What is the Value of a Failed Company’s Data in the AI Era?

When an airline goes bankrupt, there are some fairly obvious assets to be sold: aircraft, airport slots, equipment, intellectual property and perhaps its brand. But the bankruptcy of Spirit Airlines has highlighted another potentially valuable asset – the accumulated data exhaust of the enterprise itself.

Google recently won an auction to acquire a substantial collection of Spirit’s business data for $10 million, although the transaction remains subject to bankruptcy court approval and has subsequently faced both objections from former employees and a higher offer from AI company Micro1.

Google has said the data could be used to improve its products and AI models. That raises a broader question for the alternative data industry: what is an entire company’s accumulated operational history worth in the AI era?

From corporate data to training data

Much of the data traditionally used to train large language models consists of published output: websites, books, articles, academic papers and publicly available code. An enterprise corpus is fundamentally different.

Emails and messages capture discussions as they happen. Calendars reveal how activity is organised. Project records document implementation. Code embodies technical decisions. Operational and financial information records what happened afterwards.

Taken together, such information potentially captures chains of activity that are difficult to reconstruct from public data: problems being identified, decisions being debated, actions being taken and consequences emerging.

That makes enterprise data potentially valuable not simply for teaching models what businesses say, but for helping them understand how organisations actually behave.

There is, however, an important qualification in Spirit’s case. This is not the accumulated knowledge of a conspicuously successful organisation. Spirit filed for bankruptcy. Any attempt to treat its historical corpus as a blueprint for how businesses should operate would therefore be deeply questionable.

But failure may be part of what makes the data interesting.

Published corporate material inevitably suffers from a degree of survivorship and hindsight bias. Successful strategies are documented, explained and frequently rationalised after the event. A longitudinal dataset from a failed business offers something different: decisions and discussions recorded before participants knew how events would ultimately turn out.

The value is therefore not necessarily in teaching AI how to run an airline. It may lie in exposing models to the messier reality of how organisations operate – including poor decisions, inefficient processes and unsuccessful outcomes.

Conway’s Law meets AI

The inclusion of both communications data and software code adds another intriguing dimension.

Computer scientist Melvin Conway observed in 1967 that organisations tend to design systems that mirror their own communication structures – an idea subsequently known as Conway’s Law.

In simple terms, organisational architecture tends to become embedded in technology architecture.

A sufficiently rich enterprise corpus potentially contains both sides of that relationship. Communications can reveal who interacted with whom, how teams were structured, where responsibilities sat and how decisions moved through an organisation. Source code and technical documentation can show the systems those same people subsequently created and maintained.

In principle, AI could therefore encounter relationships between organisational behaviour and technological outcomes: how organisational boundaries translate into system boundaries, how requirements become software, where projects encounter friction, or how operational problems trigger technical changes.

That does not mean Google’s models will somehow determine why Spirit failed, nor that organisational or technology problems necessarily caused its bankruptcy. Drawing causal conclusions from one company’s data would be highly problematic, and many external factors contributed to Spirit’s difficulties.

The more significant question is what becomes possible when datasets of this kind are available at scale.

A new alternative data asset?

Imagine enterprise corpora covering hundreds or thousands of organisations – successful businesses, mediocre ones and failed ones – containing communications, decisions, software, operational information and eventual outcomes.

AI developers would then have something approaching longitudinal records of organisational behaviour:

organisation ? communication ? decisions ? technology ? operations ? outcomes.

For alternative data practitioners, that represents an interesting inversion of the established model.

Alternative data has largely been about observing companies from the outside, using signals such as web activity, transactions, job postings, satellite imagery or shipping movements to infer what is happening inside them.

Enterprise corpora offer the opposite perspective: historical records generated from inside organisations themselves.

There are substantial obstacles, not least around privacy, intellectual property, anonymisation and the rights of employees whose communications were never created with AI training in mind. Those issues are already emerging in the Spirit bankruptcy proceedings and could ultimately prove as important as the technology.

But the auction nevertheless provides an intriguing indication of where the economics of data may be heading.

Spirit Airlines failed to preserve the value of the enterprise itself. Yet its accumulated communications, software, operational records and institutional history have acquired a second life as an AI asset.

And paradoxically, some of that information may be valuable not despite the fact that the company failed, but because we already know how the story ended.