When your AI supplier fails, what happens to your conversations?

The BBC’s account of Character.ai’s bankruptcy raises an uncomfortable question: what happens to the private conversations held by an AI company when that company needs a buyer? For organisational leaders, the equivalent concern is readily recognisable. Replace intimate personal disclosures with acquisition discussions, staff grievances or negotiating limits, and the issue moves directly into the procurement meeting. ¹

On 12 October, Character.ai announced a Chapter 11 process to pursue a court-supervised sale. It said the service would continue during the process, but could close if no suitable bid emerged. California’s attorney general subsequently announced action seeking privacy and safety protections in any transfer. These developments establish a live concern about future custody; they do not establish that users’ conversations have already been sold. ²³

A companion-chatbot service is not a proxy for a contracted enterprise AI provider. Their terms, purposes and safeguards can differ substantially. The case nevertheless exposes a question that belongs in both settings: how far does your control over retained information survive the commercial relationship?

The archive behind the answer

Consider a hypothetical professional-services firm using an AI assistant to prepare for client negotiations. Staff upload correspondence, explore concessions and discuss why particular accounts might leave. The final briefing is concise. The conversation that produced it is considerably more revealing.

If the service retains those exchanges, it may hold the firm’s reasoning alongside its documents. Rejected proposals, uncertainty and candid assessments can be more sensitive than the polished output. An archive assembled through ordinary use could become a record that nobody consciously decided to create.

The same mechanism could affect a charity discussing donors, a public body considering service reductions or a manager preparing a difficult personnel conversation. People mentioned in those exchanges may never have used the tool themselves.

This is an exposure scenario, not evidence that all AI suppliers retain such an archive or can sell it. Its relevance depends on the particular product, settings, agreement and information involved. That is precisely why the question needs an answer before sensitive use becomes routine.

A sale is not a breach

Security reviews commonly ask how information is protected from unauthorised access. A change of ownership introduces another issue: whether information can pass to an authorised successor, and what that successor may do with it.

The distinction matters. Strong security does not, by itself, answer a question about a permitted transfer. Equally, permission to transfer information does not necessarily permit the recipient to repurpose it.

This problem predates generative AI. In RadioShack’s 2015 bankruptcy, the US Federal Trade Commission sought conditions on the proposed transfer of customer information. A subsequent multistate settlement restricted the data that could pass to the buyer and provided for much of the archive to be destroyed. That history shows both the pressure to realise value from retained data and the possibility of meaningful restrictions. ⁴⁵

Nor should readers assume that bankruptcy cancels privacy law. UK Information Commissioner guidance explicitly addresses transfers during insolvency and requires attention to lawful basis, original purposes and transparency. The guidance is currently under review following legislative changes. The treatment of any particular transaction requires current, jurisdiction-specific assessment. ⁶

Personal-data obligations also do not answer every confidentiality question. An organisation’s negotiating position may be commercially sensitive without being personal information. Procurement needs to examine the relevant confidentiality and contractual arrangements as well.

Two different exit problems

Return to the professional-services firm. Its operations director asks whether conversation histories can be exported. The supplier says yes. That is useful for continuity, but it leaves a second question unanswered: what remains with the supplier after the export?

The firm needs to distinguish its ability to recover information from the supplier’s obligations to delete or restrict retained copies. It should establish which records exist, which entities hold them, what retention exceptions apply, and what changes if the service is sold or closed.

A promise about model training is also narrower than a complete answer about retention. The review should ask separately about storage, permitted uses, onward transfers and deletion. It should cover the actual service arrangement, including relevant downstream providers, rather than stop at the brand on the screen.

None of this requires predicting which supplier will fail. An acquisition, product withdrawal or ordinary contract termination can raise related questions. An industry-wide collapse is unnecessary for the exposure to matter.

Decide while you still have choices

An approved enterprise service may have a persuasive answer to each of these questions. A lightly reviewed tool adopted by individual staff may not. Treating them as equivalent would obscure the very differences procurement should establish.

The useful response to Character.ai is to examine those differences while the organisation still has time and bargaining power. Ask what happens if the relationship ends, and then ask what happens to the information that remains.

Replacing an AI service can restore a workflow. It does not, by itself, settle the future of the conversations that helped run it.