AI Use Policy for Indian Businesses: Stop Staff Leaking Data into ChatGPT

Last updated: 29 October 2026. You’ll find the sources at the end.

On 23 October 2026 Panasonic Holdings apologised for a leak that involved no hacker at all. People working on an HR system project at Panasonic Operational Excellence and its subsidiary Panasonic Digital uploaded part of the company’s HR data to an external generative-AI service that the company had banned for staff. The data covered about 334,000 people: current and former employees, plus family members listed as emergency contacts. It is the clearest case yet for a written AI use policy.

However, most Indian companies we talk to have no written rule about this. Staff use ChatGPT, Gemini or Copilot every day, usually on personal free accounts, and nobody has told them what they may paste in. So this guide covers what went wrong at Panasonic, how the main AI tools treat your data, what Indian law expects, and a one-page AI use policy you can adapt today.

What happened at Panasonic

The facts below come from Panasonic’s own notice of 23 October 2026.

ItemDetail
WhoProject members at Panasonic Operational Excellence, working with Panasonic Digital
WhatPart of the HR system data (as of April 2026) uploaded to an external generative-AI service the company prohibits. The service is not named
PeopleAbout 75,000 current employees, 178,000 former employees, 55,000 employees’ family members and 26,000 former employees’ family members
DataEmployee number, address, phone and email (no names) for staff; name, address, phone and email for family contacts
TrainingThe setting that lets the service use data for AI training was switched off
Clean-upThe employee deleted the chat history straight after discovery. The vendor keeps deleted data for 30 days, so it held the data until 25 October
ReportedTo Japan’s Personal Information Protection Commission on 29 September 2026

Even though the service’s training setting was off, Panasonic still had to report and apologise. After all, once data on 334,000 people sat on a third-party service the company had no contract with, the company had lost control of it. Consequently, Panasonic’s fixes are a review of its rules on personal data and generative AI, more staff training, and continued monitoring.

Panasonic is not the first. In April 2023 Samsung engineers uploaded internal source code and meeting recordings to ChatGPT, and in May 2023 Samsung banned generative-AI tools on company devices and networks. Similarly, Apple, JPMorgan Chase, Verizon and several banks restricted ChatGPT the same year for the same reason. In January 2026 Politico reported that the acting head of CISA, the US cyber agency, had uploaded “for official use only” contract documents to a public version of ChatGPT.

Free accounts and business accounts treat your data differently

The biggest single risk is staff using free consumer accounts for work. On most consumer plans, the vendor can use your chats to train its models unless the user turns that off. In contrast, business plans come with a contract and, for these vendors, no training on your data by default.

ToolConsumer plan (free or paid personal)Business plan
ChatGPTOpenAI may use content for training unless you opt out (“Improve the model for everyone” in Data Controls). OpenAI does not train on Temporary Chats and deletes them after 30 daysBusiness, Enterprise and Edu: no training on your organisation’s data by default
Google GeminiWith “Keep Activity” on, chats help improve Google services, including training. Human reviewers read some chats, and Google keeps those for up to three years. Google asks users not to enter confidential informationGemini in a qualifying Google Workspace edition: no training without your permission, and no human review of chats
Microsoft CopilotConsumer Copilot uses conversations for training by default unless the user opts out. Microsoft’s opt-out list of countries leaves India outCopilot with enterprise data protection (signed in with a work account): prompts and responses are not used to train foundation models
ClaudeFree, Pro and Max users choose whether chats are used for training. Anthropic keeps data 5 years if you allow training, 30 days if notClaude for Work (Team, Enterprise), API: not used for training
DeepSeekThe privacy policy (February 2026) says personal data is stored in the People’s Republic of China and inputs may be used for training; opt-out is by email requestNo separate business plan covered here

Turning off training is not enough

Also, these settings change often. We checked each one on the vendor’s own help pages on 29 October 2026; check again before you rely on them.

Turning off training doesn’t make a tool safe for confidential data, as Panasonic shows. It only stops the vendor using your data to improve the model. The vendor still stores it, its staff may still review it, and you have no contract that covers it.

What Indian law and guidance say

There is no Indian law that bans staff from using AI tools. Several existing rules still apply.

The DPDP Act 2023 makes your company, as the Data Fiduciary, responsible for personal data processed by you or on your behalf (section 8(1)), and you may use a Data Processor only under a valid contract (section 8(2)). For example, an employee pasting customer or staff data into a personal ChatGPT account sends that data to a processor you have no contract with. In addition, section 8(5) requires reasonable security safeguards to prevent a breach, and the Schedule sets penalties of up to ₹250 crore for failing to take them and up to ₹200 crore for failing to report a breach.

The DPDP Rules 2025 were notified on 13 November 2025. Rule 6 (security safeguards) and Rule 7 (breach intimation) come into force 18 months later, around 13 May 2027. Business Standard reported a proposal to shorten that runway in January 2026, but the government has not notified it. You have time, but a policy written now is cheaper than one written after an incident.

