AI interest in bankrupt companies' data: what courts check before a lot sells

Why AI developers look at bankruptcy archives#

AI developers look at bankruptcy archives because a failed operating company can leave behind years of complete, connected records of real work: support conversations, order exceptions, engineering histories and internal decisions. Developers building agents for business tasks need examples of how that work unfolds, and a closed company's archive is complete and no longer changing.

For an estate, that interest is a possible source of recovery from records once valued mainly for the customer list, if at all. For a court, it raises familiar questions in a new setting: whether the estate can sell or license the records, whether the process was fair, and whether people whose information sits in the records are protected.

Interest is not universal. Archives built mostly on consumer personal data face more obstacles than business records of support, operations or engineering work, and an archive with broken links or unclear rights may draw no data bids at all.

Preserve the records before the case moves on#

Preserving the records comes before any sale, because debtors often stop paying for SaaS tools well before the petition is filed. By the time a trustee is appointed, help desk, CRM and chat accounts may already be suspended, with deletion dates set by the vendor's terms.

A trustee's first data steps are practical: identify every system, secure admin access, ask vendors for read-only access or a paid extension where needed, and export and verify records before accounts close. Former employees or contractors who know the systems are often the fastest route to a usable export.

Records preserved this way also serve the case itself, from claims review to litigation, so preservation is rarely wasted even when no data lot is ever sold.

What courts typically check before a data lot sells#

Courts typically check the same things for a data lot that they check for any sale of estate property, with privacy added. The table maps each check to the evidence a trustee usually brings to the hearing.

None of these checks depends on who the bidder is. A court asks the same questions whether the winning party is a competitor, a liquidator or an AI developer.

Privacy policies and the consumer privacy ombudsman#

Privacy policies matter most when a debtor promised customers it would not transfer their personal information. US bankruptcy law has a specific process for that situation: where the policy in force prohibited the transfer, a sale of that information may need court approval after a hearing, and a consumer privacy ombudsman, appointed at the court's direction, can report on the privacy issues before the court rules.

That process is aimed mainly at consumer information, but business-to-business companies are not exempt from privacy questions. Workplace chat, support conversations and CRM histories carry the names and contact details of employees and customer contacts. Removing those details before delivery, and showing the court how that was done, narrows the privacy issue to something the court can evaluate.

The privacy laws that may apply turn on the records and the people in them, and estate counsel assesses that case by case.

Selling the archive versus licensing it#

Selling an archive transfers ownership once; licensing it grants use rights while the estate keeps ownership. A non-exclusive license can allow more than one licensee for the same records, and different record families can go to different parties.

Licensing has its own trade-off in bankruptcy. Trustees want to close cases, and a license with ongoing duties for the estate, such as future deliveries, audits or support, complicates that. Licenses built for an estate tend to be paid at signing or delivery, end the estate's obligations at delivery, and place deletion or return duties on the licensee.

When an operating buyer is also purchasing the business, the sale order should say whether historical records go with the business, stay with the estate or are licensed to the buyer, so the two transactions do not collide.

Build the lot around the records, not the bidder#

A data lot is best described by its records, rights and exclusions rather than by the kind of buyer the trustee expects. Bidder-centred descriptions invite speculation; record-centred descriptions give every bidder and the court the same facts.

Published metadata standards help. The Data & Trust Alliance's Data Provenance Standards organize dataset metadata into Source, Provenance and Use groups, and the Use group includes elements for consent documentation, license to use and intended data use. A lot description that answers those headings is easier for bidders and the court to evaluate.

• 1Systems and record families, with the span of history each covers.
• 2How records link to one another, such as tickets to orders or issues to releases.
• 3What was excluded and why: privileged material, HR records, customer-owned content.
• 4How personal information will be removed before delivery.
• 5Known contractual restrictions and how they were handled.
• 6Delivery method, such as access in the estate's storage or encrypted drives.

Illustrative: a Chapter 7 trustee packages a freight brokerage's records#

Illustrative: a fictional regional freight brokerage files Chapter 7 after a major shipper leaves. The trustee finds load and carrier history in McLeod, telematics data from Samsara, damage claim files in a shared drive, and dispatch mailboxes where exceptions such as missed pickups and refused deliveries were worked out. The lender's lien covers general intangibles.

The trustee agrees with the lender on how license proceeds will be treated, carves out rate sheets for shippers whose contracts restrict their use, and excludes telematics location history because it tracks individual drivers. Driver and contact names, phone numbers and emails are removed from load notes, claims and dispatch email. The website privacy policy covered only marketing contacts, and those are excluded too.

The trustee moves for approval of a non-exclusive license covering de-identified load exception and claims records, and serves the motion on creditors and parties in the case. The motion draws no objections; the court enters an order approving the license, and the estate receives payment at delivery.

How SourceX supports trustees with a data lot#

SourceX supports trustees by treating each data lot as its own SourceX five-step transaction: Supply, Rights, Preparation, Approval and Delivery. The first stage works from metadata only, so a trustee can get an early read on whether a lot is likely to interest AI developers before spending estate funds on exports or preparation.

The output that matters most in court is the SourceX Evidence Packet. Its five headings, provenance, licensing rights, permitted use, privacy record and release authorization, line up with the authority, third-party rights, use and privacy checks in the table above, which gives estate counsel a practical starting point for the record-related parts of a motion. Large archives stay in the estate's storage or ship on encrypted drives rather than being hosted by SourceX.