We had AI write a wiki from 45k Enron emails, every fact cited
B · Useful, limited A vendor prototype on a public corpus that anyone can check, with per-fact citations and a sampled audit, but the audit is done by a second AI and model, cost and build time are not disclosed.
Summary based on the full text·Original: about 6 minutes to read
In short
A 23 October 2026 showcase by this+that, a start-up building a "company brain" product, which fed 45,047 emails from the public Enron corpus to a prototype pipeline. It produced 214 pages on the organisations Enron dealt with, holding 49,401 dated facts, each quoting a sentence from a specific email. When later mail shows a fact stopped being true, the page expires it on that date, and a slider shows what was known on any day. No person edited a page or fact.
Key numbers
- 45,047Enron emails given to the pipeline, from eight public mailboxes
- 214entity pages written about companies, regulators, banks and Enron units
- 49,401facts on those pages, each citing and quoting the email it came from
- 93%sampled facts judged fully supported by their email in a held-out audit by a second AI
- 0 of 500sampled facts judged wrong at their core
As reported by the source.
English·Web page·opens thisandthat.chat
What it reports
- Pipeline steps: find entity types, find entities, build pages, verify. It proposed 57 narrow types and a person kept 54.
- Pages went to organisations named in at least 25 messages (209 pages, from 929 organisations found) and the five most-discussed Enron units (5 pages, from 551 internal entities found).
- Also found but not paged: 78 products and platforms, 76 facilities and market points, 49 events, including noise such as catered restaurants and charity rides.
- Most-discussed entities: FERC (1,501 messages), General Electric (1,499), Pacific Gas and Electric (773), Dynegy (593).
- Audit: 93% of sampled facts fully supported by their email; 0 of 500 sampled facts wrong at their core. Facts whose source sentence could not be found word for word were discarded. Phone numbers and credentials were masked.
- Design view: facts are dated sentences, not fields; structured fields stay in CRMs and systems of record.
How they got it
Automated extraction over eight Enron mailboxes released by FERC and maintained by Carnegie Mellon since 2003, chosen because the history is settled and checkable. Verification used a held-out audit by a second AI; the detailed methodology page was not part of the text read.
Limits
- No model, cost, run time or error analysis given in the overview.
- The audit is model-on-model; no human-checked sample is reported.
- The 25-message cutoff and the type choices were human decisions that shape the output.
- Vendor showcase for a product in development.
Why it matters for companies
This is a concrete pattern for turning correspondence into a maintained, cited knowledge base where every claim traces to a source and expires when superseded. For legal, compliance, account management and onboarding, the citation-per-fact rule is the part to copy, since it makes review possible. The open questions for any real deployment are access control over whose mail is read, and who checks the audit, which here is another model.
Summary based on the full text. Read the original for the full argument and data.Read the original on thisandthat.chat
- Tags
- Tacit knowledge