The Numbers
Sumble, the sales intelligence startup founded by Kaggle's Anthony Goldbloom, closed $38.5 million across seed ($8.5M, Coatue) and Series A ($30M, Canaan Partners) rounds. The company reported 550% year-on-year growth, 19 enterprise customers including Snowflake, Databricks, Figma and Wiz, and tens of thousands of users. No ANZ presence confirmed.
Why It Matters: The 80% to 33% Problem
Most contact databases claim 80% accuracy. That sounds acceptable until you stack the steps your SDR actually takes: match company data, verify the person still works there, classify their function (ML engineer vs. data analyst), classify their level (IC vs. manager), check reporting structure. Each step at 80% accuracy compounds to 33% by step five.
Sumble's pitch: build a knowledge graph from job posts, earnings calls, filings and social signals, then use AI to maintain 99% accuracy at each step. End result sits around 95%, which matters when you are scoring 12,000 accounts or running agent-based outbound.
The Nooks Case Study
Nooks sells a power dialer and needs accurate SDR headcount to qualify accounts. Sumble analysed their book and found roughly 700 accounts that met qualification criteria but were absent from their system, representing about $15 million in potential whitespace ARR. The analysis also flagged around 1,000 accounts misclassified because SDR-count data was inflated, which had skewed territory assignments.
Nooks raised a $22 million Series A in April 2024 and grew from 10 to nearly 50 employees. No disclosed ARR or quota structure.
Data Decay at Agent Scale
Goldbloom's background is machine learning, not go-to-market. His thesis: human reps work around bad data (they call, they check LinkedIn, they triangulate). AI agents cannot. When your agent queries a database for the VP Engineering contact and gets someone who left four months ago, the sequence fails.
Sumble customers wire it into agent workflows instead of using the UI. One example: an events business uses it to count go-to-market headcount, which predicts booth staffing at conferences. The agent queries Sumble, enriches the lead, scores the account and moves on. Most usage is now headless API calls, not dashboard logins.
What This Means for Territory Planning
Most teams score accounts once a year: employee count, revenue, industry, maybe some tech stack flags. Sumble rescores daily using signals like hiring velocity (OpenAI's sales headcount grew 395% in one year, per the case study), team structure changes and tech migrations pulled from job posts. Fast growth correlates with broken processes, which makes accounts easier to sell into.
The market is ZoomInfo, Apollo and similar platforms. Differentiation is live context and agent-ready accuracy, not contact volume. Worth noting: Bloomberg employs roughly 9,000 people in lower-cost locations to clean financial data manually. Sumble's bet is that AI agents make that model obsolete for go-to-market intelligence.
The Skeptic Take
No disclosed revenue. Headcount sits between 24 and 35 employees, so this is still early. The Nooks case study is compelling, but one data point. The 80% to 33% calculation assumes independence across steps, which may not hold. And agent-first workflows are still a hypothesis for most sales organisations, not proven infrastructure.
That said: if your CRM shows 1,200 accounts in your patch and 15% of them are misclassified or stale, that is 180 accounts you are either ignoring or working incorrectly. Sumble's pitch is that the math gets worse when you add agents, because agents do not triangulate the way humans do.