AI Engineer World's Fair 2026

AI in GTM at Notion — Flora Liu

Flora Liu· Notion21:14

Loading slides…

Read the talk

AI in GTM at Notion

Flora Liu explains how Notion joins customer context, shared eligibility rules, durable agent workflows, and human judgment into one go-to-market system—and why the notes that change a sales decision belong in its architecture.

From a talk by Flora Liu

At a glance

Ideas worth remembering

  • Customer context must include the notes that change a decision, such as a lost champion, a legal blockage, or a request to stop contact.

  • Snowflake computes modeled customer entities; DynamoDB serves denormalized profiles and associated artifacts by shared IDs; Notion makes that context usable by reps and agents.

  • Shared eligibility and routing coordinate the next action across channels. Sales-assist agents prepare work for human approval, while prospect form input remains untrusted.

  • Temporal handles workflow retries, deduplication, and resumption. LLM traces help evaluate output quality, and linked engagement outcomes inform subsequent decisions.

  • Start with a strong, repeated human workflow and preserve customer-specific context you can debug. Rent general infrastructure where vendors already do the job well.

From marketing operations to a systems problem

Go-to-market work at Notion ran through a “spider web” of tools, stitched together by customer notes, proposals, and contracts. Flora Liu, an engineer moving from Product Growth to GTM Engineering, describes a small team’s effort to turn that environment into a unified system. A year earlier, she would have called it a marketing operations problem. Now it looked like one of the most interesting distributed systems problems she had worked on. 0:12

The pieces were familiar: lifecycle messaging, product recommendations, sales automation, customer data, and onboarding. What changed was the team’s sense of what it could build. A winter-break video-game project convinced Notion’s CEO that engineering could reach previously unwieldy or expensive problems. Agentic technology also expanded the small team’s ability to execute. Work that had been approached as separate automations became a candidate for a coordinated system.

Notion customers move between self-serve adoption and sales assistance. They experience one journey, while internally that journey crossed disconnected systems. After a call, a rep might switch tools to research the customer and draft a follow-up. Each switch consumed attention that could have gone toward understanding the customer’s problem. The rep’s strength was reading human signals during the buying process; managing the surrounding software imposed another job.

Marketing, sales, and customer operations used different tools and made separate decisions about the same customer. The intended replacement was a programmable, proactive, continuous decisioning system spanning self-serve growth and sales assistance: one place to decide the next step as the customer’s circumstances changed.

Customer records were spread across Salesforce, Gong, Outreach, and other vendors. Product usage lived in Snowflake; years of important context lived in documents and meeting notes. Employees already used MCPs and their own agents, but largely within their departments—“single-player mode.” Automating a small slice of the journey helped locally while leaving cross-department coordination unresolved.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

The facts that make automation dangerous

Connecting the tools exposed three different failure mechanisms:

  • Incorrect identity: Conflicting records could attach a contact to the wrong account. One bad mapping was enough to cost a rep’s trust.
  • Stale context: Every vendor added another hop. Delayed updates meant an automation could act on a customer situation that had already changed.
  • Unread notes: A buying champion leaving, a request to stop contact, or a legal blockage could sit in free-form text while the structured record still looked actionable. 4:34

The last problem changes what counts as useful customer data. Product activity alone cannot tell an automation whether outreach is appropriate. A note about legal or a lost champion can alter the next decision even when the account’s usage remains promising. If the system cannot read and process that note, scaling the workflow also scales the chance of a consequential mistake.

The project brought CX, RevOps, Product, Engineering, and Sales together to find the repeated structure beneath these workflows. The customer journey crossed all of their systems; a department-specific agent could only see part of the job.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:34 · section reference included

Four questions, with humans and agents in the same loop

Every workflow reduced to four questions, which became four architectural layers:

  • Know: What do we know about the customer? Assemble context the system can trust.
  • Decide: What should happen next? Choose a single next best step.
  • Act: How do we execute safely? Produce a concrete action, such as a lifecycle email, an in-app nudge, or a task for a rep.
  • Learn: Did it work? Observe the outcome and feed it back into decisioning. 5:43

Humans and agents use that same loop and read the same context. Agents gather information, research, draft recommendations, and write artifacts at scale. Humans judge whether a recommendation makes sense, add nuance, choose the next move, and own the customer relationship. An agent becomes another operator in the business system, with a different role from the person reviewing its work.

For sales-assist workflows, the team deliberately kept agents from speaking directly to customers. Humans approve their work by default. A contact-sales form remains untrusted user input even when an agent processes it; inserting a model into the workflow does not grant a prospect’s text permission to control customer-facing actions. 7:23

Eligibility and routing also moved into shared code. Previously, an email tool could have one eligibility check while sales used another. Central definitions for segmentation and signals let product, sales, and engineering consume the same rules. A single classifier routes the customer’s next step, with the intended effect of preventing double sends and keeping communication cohesive.

