A few weeks ago I killed my most-used feature. It was an automated pipeline that pulled listings from a commercial marketplace. By volume of records created it was the most used part of the product. It cost about $2,300 a month to run: $1,000 to $1,500 on residential proxies, and roughly $1,000 in tokens for an AI agent whose only job was repairing the parser every time the site changed its layout. The cost to serve those accounts was about ten times what they paid. Then I found that the retention data justifying the feature had been read backwards. Users who ran the automation averaged 38 properties created in week one, versus 0.7 for users who never came back. Recalculated by the share of users inside each cohort instead of average volume, it inverted: manual document upload appeared in 56% of retained users versus 5% of churned ones. The 38 was inflated by a handful of accounts running bulk batches. I nearly built a strategy on the behaviour of five people. What is left is narrower and cheaper to run. You upload an offering memorandum, a rent roll or a T-12, and get back a structured deal record plus a chat over that file, with every answer cited back to the document. Has anyone else had to learn the difference between average volume and cohort share the hard way?

CRE Pipeline

Parse commercial real estate PDFs into structured records.

Parse commercial real estate PDFs into structured records.