Get your free personalized podcast brief

We scan new podcasts and send you the top 5 insights daily.

Buying remnant corporate communications, such as internal emails from bankrupt companies like Spirit Airlines, offers real-world data but risks embedding failure modes into AI agents. Models trained on such archives may inadvertently learn the suboptimal operating behaviors and cost-cutting habits that sank those businesses.

Related Insights

A new market has emerged where defunct startups sell their entire operational histories—including codebases, internal communications, and go-to-market data—to AI labs and data brokers. This creates a new form of salvage value, turning years of failed effort into a valuable corpus for training next-generation models.

Unlike humans who debate flawed data in meetings (a "data brawl"), AI agents confidently present a single, incorrect number from bad data. This creates a "silent failure" where the error is persuasive and unnoticed, a phenomenon Salesforce's Gaurav Pathak calls "garbage in, gospel out."

Implementing AI won't fix underlying organizational problems; it will amplify whatever culture, assumptions, and biases already exist. Feeding AI biased historical data, such as past marketing materials or hiring packages, will lead to biased outputs, rapidly scaling pre-existing flaws.

AI models fail in business applications because they lack the specific context of an organization's operations. Siloed data from sales, marketing, and service leads to disconnected and irrelevant AI-driven actions, making agents seem ineffective despite their power. Unified data provides the necessary 'corporate intelligence'.

Mandating employee use of a flawed AI system, like Starbucks' chatbot, is a necessary experiment. Running trials that fail provides the most information-rich data on market tolerance and operational flaws, which is essential for guiding future, successful innovation.

Because Meta is using raw employee computer usage for AI training, its models may learn to replicate common human inefficiencies. This could lead to AI agents that browse social media or watch videos instead of working, mirroring the actual behavior of their human trainers.

A critical learning at LinkedIn was that pointing an AI at an entire company drive for context results in poor performance and hallucinations. The team had to manually curate "golden examples" and specific knowledge bases to train agents effectively, as the AI couldn't discern quality on its own.

Tech giants like Google are now bidding on the data of bankrupt companies, such as Spirit Airlines, not for their physical assets but for their unique, real-world datasets. This data is highly valuable for training AI models, creating a new and unexpected market for corporate information that was previously considered dormant.

Google outbid competitors for Spirit Airlines' internal data—not for customer info, but for mundane emails, Slacks, and meeting transcripts. This signals a new phase in AI training where unstructured corporate communications are highly valued for teaching AI agents how organizations actually function and communicate.

All data inputs for AI are inherently biased (e.g., bullish management, bearish former employees). The most effective approach is not to de-bias the inputs but to use AI to compare and contrast these biased perspectives to form an independent conclusion.

Corporate Real-World Data Carries Hidden Biases of Failed Organizational Behaviors | RiffOn