AI in GTM at Notion — Flora Liu

From Marketing Ops Problem to a Distributed Systems Challenge

Flora Liu, an engineer on Notion’s GTM engineering team, articulates a shift in perspective: what was once considered a marketing operations problem has evolved into “one of the most interesting distributed systems problems.” The traditional Go-to-Market (GTM) landscape is often a “spiderweb” of disparate tools connected by manual exchanges of notes, proposals, and contracts. Notion’s objective is to unify this fragmented system using software engineering principles, with a significant emphasis on agentic technology. The fundamental issue identified is the customer’s singular journey being internally supported by disconnected systems, leading to inefficiencies, human error, and substantial cognitive load on teams like sales representatives.

Challenges of Disconnected Systems and Unstructured Data

Notion, leveraging its platform as a collaborative brain for both humans and agents, pinpointed several critical challenges:

• Lack of a Single Source of Truth: Customer data was dispersed across numerous platforms including Salesforce, Gong, Outreach, ZoomInfo. Product usage data resided in Snowflake, and invaluable historical context—often a decade’s worth—was locked away in meeting notes and documents.
• Departmental Silos: Each department independently developed its own tools and agents, operating in a “single player mode,” which resulted in fragmented, non-holistic solutions.
• Data Quality and Latency: Issues arose from conflicting records, incorrect data mappings, and staleness caused by multiple vendor hops, eroding trust and leading to decisions based on outdated information.
• Prevalence of Unstructured Data: Critical customer intelligence, such as “the champion just left,” “don’t contact this customer again,” or “blocked in legal,” was buried in unstructured notes. Automations incapable of interpreting such context posed a risk of making critically incorrect or damaging decisions.

A Four-Layered Architecture: Know, Decide, Act, Learn

To systematically address this complexity, Notion’s team reduced every workflow to four core questions, which then became the architectural blueprint for their system:

1. Know: What information do we possess about the customer? (Establishing a trustworthy context layer).
2. Decide: What is the optimal next step for this customer? (Selecting the single best action).
3. Act: How do we safely execute this chosen action? (Triggering a concrete output: an email, an in-app nudge, or a task for a human representative).
4. Learn: Did the action succeed? (Monitoring outcomes and feeding insights back into the decision-making process for continuous improvement).

Humans and Agents on the Same Substrate

A cornerstone of Notion’s philosophy is that “humans and agents are operating on the same loop.” This implies that both entities interact with the identical context and system, eschewing an AI layer as an add-on. Agents are tasked with scalable, repetitive work such as context aggregation, research, and drafting, while humans contribute judgment, nuance, and cultivate customer relationships. This architectural approach positions agents as “another operator within the same system as humans.”

Key Implementation Decisions

• Agent Interaction with Customers: Agents are deliberately kept from direct customer communication. In sales assist workflows, human approval is a default requirement. Prospect-initiated inputs (e.g., contact sales forms) are treated as untrusted to maintain security and trust boundaries.
• First-Class Routing and Eligibility: Customer segmentation rules and signal definitions are centralized. This unification prevents redundant communications and ensures a coherent customer interaction strategy across all touchpoints.
• Ownership of the Context Layer: Notion made a strategic decision to build and own its core context layer, believing its deep understanding of its customer data model provides a critical edge. Peripheral services, such as email vendors, data enrichment tools (like Clay), and orchestration platforms, are rented.

The Technical Stack: Snowflake, DynamoDB, Notion

1. Snowflake (Compute Truth): Serves as the data warehouse where all GTM vendor data is ingested. It performs daily (and in some cases, real-time) transformations to produce a refined set of modeled, versioned entities including accounts, contacts, workspaces, eligibility data, and facts.
2. DynamoDB (Serve Truth): Acts as the key-value store, publishing a denormalized, key-addressable customer profile. This enables agents to query data in milliseconds without complex joins. DynamoDB also persists agent-generated artifacts like research snippets and summarized notes.
3. Notion (Unified Interface & Context): This is the nexus where both structured and unstructured data coalesce. Notion GTM teams utilize Notion itself to explore customer context, investigate accounts, and initiate actions, significantly reducing the need to switch between multiple applications. Notion thus functions as the literal shared source of truth for humans, automated workflows, and AI agents.

Signals and Durable Workflows with Temporal

• Signals: These are specific customer events deemed significant enough to alter the next optimal action for a customer. Signals can originate from user actions (e.g., reaching an AI usage limit, initiating contact with sales) or from external sources (e.g., a company’s funding round, shifts in their tech stack), enabling proactive engagement.
• Temporal for Orchestration: Each signal triggers a durable, multi-agent workflow orchestrated by Temporal. This platform manages retries, deduplication, and ensures workflow resilience against failures. For instance, a single malformed transcript cannot disrupt an entire batch of processing. Examples include

• Tags:
• AI Agents
• Go-to-Market
• Distributed Systems
• Feature
• Workflow Automation
• Temporal
• Human-in-the-Loop
• Snowflake
• Customer Data
• Notion
• Marketing Automation
• Context Layer
• GTM
• DynamoDB
• Sales Automation
• Revenue Operations