← All news and guides
Data protection

Is Using ChatGPT at Work DPDP-Compliant? A Find-and-Fix Guide for Employers

Your staff already use ChatGPT, Copilot and Gemini. How to find which tools, what personal data goes in, and how to fix it before DPDP duties start in May 2027.

Mono Team ·

An open-plan office in an Indian city at dusk, one person typing at a laptop, the city lights blurred behind.

Short answer: staff using ChatGPT at work can fit the DPDP Act, but only on a work account your company has a contract for, with settings you have checked, and with rules on what personal data goes in. The Act doesn't ban AI tools. It applies to the personal data in the prompts, files and transcripts. The risk is the tools you don't know about: personal accounts, free plans and browser extensions, where your company has no contract and no control. The duties start on 13 May 2027.

This guide is for employers. It covers how to find which AI tools your staff use, what personal data goes into them, which vendor settings matter, how to fix the gaps, and a one-page staff policy you can adapt. For what the Act says about AI tools in general, see our service page on DPDP and AI tools.

Why staff use of AI is a DPDP question

When an employee pastes a customer complaint into ChatGPT to draft a reply, your company is processing that customer's personal data, and the AI vendor is processing it too. Four parts of the Act follow from that:

  • You stay responsible. You remain responsible for processing that a vendor does on your behalf (Section 8(1)).
  • You need a contract. Using a data processor for an activity related to offering goods or services requires a valid contract with it (Section 8(2)). An employee's personal account gives you none.
  • You need a lawful purpose. The customer's consent, or a legitimate use listed in the Act, has to cover this use of their data (Sections 4 to 7).
  • You need safeguards and deletion. Reasonable security safeguards cover processing by your processors (Section 8(5)), and you must be able to get the data erased when the purpose is over (Section 8(7)).

Staff data counts as well. CVs pasted in for screening, appraisal notes for rewriting and payroll sheets for analysis are personal data of employees and candidates. The Act lists employment purposes among its legitimate uses (Section 7(i)), but that doesn't remove the duties on security, contracts and deletion.

Step 1: Find which AI tools staff use

A coffee cup beside a closed laptop on a wooden desk, warm late-afternoon light across the surface.

No single source shows everything. Use five, and compare them. Each one finds tools the others miss.

SourceWhat it findsWhere to look
Single sign-on and OAuth logsWhat it finds: AI apps staff signed into with a work account, and third-party apps granted access to mail, files or calendarsWhere to look: Your identity provider's sign-in logs and app list. In Google Workspace and Microsoft 365, the list of third-party apps with OAuth access.
Expense and card dataWhat it finds: Paid AI subscriptions bought on company cards or claimed backWhere to look: Search expense claims and card statements for AI vendor names over the last 12 months
Browser extension inventoryWhat it finds: Writing, summarising and meeting tools that read the page or the screenWhere to look: Your device management or browser management console, exported per device
DNS, proxy or firewall logsWhat it finds: Traffic to AI services from company devices and networks, including free and personal accountsWhere to look: Query a month of logs for known AI domains, grouped by team, not by person
A short staff surveyWhat it finds: Phone apps, personal devices and what people actually paste inWhere to look: Five questions, anonymous, with a promise that the goal is approved tools, not discipline

Be careful with the logs. Looking at which employee visited which site is itself processing staff personal data. Report by team and tool, limit who sees person-level detail, and tell staff that the review is happening. Our shadow AI audit runs this discovery for you.

Five survey questions that work

  • Which AI tools do you use for work, on any device?
  • Do you use a work account, a personal free account or a personal paid account?
  • What do you use them for? Pick from drafting, summarising, analysing data, coding, meeting notes, other.
  • Have you ever pasted in customer, candidate or colleague details? Roughly what kind?
  • Which AI tool would you most want the company to approve?

Step 2: Classify what personal data goes in

For each tool, write down whose data goes in, what kind, and for which task. A simple three-level scale is enough to decide what to fix first.

