example library.
A metric dictionary v1, a golden-questions list, a method to seed examples cheaply, and a validation routine.
Now use it. Leadership queries the nervous system in plain language, then applies the AI lens to one high-impact problem, end to end.
Leave having answered real questions without an analyst in the loop, worked one high-impact problem to a decision, and started a scored eval set.
Deliver where you already work Teach
A new dashboard nobody opens loses to a question you can ask in Slack or Teams. Decide where Mission Control lives so it actually gets used.
Live querying Lab
Each leader runs their own golden questions, cross-checks the answer (Part One, Principle 11: never trust one machine), and iterates. Track first-shot acceptance, the most honest measure of whether it's working.
Break
Short reset.
The AI lens on a high-impact problem Lab
Pick one real problem (city-level wastage, slipping fill rate on Zepto, promo ROI by platform) and work it the whole way: frame the question, query the data, cross-check, decide, name the action. This becomes the template the team reuses on every future problem.
The evaluation harness Teach Decide
Turn the golden questions into a test set that's scored continuously (did it run, was it right, was the first answer accepted). You can't improve what you don't measure, and this is how trust grows over time.
Close: who owns the answers Discuss
Name an owner for the system and a cadence for reviewing the eval scores.
Real questions answered directly, one high-impact problem worked to a decision, and a scored eval set started.
Turn the same AI lens outward. Make Baker's Dozen show up when a shopper asks an AI what to buy, then leave with a phased roadmap.
Leave with an honest audit of how Baker's Dozen appears in AI answers today, a concrete AEO playbook, and a sequenced roadmap with owners and metrics.
The protein-chips exercise Lab
Ask ChatGPT, Gemini, and Perplexity "best protein chips in India," then "high-protein cookies" and "healthiest packaged bread." Record who gets named, who doesn't, and which sources the AI cites. The gap is visible to the whole room in five minutes.
How AI answers pick brands Teach
The answer is assembled, not ranked. AI recommends what shows up consistently across the sources it trusts (Reddit India, YouTube, review sites, your own structured pages). The one peer-reviewed study (Princeton, KDD 2024) found quotes, statistics, and citations can lift AI-answer visibility by up to about 40%.
Break
Short reset.
The Baker's Dozen AEO playbook Lab
Audit and plan: entity clarity (same product descriptions everywhere, unambiguous category), write-for-extraction (protein grams, ingredients, certifications, not vague claims), presence on Reddit and YouTube and the marketplaces, earned reviews, and a way to monitor share of voice across the assistants.
The roadmap Lab
Sequence the build: warehouse, then semantic layer, then the chat layer, then AEO. Phase it, decide what's in-house versus a partner, assign owners, and set the metrics. The concrete path to India's most AI-forward D2C org.
Commit Discuss
A first review date, named owners, and the single metric that will prove this is working.
An AI-answer visibility audit, an AEO playbook, and a phased roadmap with owners and metrics.
Right people in the room. This only works if leadership and the data owners are present together. Decisions (warehouse choice, metric definitions, owners) get made live, not deferred.
Access before Session 1. Read access to Power BI, GobbleCube exports, and the production / supply-chain files needs to be sorted in advance, or Session 1 stalls on logins. See the pre-req list.
Build, don't admire. Every session ends with an artefact: the inventory, the dictionary, the worked problem, the roadmap. If a teach segment runs long, cut it, never the lab.
The order is the message. Resist jumping to the chat layer. The whole credibility of the workshop rests on building the warehouse and semantic layer first; the natural-language interface is the easy last mile.
Honesty on AEO. Lean on the one peer-reviewed source for the headline claim. Present vendor traffic and conversion numbers as directional, not fact. The honest message: AI shopping referrals are small today but growing and convert well, so it's worth starting now, cheaply.
The precedents and numbers behind this plan are in the case studies doc (Swiggy Hermes, Uber QueryGPT, the semantic-layer benchmark, the Princeton GEO study).