The ownership decision followed the same reasoning: keep control of customer context and rent general-purpose services. A lean team did not need to build an email vendor or an enrichment service; Notion uses Clay for enrichment. Its understanding of its own customers was the part the team wanted to keep.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:43 · section reference included

Snowflake computes context; DynamoDB serves it

Snowflake gathers data from the GTM vendors and computes a consolidated customer view. Daily transforms, with real-time processing in some cases, produce modeled, versioned entities: accounts, contacts, workspaces, eligibility, and facts. Ownership and timestamps identify where the information comes from and when it was recorded. 9:01

DynamoDB handles the fast lookup path. The system publishes a denormalized profile that an agent can retrieve by key in milliseconds without joins. Research snippets, summarized notes, and rolling summaries use the same IDs. A downstream reader can fetch structured customer information and generated context together, rather than reconstructing the account across several services.

How does one customer view reach both an agent and a rep? The diagram separates computing the modeled data, serving a profile, and making the context usable in Notion. The shared IDs connect structured records with research and summaries. Fast profile retrieval addresses query latency; freshness still depends on upstream processing, including transforms that run daily.

Notion brings product usage, vendor activity, research reports, and notes into the familiar workspace where GTM teams already work. Reps can investigate an account, ask questions, and initiate actions such as sending to Outreach without opening seven tabs. Humans, workflows, and agents can use the same customer context. In Liu’s phrase, the team is “using Notion to grow Notion.”

From vendor records to a shared customer view

GTM vendor dataCustomer records and activity from the GTM stack.
SnowflakeTransforms produce versioned entities, ownership, and timestamps.
DynamoDB profileDenormalized, key-addressable customer context; no joins.
Agent artifactsResearch snippets, summarized notes, and rolling summaries keyed by the same IDs.
Agent queriesQuick reads of the customer profile and associated context.
Notion customer viewUsage, activity, research, and notes in the rep’s workspace.
Snowflake models customer data, DynamoDB provides key-based reads, and Notion makes structured and unstructured context available to the people doing the work.
Read the diagram as text
  • GTM vendor data. Customer records and activity from the GTM stack.
  • Snowflake. Transforms produce versioned entities, ownership, and timestamps.
  • DynamoDB profile. Denormalized, key-addressable customer context; no joins.
  • Agent artifacts. Research snippets, summarized notes, and rolling summaries keyed by the same IDs.
  • Agent queries. Quick reads of the customer profile and associated context.
  • Notion customer view. Usage, activity, research, and notes in the rep’s workspace.
  • GTM vendor data → Snowflake: Ingest.
  • Snowflake → DynamoDB profile: Publish modeled profile.
  • Agent artifacts → DynamoDB profile: Persist under shared IDs.
  • DynamoDB profile → Agent queries: Retrieve by key.
  • DynamoDB profile → Notion customer view: Bring consolidated context into the workspace.
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:01 · section reference included

A signal turns a customer event into a task

A signal is a customer event important enough to change what should happen next. The examples fall into two groups:

  • Customer-driven signals: A customer hits an AI limit or contacts sales.
  • External signals: A company raises funding, changes its hiring activity, or shifts its technology stack. These let the system identify a possible next step before the customer asks for help. 11:20

The signal service watches the customer profile, checks whether an action is available, assigns its owner, and emits a concrete task. That owner can be a human or an agent. A rep’s task lands in a Notion database, turning an event into work someone can inspect and act on.

A customer without a signal still has a path through the system. A predictive marketing engine recommends relevant product features and automatically sends lifecycle emails, in-app nudges, or other multichannel messages to encourage adoption. This distinguishes the human-approved sales-assist workflow from automated lifecycle communication.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:20 · section reference included

Research and drafting run inside a durable workflow

The team shadowed its strongest reps and encoded repetitive sales work as a durable multi-agent workflow. Every signal becomes a workflow on Temporal. Enrichment, web search, and draft generation involve network calls that can fail or hit rate limits. Temporal handles retries, deduplication, and resuming where a failure left off, while the team writes the sequential sales logic. One malformed transcript need not take down the whole batch. 12:55

Follow the cold-outbound example from event to reviewable work. A research sub-agent runs web research concurrently. The workflow then generates three email drafts and scores them. A review agent chooses the highest-scoring draft and revises it if needed; the drafting and review process loops to improve the result. When ready, the selected email appears inside the sales task. The observable change for the rep is concrete: the task now contains researched copy to review instead of an empty starting point.

Where does the email become customer-facing? The diagram separates the automated preparation loop from human review. Scoring chooses among drafts; it does not authorize a send. The sales rep still adds judgment and approves the work.

