← All news and guides
Private AI

ChatGPT Enterprise vs Azure OpenAI vs Copilot vs Self-Hosted: Which Fits a DPDP-Bound Company

ChatGPT Enterprise, Azure OpenAI, Microsoft Copilot or self-hosted? A dated comparison of data location, training, retention and fit for DPDP-bound firms.

Mono Team ·

A long pale office corridor with four plain doors along one wall, each slightly ajar with a different shade of soft light spilling out, late afternoon.

Short answer: for most DPDP-bound companies, a managed enterprise product is fine, and self-hosting is the exception, not the rule. The DPDP Act does not ban foreign AI tools and does not make you host a model in India. What it asks is that you stay responsible for the personal data (Section 8(1)), have a contract with each processor (Section 8(2)), and keep reasonable safeguards (Section 8(5) and Rule 6). Pick ChatGPT Enterprise or Microsoft 365 Copilot for staff productivity, Azure OpenAI or a cloud model API when you build AI into your own product, and self-hosted open-weight models when you must control where inference runs, pin a model version or cap a bill that grows with usage. The choice turns on four facts: where prompts are processed, whether the vendor trains on them, how long they are kept, and who can see them. Those facts change often, so the tables below are dated 20 November 2026.

This guide is for founders, CTOs and IT heads at companies that handle personal data. It compares ChatGPT Enterprise, the OpenAI API, Azure OpenAI, Microsoft 365 Copilot, Claude (its own API or Amazon Bedrock) and models you run yourself. The core DPDP duties start on 13 May 2027. This is engineering guidance, not legal advice. Which clauses your regulator or customers expect is a question for counsel.

What does DPDP actually ask of an AI product?

The Act is technology-neutral. It applies to the personal data inside prompts, uploads, transcripts, embeddings and logs. For any AI product, five duties matter.

  • You stay responsible. Under Section 8(1) the Data Fiduciary is responsible for processing done by a Data Processor on its behalf. Buying a product does not move that duty.
  • A valid contract. Section 8(2) allows a processor to be engaged only under a valid contract. Read the vendor's data processing terms.
  • Safeguards. Rule 6 sets minimum measures: encryption or masking, access control, logs and monitoring, continuity, one-year log retention and processor clauses.
  • Erasure and retention. Section 8(7) asks you to erase data when consent is withdrawn or the purpose ends, and to make your processors erase it too. So you need to know how long each tool keeps prompts and whether you can delete them.
  • Cross-border transfers. Section 16(1) lets the government restrict transfers to notified countries. As of 20 November 2026 none are notified. Section 16(2) keeps stricter sector rules alive.

Two points follow. First, DPDP does not require data to stay in India. Hosting in India avoids a possible future Section 16 restriction and some sector rules, but it is not a legal requirement as of 20 November 2026. Second, sector rules can be stricter. The RBI's circular RBI/2017-18/153 of 6 April 2018 says payment system data must be stored only in India. If you handle payment data, that rule shapes your choice more than DPDP does. Our post on cross-border transfers and AI APIs covers Section 16, the RBI rule and its limited window for processing abroad.

Can a regulated business use public AI tools and APIs at all?

Often yes. Few rules say "never use an AI API". What stops regulated firms is usually one of these:

  • A sector storage rule, such as the RBI payment-data rule above.
  • Customer contracts. A bank or enterprise customer may bar sub-processors, require India-only processing or demand deletion proof you cannot give.
  • Missing vendor terms. The tool is a free or personal plan with no data processing agreement, and training on your data may be on by default.
  • No deletion path. Prompts and files sit where you cannot erase them.
  • Your own risk appetite. Your board may rule out any data leaving your control.

The fix is rarely a blanket ban. Approve work accounts with a contract, set the retention and training options in writing, and keep personal data that the task does not need out of prompts. Our guide to ChatGPT at work under DPDP shows how to find and fix staff use. When none of the managed options meet a hard requirement, that is the case for self-hosting.

What are the five ways to get a private ChatGPT for your company?

A row of five plain wooden doors of different sizes along a long pale corridor, each slightly open to a different shade of soft light, no signs or numbers.

"Private ChatGPT" can mean five different things. They sit on a scale from "vendor runs everything" to "you run everything".

