Stop further uploads, identify exactly what was sent and preserve the account and upload records. Then use the provider's verified support or administrative route to restrict access and request appropriate deletion or retention controls. Do not assume deleting the visible conversation removes every stored copy, and do not describe the data as used for model training until the relevant product terms and settings have been checked.
The response depends on the actual service, account tier, upload date and configuration. A consumer account, a managed business workspace and an API arrangement can have different terms. Record the arrangement that existed when the upload happened, rather than relying on a current marketing description of the tool.
What should we do before questioning the employee?
Secure the relevant business account and prevent further sharing. Preserve the file version, prompt, conversation identifier, time, user account and any available audit log. Ask the employee to keep the original material and explain the sequence. Do not direct them to delete the account before the business has preserved the evidence needed to understand the event.
Limit circulation of the incident file. Copying the customer workbook into a large management chat to ask what went wrong creates another disclosure. A restricted inventory of the fields and affected records can support the first assessment while the original is held securely.
What did the employee upload? Who could access it? What does the provider retain or use under the actual service arrangement? Evidence answering one question does not automatically answer the others.
How do we establish the scope?
Compare the upload with its source file. Check hidden sheets, comments, attachments and connected files, as well as visible rows. Identify whether names or contact details were included, whether the data was genuinely anonymised and whether confidential business material was also present. Removing a name may leave an individual identifiable through other fields.
Find out whether the conversation was shared, published, connected to another application or used to generate a downloadable output. Preserve the relevant sharing settings. Do not repeatedly run the material through the tool to see what happens. Further experiments can expand the disclosure and contaminate the record.
What should we ask the AI provider?
Use the provider's official business or privacy channel. Supply enough identifiers to locate the incident without resending all the personal data. Ask about retention, access, downstream processing, deletion options, available logs and the limits of any deletion confirmation. Request preservation of material needed for investigation where appropriate, while distinguishing that request from continued routine use.
Read the terms and data-processing agreement for the actual product. Check account settings, administrator policies and any changes made after the event. A general statement that business data is not used for training may not establish that a personal account used by the employee had the same setting. Equally, an upload does not prove that training occurred. State the verified position.
Which legal duties need assessment today?
As checked on 17 November 2026, the ordinary DPDP processing and breach duties are scheduled for 13 May 2027. The consent-manager provisions that commenced in November 2026 do not bring the whole regime forward. Existing IT Act provisions, the 2011 sensitive personal data rules and contracts need their own assessment on the present facts.
Check whether the upload disclosed sensitive personal data, breached a customer confidentiality undertaking or sent information outside an agreed processing environment. The DPDP Act's future processor-contract and security duties are useful preparation points, but should not be described as already applicable solely because the incident involves AI. A customer's contractual restrictions can be enforceable independently of that future timetable.
Could this require incident reporting?
Assess CERT-In's current directions promptly. Covered entities must report specified cyber incidents within six hours of noticing them or being told about them. The reporting material includes data leaks and breaches. Whether this upload falls within a reportable category requires a factual assessment, including what was exposed and how. Do not assume that an employee's mistake is automatically excluded.
The CERT-In FAQs allow an initial report based on available information followed by additional details. Contractual notification duties and sector requirements may have different triggers. Record the discovery time, assessment, action and acknowledgements. Waiting for a provider to answer every question does not itself suspend an independent reporting clock.
- The original uploaded file, prompt and conversation identifiers.
- The account tier, applicable terms, settings and access logs.
- The customer contract and any approved-tool policy in force at the time.
- The provider's written response on access, retention and deletion.
- The incident timeline and separate reporting or customer-notification assessments.
How should we deal with the employee and customers?
Investigate the facts before deciding whether the event was a mistake, a policy failure or deliberate misuse. Check whether the business had provided an approved tool, clear instructions and workable alternatives. Follow the applicable disciplinary process. Do not ask the employee to write an admission drafted to protect the company or to invent a technical explanation.
Where a customer must be notified, explain the data involved, confirmed facts, containment and remaining uncertainty. Avoid claiming that the matter is closed because the employee clicked delete. Protective advice should match the actual risk, such as compromised credentials or exposure of payment information, rather than a generic alarming message.
What should change after containment?
The useful control is a clear boundary around what may be uploaded, to which approved service and under whose authority. Give staff a practical way to remove unnecessary identifiers and obtain approval for uncertain cases. Review browser extensions and connected applications as well as the main chat interface.
My concern would be whether the business can reconstruct this upload and stop the next one. A policy that forbids everything but leaves staff without an approved route is difficult to apply consistently. Test the revised process using synthetic data, record the lesson and review the incident file's retention. The business data-protection guide covers the wider framework. The personal-data remedies answer and cyber-crime practice page address separate individual and incident routes.