A reactive follow-up uses a different input. A Gong transcript arrives after a call; an agent parses it, extracts sales information such as metrics, the economic buyer, decision criteria, the plan, and the champion, then drafts a grounded follow-up. Every LLM step is traced so the team can evaluate quality and improve the workflow. Durable execution keeps the process running; tracing supplies a way to inspect what it produces.

Cold outbound: prepare, compare, then hand off

Cold-outbound signalTemporal workflow — Starts a durable workflow on Temporal.
Concurrent researchTemporal workflow — A research sub-agent gathers web context.
Three scored draftsTemporal workflow — Generate alternatives and score the emails.
Review agentTemporal workflow — Selects the highest-scoring draft and makes updates.
Sales task with draftPrepared copy becomes available for the rep to act on.
Human review and approvalThe rep adds judgment, taste, and customer understanding.
Temporal contains the retryable workflow. Agents research and improve drafts; the resulting task waits for the rep’s judgment.
Read the diagram as text
  • Cold-outbound signal. Temporal workflow — Starts a durable workflow on Temporal.
  • Concurrent research. Temporal workflow — A research sub-agent gathers web context.
  • Three scored drafts. Temporal workflow — Generate alternatives and score the emails.
  • Review agent. Temporal workflow — Selects the highest-scoring draft and makes updates.
  • Sales task with draft. Prepared copy becomes available for the rep to act on.
  • Human review and approval. The rep adds judgment, taste, and customer understanding.
  • Cold-outbound signal → Concurrent research: Begin preparation.
  • Concurrent research → Three scored drafts: Draft from research.
  • Three scored drafts → Review agent: Compare scores.
  • Review agent → Three scored drafts: Improve drafts in a loop.
  • Review agent → Sales task with draft: Place ready draft in task.
  • Sales task with draft → Human review and approval: Review before customer contact.
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:55 · section reference included

Outcomes change the next decision—and the rep’s next morning

Each action has a decision log, and its outcome links back to the decision that caused it. Engagement history returns to the decision layer, where the system can choose to continue a thread, advance to another step, or pivot. Lifecycle-message performance feeds the same kind of loop. The mechanism is feedback into future choices: completing an automation leaves information for the next decision. 14:53

In the customer view, a rep sees usage and recent activity, then asks an agent questions against that same context. Notion custom agents can also access the data for recurring workflows. In the task view, the rep begins the day with prioritized work and pre-researched email drafts. Preparation happens before the rep opens the task, leaving room for the human’s judgment and “sales secret sauce.” 16:14

The ambition extends beyond helping an experienced rep move faster. The tasks and drafts expose which signals matter, which playbooks the team uses, and what good follow-up looks like. A new rep can learn from patterns captured from the strongest reps without requiring every lesson to be passed down manually. The workflow becomes part of how the team raises its floor.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:53 · section reference included

Own the context, rent the general parts, and start with good human work

Build versus buy is a decision at each layer. The team builds the customer-specific data model and internal agent behavior while renting general services such as orchestration, email, and CRM. Liu’s judgment is that internal agents can be cheaper and faster to build than people assume, especially when the team already understands the data they need. That does not make rebuilding every vendor a useful investment. 17:16

Control of context also preserves the ability to debug unusual customer models and workflows. Notion provides Markdown, databases, and hierarchies that agents can navigate, together with a workspace humans can use. Shared access to synced information lets engineers, agents, and GTM teams work from the same account understanding.

The early business signs were promising. Over thirteen weeks, Liu reports an increase in enterprise reps’ qualified opportunities. On the lifecycle-marketing side, users receiving context-aware recommendations were 63% more likely to take the next step. These are preliminary results from a system still being built; the talk does not define that next step, describe the comparison design, or quantify the opportunity increase, so the figures do not establish an equivalent gain in closed deals. 18:48

The closing implementation advice begins before the first agent:

  • Shadow your best human: The tabs and tools reveal both the chaos and the specification. Copying a mediocre process produces a mediocre agent.
  • Start with a legible workflow: Choose work that is documented and repeated, and retain human involvement where mistakes carry risk.
  • Model the primitives: Entities, context, triggers, actions, and eligibility rules turn an unfamiliar GTM process into something engineers can reason about. 19:18

The final advice is to be headless by default: make the underlying context available to agents as operators, rather than requiring every operation to pass through a human interface. Both people and agents need to read the same information; otherwise, their views become two systems that drift apart. At the time of the talk, humans remained the primary consumers of Notion’s GTM data, with agents helping around the edges. Moving agents from drafting to acting within guardrails was the direction Liu anticipated. Shared context prepares for that change while today’s sales-assist workflow keeps customer contact under human control. Liu closes with an invitation to ask questions about the work.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

17:16 · section reference included

Resources

Transcript download
Full transcript

Loading transcript…