Descriptions:
The champion just left. Do not contact this customer again. Blocked in legal. The facts that decide what a rep does next usually sit in someone’s meeting notes, and Flora Liu’s point is that an automation which cannot read them will eventually do something catastrophically wrong. Liu, an engineer on Notion’s GTM engineering team, says a year ago this looked like a marketing ops problem and now looks like one of the more interesting distributed systems problems they have worked on. The starting state was a spiderweb: customer data across Salesforce, Gong and Outreach, product usage in Snowflake, a decade of context in docs, and every department quietly building its own agents in single player mode.
The team reduced every workflow to four questions: what do we know, what should happen next, how do we execute that safely, and did it work. Those became four layers, held together by one rule: humans and agents operate on the same substrate, not an AI layer bolted on top. Snowflake computes the truth, DynamoDB serves it as a denormalized profile agents query in milliseconds, and it lands back in Notion where reps already work. Signals turn customer events into tasks, including external ones like a funding round or a shift in tech stack, and each becomes a durable Temporal workflow so one malformed transcript cannot take down a batch. Agents never speak to customers; a contact sales form counts as untrusted input. Thirteen weeks in, enterprise reps log more qualified opportunities, and users given context aware recommendations were 63% more likely to take the next step.
Speaker info:
– https://twitter.com/floppyliu
– https://www.linkedin.com/in/flofloliu/
– https://www.flofloliu.com/
Timestamps:
0:00 – From marketing ops problem to distributed systems problem
2:23 – One journey for the customer, disconnected systems inside
4:34 – Data quality, latency, and the notes nobody can parse
5:36 – Four questions that became four layers
6:39 – Humans and agents on the same loop
7:29 – Agents never talk to the customer
9:01 – Snowflake computes truth, DynamoDB serves it
11:20 – Signals: turning customer events into tasks
13:04 – Shadowing the best reps, encoded on Temporal
16:04 – A rep’s day starting from a prioritized task box
17:21 – Build or buy, decided per layer
18:52 – Early numbers and takeaways
