Sumble：AI销售数据精度衰减机制与账户评分法

Amelia called Sumble the best company you probably haven’t heard of yet. Its customers include Snowflake, Databricks, Figma, Wiz, Monaco, and a long list of AI companies like Baseten, Vercel, Modal, and CodeRabbit. SaaStr has been running it for the last six to eight weeks.

Founder Anthony Goldbloom doesn’t come from go-to-market. He previously founded Kaggle, the machine learning community that Google acquired in 2017. Nine years later it still runs under its own brand as a division of DeepMind. He built Sumble by working closely with GTM teams at the companies above to figure out what data an agent-first sales motion actually needs.

How SaaStr Uses Sumble: Mostly Headless, Through Our Agents

Before Anthony’s part, Amelia explained how SaaStr runs it. Sumble is wired into Prospector, our internal agent that handles inbound, outbound, renewals, and list enrichment. Our agents call it to find the right contact at a company, map who reports to whom, and check whether an inbound lead is the actual decision maker.

For an events business, headcount by function matters in specific ways. How many people a sponsor has in go-to-market affects how many they’re likely to bring to their booth at SaaStr AI.

Amelia also built a Claude skill that chains Exa, Sumble, and Clay, with Sumble as the first pass. Between the three, it finds almost anyone we need for outbound, renewals, or POCs. Of the tools we’ve tested, Sumble’s data has been the most consistently current.

Almost all of SaaStr’s usage is headless. Amelia rarely logs in. The agents query it directly as part of the inbound and outbound flows, and Sumble’s team noticed the steady drumbeat of usage coming from agents instead of people.

The First Wave of AI in GTM Saturated Every Channel

Anthony’s view is that the first wave of AI in go-to-market focused on sending more. AI SDRs were the hot category. Sending a message became nearly free, and email, LinkedIn, and other digital channels filled up as a result.

Sumble uses AI for the opposite job: deciding precisely who to target before anything gets sent. Anthony splits precision into two parts, granularity and accuracy.

One Walmart Job Post Revealed a Team, a Migration, and a Tech Stack

To show granularity, Anthony walked through a single Walmart job posting. From that one post you can learn:

• There’s a Catalog Engineering team, and from how it’s described alongside Walmart Global Tech, it likely reports into Global Tech
• The team maintains an enterprise-scale catalog platform
• It’s migrating its catalog orchestration service to a new platform
• The new platform likely includes NoSQL components (Cosmos DB and Cassandra are mentioned) and containerization (Docker and Kubernetes)
• The location, title, technologies, and team name are enough to make a reasonable guess about who leads the project

Now apply that to every job post Walmart has ever published, plus its financial filings, earnings call transcripts, engineering blog posts, and employee podcast appearances. You get a detailed picture of what’s happening inside the account, team by team.

Five Steps at 80% Accuracy Leaves You With 33%

The second part of precision is accuracy. Stale data is the complaint everyone has about their data vendor. Historically, keeping data fresh took enormous manual effort. Anthony pointed out that Bloomberg has had around 9,000 people in lower-cost locations cleaning data all day. Agents now make that work much cheaper.

Some Sumble customers need a precise count of people in one job function:

• Nooks sells a power dialer and needs an accurate count of SDRs at each company
• Figma needs a count of UX designers, separate from chip, fashion, jewelry, or CAD designers
• Weights & Biases needs a count of machine learning engineers

Getting that count right takes several steps in a row, and errors multiply across them. Suppose each step is 80% accurate:

• Company data
• Matching people to companies
• Freshness of people records
• Job function classification (UX designer vs. chip designer)
• Job level classification (ICs, who consume seats, vs. managers)

0.8 multiplied five times comes out to about 33%. Sumble targets 99% on each step, which leaves about 95% accuracy at the end. Anthony was clear that Sumble isn’t perfect yet, and that 99% per step is the goal.

Daily Account Scoring: OpenAI’s Sales Headcount Grew 395% in a Year

Anthony recommends two things to set up first once a team has precise data: account scoring that refreshes daily, and contact scoring.

Most companies already score accounts once a year before territory planning, using employee count, industry, revenue, and some technographics. Sumble rescores daily on much more granular attributes, and the score doubles as a starting point for the rep to understand the account.

He showed OpenAI, ranked third on Sumble’s own target list. The score breaks into three parts:

• Size: how large the opportunity is today
• Growth and momentum: how fast it’s growing. Fast-growing companies also tend to have processes that are breaking, which makes them easier to sell into
• Concentration: how central Sumble’s product is to what the company does

Under size, Sumble tracks OpenAI’s headcount of GTM engineers, salespeople, marketers, RevOps, and SDRs. It checks whether the GTM team uses Clay, Cursor, Claude Code, or Codex as a sign of being AI-forward, and whether they already use GTM enrichment. It found six teams at OpenAI using some enrichment tool. A rep can click into each team, see sub-teams and tools like Clay or LinkedIn Sales Navigator, and drill down to the individual people. Clicking the GTM engineer count lists every GTM engineer at OpenAI.

Under growth, OpenAI’s sales headcount grew 395% over the past year and marketing grew 269%.

Concentration scored low. Only 5.4% of OpenAI’s headcount is sales reps. It isn’t a heavily sales-led company, so sales tooling isn’t a top priority for Sam Altman or Greg Brockman, and that pulls the score down.

Contact Scoring Ranks the People, Within an Account and Across All of Them

Account scoring tells you which companies matter. Contact scoring ranks the people inside each account, so a rep immediately sees who to get in front of and marketing knows who to put in a campaign. Sumble also ranks contacts across accounts, producing a list of the top 1,000 people in the world it most wants to meet.

A Slack Alert That Says #102 Account, #6 Person

Anthony showed a Slack alert from Sumble’s own channel. The head of revenue operations at Sumble’s 102nd-ranked account had started a product-led session, and that person was ranked sixth at their company. The AE on the account knew right away this lead wasn’t one to let sit. If the alert had said account #10,000 and person #400, she could reasonably leave it for later.

Precise rankings also drive marketing decisions:

• Who to invite to a dinner in New York
• Who belongs in a 10,000-person contact-level ad audience
• Which events your top 10 targets are speaking at. If the fifth-ranked person on your list is speaking at a conference in Phoenix, email them asking to meet there. If they say yes, book the trip and the ticket.

Nooks Found 700 Accounts and Doubled Its Enterprise AE Team

Nooks started with Sumble on territory assignments. An accurate SDR count per company surfaced 700 accounts Nooks didn’t know about. That was enough to justify doubling its enterprise AE team. Anthony noted this came from territory planning alone, before any of the advertising or inbound conversion use cases.

What Our Agent Pulled on a BambooHR Inbound

Amelia then showed how Sumble works inside Prospector on a real inbound from BambooHR. SaaStr routes inbound into a queue in 10K instead of Slack alerts. She clicks “generate pitch” and the agent runs in order.