Databricks built an AI analytics assistant named Marge that lets its marketing team ask natural-language questions and get governed answers in seconds. The system has pushed data-driven decision-making up threefold across the organization, with more than 85% of the marketing department now using it regularly.
Marge handles over 800 questions each month and has fielded more than 5,000 total since launch. Flagged incorrect responses dropped by 25% as the system improved. The team scaled access to insights without adding analytics headcount, freeing analysts to focus on higher-value work like experimentation and model design.
Why self-service analytics breaks down for marketing
Marketing data lives across campaign platforms, web analytics, CRM systems, event tools, advertising channels, and sales data. Each system carries its own definitions, identifiers, and reporting logic. Even when dashboards exist, marketers often struggle to know which metric is current, which source is authoritative, or how to answer a question the dashboard was not designed to handle.
Databricks faced the same friction. Data and reporting sat in multiple places. Dashboards were sometimes stale or inconsistent. Marketers did not always know which numbers to trust. Analytics and engineering teams spent significant time on repetitive requests, and demand for insights grew faster than the team could support.
The company first created a Marketing Lakehouse within its company-wide Databricks lakehouse, shared across Marketing, Finance, Sales, and Product. It became the governed source of truth for go-to-market data, aligning campaign and sales information around consistent definitions and metrics. That foundation made Marge possible.
What Marge actually is and how it works
Marge is a conversational analytics assistant built with Genie Agents, grounded in the data and business context inside the Marketing Lakehouse. It understands the definitions Databricks marketers use for regions, products, channels, campaigns, fiscal periods, and pipeline.
Marketers ask questions in plain English - "How did my email campaign perform?" or "Which programs influenced pipeline this quarter?" - and Marge translates those into analytical queries against governed enterprise data. Unity Catalog provides centralized governance, lineage, and role-based access controls, so each user sees only the data they are authorized to access.
Marketers access Marge through Genie One, an AI cowork experience that brings together dashboards, Genie Agents, apps, and deeper analysis. The system automatically routes each request to the appropriate Genie Agent. Focused agents exist for domains such as web performance, digital analytics, and marketing planning, each operating with narrower, more relevant context. For complex questions, Agent mode evaluates multiple steps and produces a deeper analysis.
Four mechanisms that make Genie trustworthy
Trust was the top priority from the initial pilot through broad rollout. No AI system is perfect, so the team focused on four mechanisms to improve accuracy, reliability, and transparency.
First, document the data and its relationships. Genie needs the same context a new analyst would need. The team uses Unity Catalog to centralize metadata, table descriptions, column annotations, lineage, and access controls. AI-generated descriptions gave a useful first pass, but marketing stakeholders and data experts reviewed and enriched them with business context only people inside the organization would know - the precise meaning of a marketing-qualified lead, the dates included in each fiscal quarter, how campaigns map to products and regions, and which fields represent cost, engagement, or attribution.
Second, encode verified answers and example queries. For common, high-value questions, the team provides trusted assets and verified logic that domain experts have reviewed. These cover areas such as conversion rates, customer lifetime value, and event registration. Users can see when an answer is based on verified logic, which adds an important signal of trust. Example question-and-query pairs teach Genie how to handle known scenarios and generalize the same pattern to similar questions.
Third, teach Genie the language of the business. Every organization has terminology that looks simple but carries specific meaning. "Pipeline," "region," "fiscal year," and "campaign" may each have definitions that differ from one company to another. The team gives Marge clear behavioral guidance for interpreting these terms and handling ambiguity. Users may say "spend" or "investment" when the underlying field is named "cost" - Marge needs to understand those words refer to the same concept. The system is also instructed to ask clarifying questions when a request is missing critical information, such as the time period, channel, or region.
Fourth, create continuous feedback and evaluation loops. Every response gives users an opportunity to provide positive or negative feedback. The marketing analytics team reviews ratings and comments in a monitoring dashboard, investigates issues, and updates the agent as needed. Benchmark questions are run against known answers to evaluate performance systematically. A meaningful decrease in benchmark accuracy signals that the data model, definitions, or agent context may need attention.
How adoption spread across the marketing organization
Technical quality was only half of the work. Marketers also needed to believe that Marge understood their needs and could fit naturally into their daily workflow.
The team started with one focused use case: email campaign performance. They included only the essential campaign, recipient, and engagement data required to answer those questions. Starting narrow made it easier to validate accuracy, build confidence, and show value quickly. Account data, attribution, and other domains were added later based on what users requested.
They also embedded Marge into an existing workflow. Every marketing analytics ticket now receives an automated response asking, "Have you asked Genie?" Analysts only engage after the requester indicates they have tried Genie. This simple change directs basic questions to self-service analytics and preserves analyst time for more complex work. The team describes the difference as moving analysts away from 101- and 201-level requests so they can focus on 301- and 401-level analysis.
Quick action on feedback proved essential. The team continuously updates metadata, examples, trusted answers, and instructions based on what users report. Rapid improvements show marketers their input matters and help Marge become more useful with every iteration. The feedback loop also helps the team expand with discipline - they add data and capabilities in response to demonstrated demand rather than trying to anticipate every possible question at launch.
A practical rollout sequence for martech teams
For teams beginning a similar journey, the recommended sequence is straightforward. Choose one bounded, high-frequency use case - a question marketers ask often, supported by data you understand well. Select the minimum required data, including only the tables and fields needed to answer that first set of questions. Document the business context: relationships, metrics, fiscal periods, regions, and domain-specific language.
Add verified logic for important questions using metric definitions, example SQL, and trusted assets where consistency matters most. Pilot with a small user group and observe their actual questions, confusion points, and language before expanding. Measure quality and behavior - usage, feedback, flagged answers, and benchmark accuracy. Embed Genie into an existing workflow so it becomes the natural first stop for common analytics questions. Then expand based on demand, adding new data and focused Genie Agents as users demonstrate a need for them.
Ongoing maintenance is relatively light. At Databricks, one BI manager spends roughly one hour per week reviewing feedback and maintaining Marge. That small, consistent investment helped reduce the rate of flagged incorrect answers by 25%.
Why this matters for marketing professionals
Marge began as a prototype for 10 users and one use case. It now supports more than 85% of Databricks' marketing organization and has answered over 5,000 questions. The path from pilot to scale was built one trusted answer at a time - and the underlying lesson is that self-service analytics depends entirely on a strong data foundation. AI makes governed context easier to access, but it does not compensate for conflicting source data and unclear business logic. For marketing teams exploring AI for Marketing, the playbook is concrete: start small, embed the tool into existing workflows, and let user demand - not speculation - determine what gets built next. Managers looking to lead this shift can follow a structured AI Learning Path for Marketing Managers that maps directly to the skills these rollouts require.
Your membership also unlocks:
