Google Has More Data Than Almost Anyone. So Why Is It Bidding $10 Million On Old Emails?

Summary

Google’s reported $10 million bid for Spirit Airlines’ work records reflects AI companies’ belief that workplace data can train agents to perform business tasks. The speaker’s central caution is that records of work are not the same as the work itself. Emails, chats, documents, and meeting transcripts may show activity without capturing the judgment, coordination, company context, and outcomes that created value.

Business implications

• Work is often hard to verify. Software that runs is relatively easy to assess; the quality of a document, decision, or customer resolution is harder to prove. That makes it difficult to define what an agent should learn and what “the job is done” means.
• Activity traces can misrepresent performance. In an accounts-payable example, dozens of messages and several meetings might lead to an invoice resolution. The valuable contribution could be one person spotting a wrongly applied credit—not the entire message trail. A ticket may also appear successfully closed, only for a duplicate payment to emerge weeks later.
• Company context matters. Employees rely on tacit knowledge about customers, exceptions, systems, and approval authority. A model trained on generic records may imitate visible processes while missing the reasoning needed to resolve unusual cases.
• Data purchases may have selection bias. The speaker argues that companies most motivated to sell archives may be financially distressed or out of other options. Spirit’s records therefore may not provide a clean picture of successful, repeatable operations.
• Automation can create resistance if it is framed as replacement. The speaker says leaders have privately estimated that as many as one-third of their teams may quietly resist or sabotage AI rollouts. This is presented as an anecdotal estimate, not a measured industry-wide KPI.
• The business case should be tied to outcomes. The speaker frames employees’ contributions in terms of increasing revenue, reducing costs, or otherwise improving real business value—not the volume of messages or documents produced.

Deals, players, and metrics mentioned

• Spirit Airlines archive: Google reportedly won an August auction with a $10 million bid, beating Mercor’s $7.5 million offer. The records reportedly include 100 million emails, 500 million Teams entries, code, and business documents. The proposed sale was not yet complete; a hearing was scheduled for October 14.
• Mercor and Deep Tune: Mercor was described as planning to acquire Deep Tune, a learning-platform company. The combination aims to pair the software agents operate in with ways to check whether they completed tasks successfully.
• Mercor’s evaluation approach: Mercor says experts define tasks and success criteria, and it has posted paid work to set up agents and assess whether they can do the job. The key challenge is deciding what counts as success—for example, whether closing an accounting ticket is enough.
• Grepped.ai: The speaker describes this emerging data broker as packaging small-business data for resale to AI labs, with prices ranging from $20,000 to $5 million.
• No direct financial impact metrics—such as revenue lift, cost savings, or productivity gains from these agents—are provided in the subtitles.

Frameworks and actionable playbooks

Define work by outcomes, not artifacts

• Specify the business result a task should achieve.
• Identify the decisions, judgment, and context behind that result.
• Distinguish outcome-producing work from coordination or documentation that merely accompanies it.

Evaluate agents on representative tasks

• Select typical real-world cases, including exceptions and messy handoffs.
• State what a correct result depends on and how it will be verified.
• Check for downstream failures, not just apparent completion—for example, whether a later duplicate payment exposes a bad resolution.

Assess a data sale before proceeding

• Clarify the proposed scope and use of the records.
• Review whether confidential or personal employee information is included or inferable, including through links between systems.
• Ask whether the archive is genuinely useful for improving the company’s own performance.
• Recognize and compensate employees’ expertise if they are needed to turn raw records into training materials or success criteria.
• Validate that the resulting agent works before selling or relying on the data; otherwise, the archive may lose value without producing useful capability.

Deploy AI against suitable workflows

• Start with repetitive, relatively boring work that can be handled by a well-defined system.
• Avoid assuming that an agent must reproduce human Slack, email, and meeting patterns; redesign the workflow where possible so the agent reads inputs, processes them, and returns a defined output.
• Keep humans involved where judgment and company-specific context are central.

Build a reliable source of truth

• Invest in structured, high-quality documentation—such as usable Markdown files—that both employees and agents can consult.
• Give teams time to prepare clear assessments and reference material.
• Use QA teams to evaluate what agents produce. The speaker presents this foundational information work as more valuable than simply feeding agents a large archive of unstructured records.

Presenter and sources mentioned: Presenter Nate B. Jones. Organizations and examples discussed include Google, Spirit Airlines, Mercor, Deep Tune, Grepped.ai, Glean, and the Spirit Airlines flight attendant union.