Research / Data partnerships

Mikra Research10 min read

Who controls a closed company’s records, what the law says about selling them, what survives a wind-down, and how buyers judge if an archive is licensable.

When a company closes, its people leave and its lease ends. Its records do not. Years of support tickets, contracts, lab notebooks, code, and correspondence sit in a storage bucket or a liquidator’s inventory, still owned by someone, still governed by the promises made when they were collected. Courts, regulators, and AI buyers have each developed views about what can be done with them. This post sets out those views, which records tend to survive, and how a careful buyer decides whether an archive can be licensed at all.

Who holds the data after the doors close.

The answer depends on how the company closed.

In a U.S. bankruptcy, filing creates an estate that comprises “all legal or equitable interests of the debtor in property as of the commencement of the case” [14]. Databases, documents, and code are property. In Chapter 11, the debtor usually keeps running the business and sells assets under court supervision; in Chapter 7, a trustee liquidates. Either way, sales are governed by 11 U.S.C. § 363 [4].

Outside bankruptcy, many smaller companies wind down through an assignment for the benefit of creditors, a state-law process that “transfers all of the assets from a debtor to a trust for liquidating and distributing its assets” to an assignee the debtor chooses, without judicial oversight [15]. The assignee then controls the records.

A company that simply dissolves is not gone immediately. Under Delaware law, a dissolved corporation is “continued, for the term of 3 years” to defend suits and “gradually to settle and close their business,” and the Court of Chancery may appoint directors as trustees or outside receivers to take charge of its property [16]. In practice, the directors or a founder often hold the keys during that period, and sometimes long after.

So the holder may be a debtor in possession, a trustee, an assignee, a receiver, or the former directors. The first question for any buyer is which of these it is dealing with, and whether that person has authority to license.

What the law says about selling it.

The modern rules trace to one case. Toysmart.com, a Waltham, Massachusetts toy retailer, had promised customers their personal information would “never be shared with third parties,” then tried to sell its customer list in bankruptcy. The FTC’s July 21, 2000 settlement allowed a sale only to a “Qualified Buyer” in a related market that bought the whole website, agreed to honor the original privacy statement, and obtained opt-in consent before changing it; if no such buyer appeared within a year, the data had to be destroyed [2]. None did. In February 2001 a Disney subsidiary paid $50,000 to have the list of roughly 250,000 customers destroyed, with an affidavit to the court confirming it [3].

Congress wrote that outcome into the Bankruptcy Code in 2005. Where a debtor had a privacy policy prohibiting transfer of personally identifiable information to unaffiliated persons, a trustee may sell that information only if the sale “is consistent with such policy,” or “after appointment of a consumer privacy ombudsman in accordance with section 332, and after notice and a hearing, the court approves such sale” [4]. The Harvard Law Review noted in March 2025 that the mechanism has limits: roughly 33 ombudsmen have been appointed since the law passed, their reports are non-binding, and they cannot police compliance after the sale closes [5].

The FTC restated those conditions in the RadioShack bankruptcy. On May 18, 2015, it told the court that information on “tens of millions of consumers” was among the assets for sale, that RadioShack had promised not to sell it, and that any buyer should be in substantially the same line of business, take the data bundled with other assets, honor the existing policy, and obtain affirmative consent before any materially different use [1]. The same year, Caesars Entertainment attributed zero equity value to its Total Rewards loyalty database in bankruptcy while dissident creditors argued it was worth more than $1 billion [12]; the program held roughly 45 million members and 17 years of data [13].

The 23andMe case tested all of this against genetic data. The company filed for Chapter 11 in March 2025. Twenty-eight state attorneys general argued that 23andMe “lacked legal authority to treat biological samples and genotype data as freely transferrable assets” and that explicit consent was required [6]. The court-appointed ombudsman, Professor Neil Richards, reported on June 11, 2025 that the privacy statement had been updated 22 times since 2007 and did not mention transfer in bankruptcy before June 2022, that many inactive customers never saw later changes, and that a transfer could run afoul of state genetic privacy laws absent explicit consent; he recommended affirmative consent and deletion for customers who declined [7]. On June 27, 2025, the court approved a $305 million sale to TTAM Research Institute, a nonprofit formed by the company’s co-founder, on terms negotiated with the attorneys general: no transfer of consumer DNA, data kept under the same privacy policies, and deletion rights that state regulators could verify [8]. The breach that preceded the bankruptcy was settled separately in July 2026 with 42 attorneys general for $150 million, of which about $18 million was actually distributed because of the bankruptcy [9].

The pattern is repeating. Character.AI filed for Chapter 11 in October 2026 with a $20 million stalking-horse bid from Disney for a nonexclusive perpetual license to some of its intellectual property, and expects to close a sale of its assets by late November or early December [10]. On October 15, 2026, California and a coalition of other states filed notice citing the “significant volume of sensitive consumer information” the company holds, including private conversations with minors, and reserved the right to challenge any sale [11].

Two constraints sit outside bankruptcy. Under the California Consumer Privacy Act, a transfer of personal information as an asset in a merger, acquisition, or bankruptcy is generally not a “sale,” but only if the recipient does not use the data in a way materially inconsistent with the notice given at collection [17]. Under the GDPR, personal data may be kept in identifiable form “for no longer than is necessary for the purposes for which the personal data are processed” [18], and any new purpose must pass a compatibility test that weighs the link between purposes, the context of collection, the sensitivity of the data, the consequences for the individual, and safeguards “such as encryption or pseudonymisation” [19].

The thread across twenty-six years: the promise made at collection travels with the data. A buyer inherits it.

What usually survives, and what it’s useful for.

Not everything a company kept is a liability, and not everything is a dataset. Most archives contain a mix.