CERT-In and government guidance

CERT-In’s advisory CIAD-2025-0013 (March 2025) on generative AI tells users to “only use AI tools approved and verified by your organization” and not to share anything they would not be comfortable seeing made public. Then a later advisory, CIAD-2026-0020 (April 2026), suggests restricting outbound traffic to AI services so staff cannot use unsanctioned tools, and tells MSMEs to avoid unverified AI tools in production.

The government has also applied the same thinking to itself. Reuters reported in February 2025 that a Finance Ministry note told employees to avoid AI tools such as ChatGPT and DeepSeek on office computers and devices. Meanwhile, MeitY’s AI Governance Guidelines (November 2025) prefer existing laws over a new AI law and ask organisations to adopt privacy and security measures proportionate to risk.

If you follow ISO 27001:2022, an AI use policy fits under Annex A control 5.10 (acceptable use of information and other associated assets) and 5.23 (information security for use of cloud services).

A one-page AI use policy you can adapt

Keep your AI use policy short enough that people read it. Replace the bracketed parts.

[Company name] AI use policy (version 1.0, [date], owner: [name, role])

  1. Approved tools. Staff may use only these AI tools for work: [e.g. ChatGPT Business through the company workspace, Microsoft Copilot signed in with your work account]. Do not use personal or free AI accounts for company work.
  2. Never enter into any AI tool, approved or not: customer or employee personal data (names with contact details, Aadhaar, PAN, bank or card details, health data, salary), passwords, API keys or other secrets, source code from [repositories], contracts, financial results before release, or anything marked confidential.
  3. You may enter public information, your own drafts with personal and confidential details removed, and generic technical questions.
  4. Check the output. Because AI answers can be wrong, you are responsible for anything you send, publish or deploy, including code.
  5. Agents and integrations (tools that connect to email, drive, code repositories or cloud accounts) need written approval from [IT/security] before setup, with the narrowest access possible.
  6. Mistakes. If you paste something you shouldn’t have, tell [contact, email] within [1 hour]. Do not just delete the chat, because we need to assess it, and we may have to report some incidents to CERT-In within 6 hours.
  7. Review. [Owner] reviews this policy every [6 months] and when we adopt a new tool.

Our free security policy generator can produce a full acceptable-use policy around this, with your company name and contacts filled in.

Back your AI use policy with controls

However, an AI use policy alone relies on people remembering it. Add at least one technical control.

  • Give people an approved option. If staff need AI to do their job, a business account with a contract is the control that removes most of the risk. Blocking everything, on the other hand, usually pushes people to their phones.
  • See who is using what. If your office traffic goes through a proxy, its logs show which machines reach AI sites. For a Squid access log (tested on a sample log, Ubuntu 24.04):
grep -E 'chatgpt\.com|openai\.com|gemini\.google\.com|claude\.ai|deepseek\.com|copilot\.microsoft\.com|perplexity\.ai' /var/log/squid/access.log \
  | awk '{print $3, $7}' | sed 's/:443//' | sort | uniq -c | sort -rn

Sample output:

      2 192.168.1.21 chatgpt.com
      1 192.168.1.40 chat.deepseek.com
      1 192.168.1.34 gemini.google.com
      1 192.168.1.34 claude.ai
  • Block the tools you have not approved. In Microsoft Defender for Cloud Apps you can filter the app catalogue to Generative AI and mark apps as Unsanctioned, which blocks them on devices onboarded to Defender for Endpoint. Chrome Enterprise Premium (a paid add-on) can block, warn or log visits to its “Generative AI” category. However, these steps come from Microsoft’s and Google’s documentation, and we have not tested them in our lab.
  • Stop sensitive data at the browser. Microsoft Purview DLP can block sensitive text and file uploads to unmanaged AI apps in Edge for Business on Intune-managed Windows devices. Google Workspace DLP for Gemini, however, needs an Enterprise, Frontline or Education edition, which leaves out Business Starter, Standard and Plus.
  • Run a 15-minute session with real examples such as Panasonic and Samsung, and repeat the reminder when you add a new tool. People remember cases better than a long document.

Quick checklist

  • Written AI use policy, signed or acknowledged by every employee
  • At least one approved AI tool on a business plan with a contract
  • List of data that must never go into any AI tool
  • A named person staff tell when they paste something by mistake
  • Proxy or DNS logs reviewed for AI sites once a month
  • Unapproved AI apps blocked on managed devices where your licences allow it
  • Policy reviewed every six months

If you are preparing for ISO 27001, our ISO 27001 readiness assessment covers acceptable use, cloud services and supplier controls, including AI tools.

Related guides

Sources

Preparing for an ISO 27001 or IT audit? Our ISO 27001 readiness assessment finds your gaps before the auditor does, and our free security policy generator drafts the policies you need. Contact us for a quote.

Tags: , , , , ,
Post author avatar

Bhaskar Soni

Bhaskar Soni is a Security Engineer and the founder of Youngster Company, based in Ahmedabad, Gujarat. He works hands-on with Linux server hardening, vulnerability assessment and penetration testing (VAPT), ISO 27001 compliance and OSINT investigations. He writes practical, tested guides for IT professionals and Indian businesses, and every article on this site is written or reviewed by him.

Leave a Reply