What You Will Build

You will connect your company’s data warehouse (Snowflake, BigQuery, Postgres, or similar) to Zoë Self-Learning AI, the self-learning AI data analyst from Zenlytic. Within an hour, Zoë will automatically scan your schemas, build a governed semantic layer, and begin answering natural-language questions with full citations. You will then train Zoë on your specific business logic, query it both via chat and programmatically through the API, and establish a secure, Git-versioned analytics workflow that eliminates the need for a dedicated data team for ad-hoc reporting.

Prerequisites

  • Admin access to a cloud data warehouse (Snowflake, BigQuery, Databricks, Postgres, or Redshift)
  • A Zenlytic account (start with the self-serve free tier for teams up to 10 users at zenlytic.com)
  • Basic familiarity with SQL schemas and API keys (no modeling required)
  • A set of 4-5 high-impact business questions you want answered (e.g., “What were last month’s revenue trends by region?”, “Why did churn spike in Q2?”)

1. What Is Zoë and Why Technical Founders Should Care

Zoë is a self-learning AI data analyst that connects to your data warehouse and builds a governed semantic layer automatically, reducing setup from months to under an hour. Unlike text-to-SQL tools that merely translate a question into a query, Zoë uses a multi-model AI architecture to understand your business context, handle open-ended strategic questions, and continuously improve with every interaction. It is not a copilot; it is an autonomous data worker that onboards itself.

For technical founders who are drowning in Slack requests for “just a quick query,” Zoë is the off switch. It eliminates the classic 6-month setup tax of enterprise BI, requiring zero YAML configuration, zero dbt models, and zero dedicated data engineers to maintain. Backed by a $9 million Series A from M13, Zenlytic holds a 4.9 out of 5.0 rating on Gartner Peer Insights with a 100% likelihood-to-recommend score from data and analytics leaders. The promise: data to decisions in 59 minutes or less.

If you are tired of your engineering team spending half its time on ad-hoc SQL tickets, Zoë is the tool that finally makes self-service analytics real. And because it learns your business logic through conversation, it gets smarter the more you use it.

2. Under the Hood: Zoë’s Multi-Model Architecture and Semantic Layer

Zoë does not treat every question the same. Its multi-model AI architecture selects the optimal reasoning route per query: fast retrieval for simple lookups, detailed analysis for trend breakdowns, or Python-based computation for forecasts and advanced calculations. This means the same AI that answers “Show me last week’s signups by channel” can also tackle “Which products should we advertise more?” without breaking a sweat.

The cognitive layer automatically maps technical database fields to business-friendly metrics, storing everything in a Git-versioned context layer. Every metric definition, filter, and priority is tracked, auditable, and reversible. This is a game-changer for engineering teams: you get the same code-review workflow you use for application code, now applied to your analytics logic. When Zoë learns a new meaning (e.g., “revenue” should exclude test transactions), that change is committed to the Git log. No more tribal knowledge buried in legacy dashboards.

The semantic layer itself is built in minutes, not months. Zoë scans your warehouse schemas and query history to identify relevant tables, infer relationships, and propose metric definitions. You approve or correct them through conversation. This eliminates the painful manual process of writing YAML or maintaining dbt models for every business concept. For a deeper dive into how this compares to traditional BI, read our guide on how AI turns business data into actionable decisions.

3. Connecting Your Warehouse to Zoë: A Step-by-Step Guide

Connecting your warehouse is the fastest part of the process. Zoë supports Snowflake, BigQuery, Databricks, Postgres, Redshift, and most others. No data modeling is required; you just provide read-only credentials.

Step 3.1: Connect via the UI (No Code)

In the Zenlytic app, navigate to Settings > Data Sources > Add Connection. Select your warehouse type, enter the connection details (host, database, user, role), and test the connection. Once successful, Zoë immediately begins scanning schemas.

Step 3.2: Connect Programmatically via the API

For automated or infrastructure-as-code setups, use the natural language query API to register your warehouse. Below is an example curl command that creates a new Snowflake connection:

curl -X POST https://api.zenlytic.com/v1/connections \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "snowflake",
    "name": "Production Warehouse",
    "credentials": {
      "account": "your_account",
      "user": "your_user",
      "password": "your_password",
      "role": "READ_ONLY",
      "warehouse": "COMPUTE_WH",
      "database": "PROD"
    }
  }'

Expected response:

{
  "id": "conn_abc123",
  "status": "connected",
  "scan_started_at": "2026-06-15T14:30:00Z"
}

Zoë will begin scanning right away. You can check progress with a GET request to /v1/connections/{id}/status. Typically, the first validated answer arrives in under 59 minutes. During this time, Zoë identifies relevant tables, suggests primary keys, and builds an initial semantic layer. You will receive a notification when the layer is ready for review.

Pro tip: Ensure your tables have clear primary keys and consistent naming to speed up automatic relationship detection. If you need more control, you can later fine-tune the schema mapping via conversation or the API.

4. Training Zoë on Your Business Logic Without Writing Code

Once connected, the real magic begins: training the train AI data analyst to understand your unique business metrics. You do this entirely through natural language.

Step 4.1: Ask a Few Representative Questions

Start with your pilot questions. For example:

  • “What were last month’s revenue trends by region?”
  • “Why did user churn increase in Q2?”
  • “Show me weekly active users for the past 90 days.”

Zoë will generate SQL, run it, and return an answer with a chart and citations to the source tables and filters used. You then have the option to correct any misinterpretations. If Zoë calculated revenue including test accounts, you can say: “Revenue should exclude users with role ‘test’.” Zoë updates the metric definition, commits the change to Git, and remembers it forever.