Record type Typical holder AI use Rights question
Customer-support transcripts and tickets Debtor, assignee, or SaaS vendor account Dialogue and resolution data; evaluation sets Privacy policy at collection; customer PII; de-identification required [4][18]
CRM and sales records Trustee or assignee Rarely useful; mostly PII Explicit “no sale” promises bind the buyer [1][2]
Contracts, policies, and internal documents Directors during wind-down; counsel Domain text; document understanding Confidentiality clauses; counterparties’ information
Source code and engineering history Assignee or acquirer of IP Code and review data Open-source licenses; employee and contractor IP assignments
Research, lab, and clinical records Receiver or sponsor Scientific reasoning; structured experiments IRB consent scope; HIPAA de-identification [20]
Editorial, photo, and publication archives Rights holder or successor Text and image corpora Author and contributor contracts [26]
Financial and tax records Officers, as required by statute Little direct use Retention periods; not licensable in identifiable form [21]

The licensing market for archives is real and public. The Associated Press spent most of 2025 making tens of millions of assets spanning five decades machine-readable, rights-cleared, and available by API, after licensing content to OpenAI in 2023, Google in early 2025, and Microsoft’s content marketplace in November 2025 [24][25]. Wiley said in 2024 that it expected $44 million from AI licensing of its books, with no specific opt-out for authors [26]. Shutterstock reported $104 million in revenue from licensing its library for AI training in 2023 [27]. These are living companies, but the asset in each case is the back catalog, which is exactly what a closed company still has.

Whether that asset ever appears on a balance sheet is another matter. Doug Laney, who coined “infonomics,” has argued that most firms hold “a latent asset that isn’t being measured or managed well,” and that accounting standards bodies “have not yet come round to allowing for information to be categorised as an asset on the balance sheet” [23]. That is one reason archives are overlooked in a wind-down: they were never carried as a number.

How a responsible buyer evaluates an archive.

These checks decide whether an archive is licensable, in the order a buyer should run them.

Provenance. Who collected the data, from whom, and under what terms. The buyer reads the privacy policy in force at collection, not the current one. The 23andMe ombudsman’s finding turned on what customers before June 2022 had been told [7].

Authority. Whether the person offering the archive can bind it. A Chapter 7 trustee can; a former executive who kept a backup usually cannot. For a dissolved corporation, the buyer confirms the wind-down period and who was made responsible [16]. For an assignee, the buyer reads the assignment [15].

Consent scope. Whether the intended use fits the purpose for which the data was collected. Under the GDPR that is the Article 6(4) test [19]. Under the FTC’s long-standing position, any materially different use needs affirmative consent [1]. For most consumer records, training a model was not within the original purpose.

De-identification. Whether identifying fields can be removed so that what remains is no longer personal data. The HIPAA standard is the clearest reference: information is de-identified when there is “no reasonable basis to believe that the information can be used to identify an individual,” shown either by expert determination that the risk is very small or by removing eighteen categories of identifiers, including names, dates finer than a year, contact details, account numbers, device identifiers, and biometrics [20]. The GDPR lists pseudonymisation among the safeguards that help a new purpose pass [19]. For most archives, this is the step that converts a liability into a dataset.

Retention conflicts. Whether the holder must keep records the buyer would want deleted, or destroy records the buyer would want to keep. The IRS asks businesses to keep most records three years, employment tax records four, and some indefinitely [21]. The FTC Disposal Rule requires that consumer-report information be destroyed so it “cannot be read or reconstructed” [22]. Litigation holds override both.

Chain of custody. Whether the archive the buyer receives is the one that was collected. Backups copied by departing staff, exports from a lapsed vendor account, and data assembled by a third party are each weaker than records transferred by the lawful holder with a documented inventory. This is why Toysmart’s list was held by counsel until the court ordered its destruction, with an affidavit to prove it [3].

If you hold one.

If you are a founder, a director in a wind-down, a liquidator, or a trustee, a few practical steps.

Inventory what exists before anything is deleted: systems, date ranges, record counts, and where each lives, including vendor accounts that will lapse. Pull the privacy policy and terms of service in force for each period. Find the employee and contractor IP assignments and the open-source manifests for any code. Note what is under a retention obligation or a litigation hold.

Then separate personal data from the rest. Operational documents, engineering history, research protocols, and editorial content are often licensable with ordinary contractual care. Customer records are licensable only after a rights review and de-identification, and sometimes not at all.

Mikra licenses archives from closed companies as part of our data partnerships. Our own practice: we work with companies that have closed or are winding down and had 30 or more office-based employees in the United States or Canada at their peak. The owner uploads the records to Mikra, we de-identify them before any customer receives anything, and we delete the raw upload within 120 days. The owner keeps ownership and is paid a share of the revenue the data earns, on terms set in the agreement, in writing, before any export. There is no upfront payment and no guaranteed amount: we pay only from what customers pay us. Patient records and a law firm’s client files move only through a separate track, where an independent provider de-identifies them before we receive anything. The inventory above is what we ask for first.

Caveats.

The law here is unsettled. The consumer privacy ombudsman mechanism is non-binding and rarely invoked [5]. The 23andMe sale was approved on negotiated terms, not on a ruling that the states’ consent argument was correct [6][8]. The Character.AI case is pending as of this writing [10][11]. State privacy statutes differ, and the GDPR follows European residents’ data wherever the company was incorporated. The Caesars valuation came from creditors in a contested proceeding, not an arm’s-length sale [12]. The publisher figures are for operating companies and do not establish a price for a closed company’s archive.

This is not legal advice.

Hold data like this?

See if your
data fits.

Tell us what your company holds and a program manager will reply with an assessment and a call.