LevelExamplesRule for staff
No personal dataExamples: Public text, generic code, marketing copy with no namesRule for staff: Any approved tool
Ordinary personal dataExamples: Names, emails, phone numbers, support tickets, CVs, call transcriptsRule for staff: Approved work tools only, and only what the task needs
High-risk personal dataExamples: Aadhaar numbers, PANs, bank and card details, health information, children's data, salary dataRule for staff: Not in general chat tools. Mask first, or use a tool set up for it.

The Act itself doesn't create a "sensitive" category. The third level is our working list of data where a leak does the most harm, or where sector rules add duties. Children's data has its own duties under Section 9.

Step 3: Check the vendor and account settings

The same product can sit under very different terms. A consumer account is a contract between the vendor and your employee. A business plan is a contract with your company. That difference decides most of what follows.

What the main vendors say, as of November 2026

Vendor and planModel trainingRetention and admin control
OpenAI: ChatGPT Free, Go, Plus, ProModel training: May be used for training. Users can opt out under Settings, Data controls, "Improve the model for everyone".Retention and admin control: Controlled by the user, not your company
OpenAI: ChatGPT Business, Enterprise, EduModel training: Not used for training by defaultRetention and admin control: Enterprise and Edu let admins control retention. SAML SSO is available.
OpenAI: APIModel training: Not used for training unless you opt inRetention and admin control: Abuse monitoring logs kept up to 30 days by default. Zero Data Retention needs OpenAI's approval.
Microsoft Copilot with Microsoft 365 work accountsModel training: Prompts, responses and Microsoft Graph data aren't used to train foundation modelsRetention and admin control: Stored under your Microsoft 365 commitments. Admins set retention in Microsoft Purview.
Google Gemini with qualifying Workspace editionsModel training: Not used for model training outside your domain without permissionRetention and admin control: Admins set history retention: Gemini app up to 36 months; Gemini in Workspace apps 90 days to indefinite.

These are summaries of each vendor's public pages, checked on 18 November 2026. Vendors change plan names and terms often. Read the current terms and the data processing addendum before you rely on them, and have your lawyers check whether the vendor acts only on your behalf.

Settings to record for each approved tool

  • Account type. Work accounts on a business plan, created through single sign-on. No personal logins.
  • Training. Off, and confirmed in the admin console or contract.
  • History retention. Set to the shortest period that works, and recorded.
  • Connectors and plug-ins. Only the ones you approved. Each connector to mail, drives or CRM widens what the tool can read.
  • Data location. Where prompts are stored and processed, from the vendor's terms.
  • Offboarding. Leavers lose access when their SSO account is disabled, and their history is handled under your retention rule.
  • Audit logs. Admin and usage logs exported to your own log store. Rule 6 asks for access logs kept for one year; our log retention guide covers the settings.

Step 4: Fix it without banning AI

A small meeting room with glass walls at night, empty chairs around a table, office lights softly blurred.

Bans tend to push use onto personal phones, where you see nothing. A better fix gives staff a good approved tool and makes the risky path harder.

  • Approve a short list. One general chat tool and the AI built into your office suite covers most needs. Put them on work accounts with a contract, and record the settings.
  • Move people across. Tell staff which tool replaces which, and give them a date. Then block sign-ups with work email to unapproved AI services where your tools allow it.
  • Mask before it leaves. For your own apps and AI API calls, remove or replace names, phone numbers, PANs and Aadhaar numbers before the prompt reaches the model, and log what was masked. Masked data that can be linked back to a person is still personal data in your systems.
  • Keep the most sensitive data in-house. For data that should never reach a third-party AI vendor, run models on infrastructure you control. See Own Your AI.
  • Cover deletion. Make sure you can find and delete a person's data in AI chat histories, uploaded files and any vector stores, while keeping the logs the Rules require.
  • Add AI vendors to your data map. Each approved tool goes in your processor register with the contract, settings and owner.

Most of this is engineering work, and it fits into a wider programme. A DPDP gap assessment includes AI tools alongside every other system, and DPDP implementation builds the masking, logging and deletion.

