Small data errors add up: 80% accurate steps end at 33%

If every step in a chain of data work is 80% accurate, the final answer is right only about a third of the time. That is the warning from Anthony Goldbloom, founder of the sales data company Sumble, as reported by SaaStr. It matters more now that AI agents, which are software helpers that act on their own, use this data to decide who to contact.

How the math works

Getting a precise count of people in one job function takes several steps in a row. Errors multiply across them. Multiply 0.8 by itself five times and you get about 33%. Sumble aims for 99% on each step, which leaves about 95% at the end. Goldbloom was clear that Sumble is not perfect yet and that 99% per step is the goal.

He also pointed to how stale data used to be fought. Bloomberg has had around 9,000 people in lower-cost locations cleaning data all day. Agents now make that work much cheaper.

What Sumble does

Goldbloom previously founded Kaggle, a machine learning community that Google acquired in 2017. He built Sumble with sales teams at companies such as Snowflake, Databricks, Figma, and Wiz. His view is that the first wave of AI in sales was about sending more messages. Sumble uses AI to decide precisely who to target before anything is sent.

It scores accounts daily, not once a year, and it ranks the people inside each account. Two examples from the talk:

• OpenAI ranked third on Sumble's own target list. Its sales headcount grew 395% in a year and marketing grew 269%.
• A lead who ranked sixth at an account ranked 102nd overall got an instant alert, so the rep knew not to let it sit.

What happened at Nooks and SaaStr

Nooks started with Sumble on territory assignments. An accurate count of sales development reps at each company surfaced 700 accounts Nooks did not know about. That was enough to justify doubling its enterprise sales team. Goldbloom said this came from territory planning alone.

SaaStr has used Sumble for six to eight weeks. It is wired into Prospector, SaaStr's internal agent for inbound leads, outbound outreach, renewals, and list enrichment. Almost all usage is headless, meaning agents query it directly and people rarely log in. SaaStr's team says Sumble's data has been the most consistently current of the tools it tested.

What to watch for

The 99% figure is a target, not a result. The accuracy claims here come from the company and its customers, and the material gives no independent test. Still, the arithmetic applies to any data you buy or any AI tool that picks leads for you.

• Ask how accurate each step is, not only the overall number.
• Ask how often the data is refreshed.
• Check a small sample yourself before an agent acts on it at scale.