OptionExamplesWho runs the modelYour data pathBest for
1. Vendor workspaceExamples: ChatGPT Enterprise, Microsoft 365 CopilotWho runs the model: The vendorYour data path: Staff prompts go to the vendor under an enterprise contractBest for: Staff productivity, fast start
2. Hyperscaler model API in your tenantExamples: Azure OpenAI in Microsoft Foundry, Claude on Amazon BedrockWho runs the model: The cloud providerYour data path: Your app calls a model hosted inside the provider's cloud under your accountBest for: AI features in your own product
3. Model vendor API, your own interfaceExamples: OpenAI API, Claude APIWho runs the model: The model vendorYour data path: Your app calls the vendor's API directlyBest for: Product features where the vendor's terms are enough
4. Open-weight model in your cloud accountExamples: An open-weight model on a VPC GPU instanceWho runs the model: YouYour data path: Prompts stay inside your cloud networkBest for: Control over inference location and model version
5. On-premises or Indian GPU cloudExamples: Your own servers, or rented GPUs from an Indian providerWho runs the model: YouYour data path: Prompts stay on hardware you control or rentBest for: Strict data control, steady heavy load

Options 1 to 3 are products you buy. Options 4 and 5 are systems you operate. The cost model for 3 against 4 and 5 is in our private LLM against API cost model in rupees. The hosting trade-offs for 4 and 5 are in VPC against on-premise against Indian GPU cloud.

Where does each option process and store your data?

This is the question that most often decides the choice, and the answers are not what people assume. Storage and processing are separate. A vendor may store data in India and still run the model elsewhere. The summary below is dated 20 November 2026. Our post on cross-border transfers and AI APIs has the fuller provider table, including logs and the legal background. Re-check each vendor page before you rely on it.

ProductCan the model run in India?What to check
ChatGPT Enterprise and EduCan the model run in India?: No. India is available for storage at rest, but inference residency covers only the EEA and Switzerland, the US and the UAEWhat to check: Whether your contracts need India processing
OpenAI APICan the model run in India?: Partly. India processing on in.api.openai.com covers only gpt-6-luna, on Standard pay-as-you-go, without hosted toolsWhat to check: Eligibility through OpenAI sales, and the 10% data residency uplift on eligible models released on or after 5 March 2026
Azure OpenAI in Microsoft FoundryCan the model run in India?: Depends on deployment type and model. Standard deployments process in the geography you choose. Global may process in any geography where the model is deployed. DataZone processes within the US or EU zoneWhat to check: Which models your Indian region offers
Microsoft 365 CopilotCan the model run in India?: Not fixed. Outside the EU, queries may be processed in the US, EU or other regions (page updated 4 November 2026)What to check: What applies to your tenant, in writing
Claude API (Anthropic, first party)Can the model run in India?: No. inference_geo is global or us, and workspace geo is US onlyWhat to check: Use Bedrock if you need India processing
Claude on Amazon BedrockCan the model run in India?: Yes for Claude Opus 5, Sonnet 5 and Haiku 4.5, through the India profile (since 29 September 2026), which routes only between ap-south-1 and ap-south-2What to check: That your exact model is in the profile
Self-hosted open-weight modelCan the model run in India?: Wherever you put itWhat to check: Your own architecture

Three details trip people up.

  • ChatGPT Enterprise can store in India but not run there. Prompts are processed in the inference regions listed above. If your customer contract says "processed in India only", ChatGPT Enterprise cannot meet it as of 20 November 2026.
  • Azure region is not the same as deployment type. A resource in an Indian region with a Global deployment may still process prompts elsewhere. Stored data stays in the resource's geography either way. Model availability by Indian region changes, so check it for your region.
  • Copilot's processing location is not fixed for non-EU tenants. Microsoft has announced in-country processing for India, but confirm what applies to your tenant in writing before you promise anything to a customer.

Do these vendors train on your data, and how long do they keep it?

For the enterprise tiers, the training answer is the same across vendors: no by default. Retention is where they differ, and it is where DPDP erasure and Rule 6 logging meet your contract.