A one-page staff AI use policy: outline

Keep it to one page, so people read it. Adapt the outline below and have your lawyers review the final text.

  • 1. Purpose. We want you to use AI tools. This policy says which ones and what data may go in, so we protect our customers, candidates and colleagues.
  • 2. Approved tools. A named list, each on a work account through single sign-on. Anything not on the list needs approval from a named owner first.
  • 3. Never on personal accounts. No company or customer information in personal or free AI accounts, on any device.
  • 4. What may go in. No personal data: any approved tool. Ordinary personal data: approved tools, only what the task needs. High-risk data such as Aadhaar, PAN, bank, health, salary or children's data: not in chat tools.
  • 5. Check the output. You are responsible for what you send. Check facts, names and numbers before using AI output.
  • 6. Meetings. Tell attendees before an AI note-taker records a call. Use only the approved note-taker.
  • 7. Browser extensions and plug-ins. Install only approved ones. Many read every page you open.
  • 8. Mistakes. If you pasted something you shouldn't have, tell the security contact the same day. Reporting early is not a disciplinary matter.
  • 9. Owner and review. A named owner, a review date, and where to ask questions.

Checklist

  • 1. AI tools found from SSO, expenses, browser extensions, DNS or proxy logs and a staff survey.
  • 2. Each tool listed with the personal data that goes in, at one of three levels.
  • 3. Vendor terms read for each tool, and flagged to your lawyers where needed.
  • 4. A short list of approved tools on work accounts, with training off and retention set.
  • 5. Unapproved tools replaced, with a date, and blocked where possible.
  • 6. Masking in front of your own AI API calls, tested on Indian identifiers.
  • 7. AI vendors added to the processor register and data map.
  • 8. Deletion covers AI histories, files and vector stores.
  • 9. AI admin and usage logs kept for one year.
  • 10. A one-page staff policy published, with training and a named owner.

Want a quick read on where you stand? Take the DPDP self-check. To have the discovery and fixes done for you, see our DPDP and AI tools service.

This article is general information, not legal advice. Our partner law firm gives the written legal opinion on how the Act applies to you.

Sources: Digital Personal Data Protection Act, 2023, Sections 4 to 9. Digital Personal Data Protection Rules, 2025, Rule 6. OpenAI: Enterprise privacy, How your data is used to improve model performance, and Data controls in the OpenAI platform. Microsoft Learn: Data, privacy and security for Microsoft Copilot. Google: Generative AI in Google Workspace Privacy Hub. Last reviewed 18 November 2026.

FAQ

Can employees use ChatGPT under the DPDP Act?

Yes, if the company sets it up properly. Staff should use a business plan on work accounts that the company has a contract for, with model training off and retention set. Personal data should go in only where the task needs it and a lawful purpose covers it. Personal accounts are hard to fit, because the company has no contract with the vendor.

How do we find out which AI tools our staff use?

Combine five sources: single sign-on and OAuth app logs, expense and card data, the browser extension inventory from device management, DNS or proxy logs for AI domains, and a short anonymous staff survey. Report by team and tool, because person-level browsing data is itself employee personal data.

Does ChatGPT train on our company's data?

As of November 2026, OpenAI says it doesn't train on inputs or outputs from ChatGPT Business, Enterprise, Edu or the API by default. Free, Plus and Pro accounts may be used for training unless the user turns off "Improve the model for everyone". Check the current terms before relying on this.

What should a staff AI use policy cover?

Which tools are approved, that personal accounts are not allowed for work data, what personal data may go into which tool, checking output, AI note-takers in meetings, browser extensions, how to report a mistake, and a named owner with a review date. Keep it to one page.

Should we ban ChatGPT at work?

We don't recommend it. Bans usually move AI use onto personal phones and accounts, where the company can't see or control it. Approving a few tools on work accounts, with clear rules on personal data, gives you more control.

Find out where personal data meets AI in your company.