This+that turned 45,047 Enron emails into a cited AI wiki for $11,000

A pipeline read 45,047 old Enron emails and wrote a wiki about them. Every fact cites the message it came from.

The team behind it is This+that. The company builds what it calls a "company brain" for teams. It ran the Enron emails through a four-step pipeline. The output was 214 pages about organizations and their people. The page lists 49,401 facts, each with a link to the source email.

What the pipeline does

The four steps are listed on the page. First the system finds entity types. Then it finds the entities themselves. Then it builds a page for each one. Last, a second AI checks the work.

The input was eight mailboxes from the public Enron corpus. The corpus has been maintained by Carnegie Mellon since 2003. The pipeline looked only at the emails. The team said on Hacker News that it stopped the model from adding outside knowledge. It said that step took real effort.

The team found 551 entities inside Enron and 929 outside it. It wrote pages only for organizations mentioned in 25 or more messages. Products, facilities, market points and events were counted but not given pages.

The audit

A second AI checked a random sample of 500 facts. This+that reports that 93 percent were fully supported by the source email. It reports zero facts wrong at their core. The team also wrote on Hacker News that 62,205 candidate facts went through the pipeline. Of those, 37 quoted a sentence the system could not find in the message. Those were cut.

Readers pushed back on the Hacker News thread. One said the model "kept adding details from the public record that the email didn't contain". Another doubted that people would read AI-written text. The team did not dispute these points. It said the result is a prototype, not a product users can run on their own data yet.

The cost

Jeff from This+that said on Hacker News that the run cost about $11,000 at list prices. That is for 45,047 emails of input, 214 pages of output, and the second-AI audit on top. The company did not name the model or the inference vendor on the page. Nor did it break the cost down by step.

For a team sitting on a pile of emails, tickets or old documents, this is a worked example. It shows what a cited AI wiki can look like, and roughly what it costs. The method is not open source. The lesson is simple. Cite every fact to a source document. Have a second model audit a sample before you trust the output.

Source

We had AI write a wiki from 45k Enron emails, every fact cited, This+that. Show HN discussion at news.ycombinator.com/item?id=50207620.