Step 4.2: Define Custom Metrics via Conversation

You can also define entirely new metrics without SQL. Just say: “Create a new metric called ‘Net Revenue Retention’ that subtracts churned revenue from expansion revenue, divided by starting revenue.” Zoë will interpret your intent, propose the calculation, and let you confirm. This is where the multi-model architecture shines: for complex logic, Zoë may write a Python script to compute the metric and store the result in the semantic layer.

The feedback loop is critical. The first few queries might be 80% accurate. By correcting misinterpretations early, you lock in the last 20% of accuracy quickly. Each correction becomes a training example, making Zoë smarter for all future users.

5. Querying Zoë via API and Natural Language: Real Examples

Zoë is not limited to the chat interface. You can submit natural language query API calls from your own applications, Slack commands, or automation scripts. This is invaluable for embedding analytics into your product or automating regular reports.

Step 5.1. Basic API Query

To ask a question programmatically, send a POST request to the /v1/ask endpoint:

curl -X POST https://api.zenlytic.com/v1/ask \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"question": "Why did churn spike in Q2?"}'

Expected response (truncated):

{
  "id": "qry_456def",
  "answer": "Churn increased 15% in Q2 compared to Q1. The primary driver was a 30% drop in retention among users on the basic plan, particularly in the EMEA region (churn rate 22% vs 12% in AMER). Suggested root cause: the may pricing change that removed a key feature for basic users.",
  "visualizations": [
    { "type": "bar", "title": "Churn Rate by Plan and Region", "data": [...] }
  ],
  "citations": [
    { "table": "prod.subscriptions", "filter": "status = 'canceled' and cancel_date between '2026-04-01' and '2026-06-30'" },
    { "table": "prod.users", "filter": "plan = 'basic'" }
  ],
  "sql": "SELECT region, plan, COUNT(*) AS churned FROM ..."
}

Notice the citations. Every answer includes the exact tables and filters used, so you can reproduce the result manually if needed. This is trust by transparency.

Step 5.2. Complex Computation: Python Forecasting

For questions like “Forecast next month’s revenue by channel,” Zoë may spin up a Python environment inside its secure sandbox, run a time-series model, and return the result as a visual and a data table. The API response will include a computation field showing the script executed.

Example script snippet returned:

# Auto-generated by Zoë
import pandas as pd
from statsmodels.tsa.holtwinters import ExponentialSmoothing
# ... model fitting and forecast

You never have to write this code yourself. Zoë handles the fit, the validation, and the visualization. And because everything is auditable, your engineering team can inspect the logic later.

This API capability pairs beautifully with automation tools. See our guide on no-code AI agents that actually work for ideas on how to tie Zoë’s answers into Slack bots or recurring email reports.

6. Governance, Security, and Git-Backed Audits

Zoë’s AI analytics governance is built for engineering teams that demand control. All semantic layer changes are versioned in Git. Every time you correct a metric, add a filter, or define a new calculation, Zoë commits the change with a descriptive message. You can roll back to any previous state, review changes via pull request, and enforce code review policies.

Security is enterprise-grade: SOC 2 Type II certified, compliant with GDPR and HIPAA, with support for SSO/SAML, role-based access control, and even agent-level permissions. This means you can give marketing full access to campaign data while restricting finance metrics to the CFO only, all managed through the semantic layer.

Every answer from Zoë includes explicit citations. This is the opposite of a black-box AI. Stakeholders can see exactly where the data came from, what filters were applied, and how the metric was calculated. This transparency builds trust and makes audits straightforward. No more “the dashboard says X but I don’t know why.”

7. Common Pitfalls and Best Practices for Technical Teams

Even with a self-learning AI, self-service analytics best practices still apply. Here are the mistakes I see most often with technical teams adopting Zoë.

Pitfall 1: Neglecting Data Quality Before Onboarding

Zoë is powerful, but garbage in, garbage out. If your tables have inconsistent naming (e.g., revenue in one schema, Rev in another) or missing primary keys, your semantic layer will be brittle. Clean your core tables first: ensure clear keys, consistent column types, and remove obvious junk data. Zoë will thank you with faster, more accurate schema detection.

Pitfall 2: Skipping the Feedback Loop

Do not treat Zoë as a “fire and forget” tool. The first day of use will uncover misinterpretations. If you ignore them, those 80% accurate answers will stay at 80%, and trust will erode. Invest an hour in the pilot phase: have a few power users ask their real questions and correct every mistake. That feedback locks in the last 20% of accuracy and makes Zoë a trusted advisor.

Pitfall 3: Expanding Too Fast

Start with a small cross-functional team (maybe your growth and finance leads) and a focused set of four or five high-impact questions. Define clear ROI goals, like reducing ad-hoc SQL tickets by 80% or cutting report generation time from 10 hours to 5 minutes. Once Zoë proves reliable on that core dataset, add more tables (marketing spend, support tickets) and repeat the validation loop. Gradual expansion prevents confusion and maintains trust.

Finally, keep a lean data-ops function even after Zoë takes over. Someone should periodically validate new insights and update metadata. But that one person can replace a whole team of report writers.

Next Steps

You now have everything you need to deploy Zoë in production this afternoon. Start by connecting your warehouse, asking those pilot questions, and correcting a few answers. Within a week, your team will be getting trusted insights in seconds instead of waiting days for SQL tickets. For more on scaling this approach, read about connecting AI agents to your business for real automation, or check out our no-code playbook for building AI-powered tools.

Cover photo by Steve A Johnson on Pexels.