Building an institutional memory from 45,000 emails: AI extracts 49,000 facts and drops them when they expire

Converted a prototype AI model developed by the This and That team into a massive archive containing 45,000 and 47 email messages for Enron into a living institutional encyclopedia, extracting 49,000 and 401 documented facts distributed across 214 pages covering the companies, banks and regulators that dealt with the collapsed company. The experiment revealed a technical pathway that goes beyond traditional search engines and chatbots, as the algorithms took on reading the complex correspondence, building entity pages and documenting each incident with a direct reference to its original message without any human editorial intervention in phrasing the facts.

The core innovation lies in handling the temporal dimension of knowledge, as the validity of information automatically drops from the real-time record once a later message reveals that it is no longer correct.The team equipped the encyclopedia with an interactive timeline that allows viewing the state of institutional knowledge on any chosen date, addressing the problem of mixing concurrent facts with outdated decisions. The system focused on retrieving the missing space in corporate knowledge management, namely the vital information that changes continuously, cannot be captured by the static fields of customer-relationship-management systems, and remains buried in mailboxes.

Independent audit results carried out by a second AI model on a random sample showed that 93 % of the extracted facts were fully supported by the email text attributed to them, while the rate of material errors was zero across 500 facts examined in detail. In entity identification, the code path proposed 57 distinct classifications, of which the human supervisor retained 54, applying a rule that assigns separate pages to institutions mentioned in at least 25 messages, alongside five main sections within Enron that covered three business units, a trading platform and two technical systems for deal management.

This practical achievement puts an end to organizations’ reliance on superficial text retrieval, offering an organizational alternative based on rigorously audited and accountable temporal records.The shift carries strategic significance for business environments in the Gulf, Egypt and the Levant, where large entities face a chronic documentation gap caused by talent turnover and the migration of institutional knowledge with individuals. Instead of spending budgets on chatbots that hallucinate when tracking contracts and correspondence, this approach opens the way to automating the organizational memory of holding companies, banks and regulators, reducing legal-compliance costs and accelerating internal investigations by turning neglected archives into temporally documented knowledge assets.

This leap requires technology leaders and data managers in the region to revisit knowledge-management strategies, moving from passive email archiving to building live, dynamic entity networks. The pathway consists of achieving disciplined integration between automated extraction capabilities and human framing, with AI monitoring events, linking them to their sources and tracking expiration dates, while the human team defines entity classifications and importance thresholds, ensuring the creation of an accurate institutional memory that protects the organization and supports its executive decisions.