When an AI provider shuts down, gets acquired or becomes insolvent, the information entrusted to it does not necessarily disappear. Whether that information can be transferred depends on applicable law, contractual terms, processing purposes and the legal proceedings involved. For enterprises, the implication is straightforward: AI vendor due diligence must cover the day you stop using the service, not just the day you deploy it.
The report supplied about Character.AI alleges bankruptcy and the possible sale of user conversation data. No court documents or primary-source link accompanied it here, so those details should not be treated as verified facts. But the underlying risk does not depend on this particular case: information accumulated by a provider can remain there long after the business relationship ends.
Warning: Bankruptcy does not erase privacy obligations. An asset-sale clause is not blanket permission to use information for any purpose. Applicable law, contractual commitments and decisions in the specific proceedings still need to be examined.
That distinction matters. The claim that every privacy protection disappears when a company collapses is not a sound basis for business decisions. Equally, a marketing promise that your data is safe is no substitute for legal and technical scrutiny.
A conversation is more than text. It is a snapshot of the business
An AI conversation can reveal much more than an uploaded document. It might expose a manager's thinking before a dismissal, a weak point in a customer negotiation or the real reason a deal was rejected.
Systems with personalities and ongoing conversations can create a sense of intimacy that encourages disclosure. Inside an enterprise, this happens even without a virtual character. An employee who receives useful help over time starts treating the system like a colleague. They add names, circumstances and background details to improve the response, sometimes disclosing information about people who have never used the service.
The question, then, is not simply whether the provider uses content to train models. You also need to establish what is retained in conversation histories, activity logs, backups and support systems, and who can access each of them.
A policy against training on enterprise data matters. It is not a complete data management policy.
What should you check before signing with an AI provider?
Asking whether the product is secure is not enough. Security, privacy and business continuity are separate questions. A system can be well protected against intrusion while retaining information for too long or permitting contractual transfers that do not meet your organization's requirements.
- Data use: What purposes is the content processed for? Review the agreement and the terms of the subscription tier.
- Retention: How long is each type of information retained? Request retention policies for each component.
- Change of control: What happens during an acquisition or liquidation? Review transfer provisions and notification requirements.
- Deletion: What is deleted, and what remains? Check the scope, exceptions and timelines.
- Exit: How will operations continue without the service? Review export capabilities and a transition plan.
The answers must apply to the subscription tier and configuration you actually purchase. A personal account, an enterprise workspace and API access do not necessarily operate under the same terms. A familiar brand or preferred-vendor designation does not remove the need for review.
Enterprises need to assess applicable privacy laws, data security requirements and rules governing international data transfers. Sector-specific obligations and laws in other jurisdictions may also apply. Choosing a service marketed as GDPR-compliant does not complete that assessment.
AI agents need a data map, not just a tool inventory
A chat tool generally receives what an employee enters. An agent may have permission to read email, retrieve customer data and write to operational systems. When it connects to several systems, information may be retained in several places: by the model provider, in the orchestration platform, in monitoring tools and in the enterprise database.
This is not an argument against agents. It is an argument for building the internal capability to manage them. Operational value increases when an agent gets the information needed to complete its task, rather than unrestricted access to everything that can be connected.
- Map the data flow: Document what information leaves each system, which provider receives it and where it is stored.
- Minimize before sending: Remove identifying details and limit context to what the task requires.
- Define access and retention: Set role-based access and retention periods aligned with business needs and legal obligations.
- Build an exit path: Plan for export, deletion, permission revocation and provider replacement without bringing the process to a halt.
- Test and verify: Confirm that exports are usable, connections have been disabled and deletion has been completed to the promised extent.
The sequence is deliberate. First understand the business process and the sensitivity of its information. Then choose the technical settings. Deep AI expertise, operational experience and legal understanding need to meet in the same discussion. A successful agent demo is not evidence of a dependable enterprise foundation.
Keep humans in the loop, but not on every message
Routing every interaction for manual approval is not the answer. That would eliminate much of the efficiency the agent was introduced to deliver.
Instead, define a policy that distinguishes routine activity from exceptions. Processing limited, approved information can happen automatically. An attempt to send sensitive information to a new destination, expand permissions or make a material change to provider terms should trigger a pause and human review.
Technical controls also need to be supported by employee literacy. Employees must know not only how to write an effective prompt, but also which information must never be included. These are complementary efforts: responsible use of AI tools and the deployment of managed agents.
Deleting an account is an action, not proof that the data is gone
Deleting an account may be the right step when you stop using a service. But do not assume it immediately removes every copy. Backups, legal retention duties and other exceptions may apply. Nor can you assume what deletion means for information already used in training.
Tip: Check the full deletion chain. Ask what happens to conversations, files, activity logs, backups and copies held by subprocessors. A general answer about account deletion may not cover all of them.
Before deleting anything, also check whether your organization must preserve particular records. The goal is not indiscriminate deletion. It is to stop unnecessary retention without compromising recordkeeping obligations or business continuity.
The question leadership should ask now
At your next AI review, ask for an answer to one question: If our main provider stops operating tomorrow, do we know where our information is, who is entitled to hold it and how we will keep working?
If the answer depends on one employee's memory or a salesperson's assurance, a layer of governance is missing. Responsibility should be shared by IT, information security, legal and the owner of the business process.
You do not need to bet on the collapse of the AI industry to justify this investment. Providers are acquired, products close and terms change even in healthy markets. An enterprise that manages its information and vendor dependencies properly can adopt AI with greater confidence, capturing value without surrendering unnecessary control over its assets.