Engineering a Unified GTM System at Notion

A Unified GTM Architecture

Notion moved away from fragmented, department-specific tooling by building a centralized GTM system that treats humans and agents as equal operators on a shared substrate. The architecture is defined by four core layers: a context layer (Know), a decision engine (Decide), an execution layer (Act), and a feedback loop (Learn). By centralizing data in Snowflake and serving it via DynamoDB, the team eliminated data latency and conflicting records that previously caused agents to act on stale or incorrect information.

Engineering the Workflow

To ensure reliability and scalability, the team modeled GTM workflows as durable tasks using Temporal. This allows for complex, multi-step processes—such as researching prospects, drafting emails, and scoring outputs—to be retried automatically without failing the entire batch.

• Data Normalization: Ingested data from Salesforce, Gong, Outreach, and product usage into Snowflake to create versioned entities (accounts, contacts, workspaces).
• Key-Value Serving: Published denormalized profiles to DynamoDB, allowing agents to query customer context in milliseconds without expensive joins.
• Signal-Driven Execution: Defined "signals" (e.g., funding rounds, tech stack shifts, or product usage limits) as triggers that automatically initiate specific workflows.
• Human-in-the-loop: Designed agents to perform research and drafting while keeping humans as the final approvers, treating contact forms as untrusted input to maintain security boundaries.

Build vs. Buy Strategy

Notion adopted a per-layer strategy for tool selection. They chose to rent general-purpose infrastructure like email vendors and enrichment services (e.g., Clay) while building their own context layer. This ensures the company retains ownership of its unique data models and workflows, which are stored in Notion databases and markdown, formats that are natively readable by both human employees and AI agents.

Performance Impact

In the first 13 weeks of deployment, the system demonstrated measurable improvements in sales and marketing efficiency:

• Enterprise sales representatives logged a higher volume of qualified opportunities.
• Users receiving context-aware recommendations were 63% more likely to take the next step in the product journey.