ProductTraining on your dataRetention and logging points to check
ChatGPT Enterprise and EduTraining on your data: OpenAI says it does not train its models on your data by default (enterprise privacy page, updated 8 January 2026)Retention and logging points to check: Admins can control retention. SAML single sign-on is available. Confirm how long deleted chats persist
OpenAI APITraining on your data: Not by defaultRetention and logging points to check: Abuse monitoring logs kept up to 30 days. Zero data retention needs OpenAI's approval. Chat Completions cannot set store=true outside the US
Azure OpenAI in Microsoft FoundryTraining on your data: Prompts, completions and embeddings are not available to OpenAI and are not used to train foundation models without your permissionRetention and logging points to check: Flagged prompts may be stored in your geography and reviewed by Microsoft staff for abuse. You can apply for modified abuse monitoring, but per the 17 November 2026 change log, it will not be available for models made available on or after 1 November 2026, from a date to be announced. Existing approvals continue for earlier models. Ask for the retention period in writing
Microsoft 365 CopilotTraining on your data: Prompts, responses and Graph data aren't used to train foundation LLMsRetention and logging points to check: Interactions are stored with your Microsoft 365 content. Admins set retention in Microsoft Purview. Copilot has opted out of abuse-monitoring human review
Claude APITraining on your data: Not used for training without express permissionRetention and logging points to check: Default deletion within 30 days (Anthropic privacy page, updated 1 July 2026). Zero data retention is available for eligible models
Claude on Amazon BedrockTraining on your data: Governed by your AWS termsRetention and logging points to check: AWS's data retention page says Fable 5.x models need the aws_review setting, with up to 30 days of retention. Read the page for each model you choose

Whatever you choose, ask for the retention period, the deletion method and the sub-processor list in writing. Those are what you need to answer a customer questionnaire and to wire the vendor into your erasure process.

How is Microsoft 365 Copilot different from the others?

Copilot is not a chat box with a model behind it. It is an assistant wired into Microsoft 365. It reads email, chats, meetings and files through Microsoft Graph, and it shows each person only content they have at least view permission for. That is good news and a warning.

  • The good news. Copilot honours the permissions, sensitivity labels and usage rights you already set.
  • The warning. If your SharePoint sites and shared drives are over-shared, Copilot will surface what people could always reach but never found. Fix permissions before you roll out seats.
  • Agents and web search. Agents have their own privacy terms, and web search sends a generated query to Bing. Admins can limit web search through Microsoft 365 privacy controls.

So Copilot suits firms that already live in Microsoft 365 and have clean permissions. If your knowledge sits in many systems, you will likely build your own assistant on a model API and enforce permissions yourself. Our post on permission-aware RAG shows how.

When is self-hosting the right answer?

Self-hosting moves every control to you, and every job. You pick the model, the servers, the network rules, the logs and the upgrades. It makes sense when at least one of these is true:

  • A hard residency rule. A customer contract or sector rule says prompts must be processed in India, and no managed option you can use meets it.
  • A pinned model. You need the same model version for years, for audit or validation, and cannot accept a vendor retiring it.
  • A bill that must not move. Your load is steady and heavy enough that rented GPUs beat per-token prices.
  • Data you will not send out. Some documents never leave your network.

It does not make sense just because "private" sounds safer. A self-hosted system without access control, logging and a delete path is private only in name. You also need people to run it. Plan for at least one engineer who is comfortable with GPUs, plus cover for leave and incidents. See do you need an MLOps team. If you cannot staff them, buy a managed option, or ask a partner to build and run it. This is what our Own Your AI service is for.

Which option fits which company?

Two pairs of hands arranging small wooden blocks of different sizes on a plain table in soft daylight, the blocks forming two separate groups, no text or markings.
Your situationUsual fitWatch for
Staff drafting, summarising and searching in Microsoft 365, clean permissionsUsual fit: Microsoft 365 CopilotWatch for: Over-shared sites, agents and web search settings
Staff want a general assistant, no Microsoft 365 dependencyUsual fit: ChatGPT EnterpriseWatch for: Inference not in India, retention settings, uploads
Building AI features into your SaaS or app on customer personal dataUsual fit: Azure OpenAI or Bedrock in your cloud account, or a direct API with a signed DPAWatch for: Deployment type, retention and abuse-monitoring terms, sub-processor lists for your customers
Payment data under RBI storage rulesUsual fit: Option 4 or 5, or a managed option that stores the data in India and, if it processes abroad, meets the RBI FAQ's deletion and return windowWatch for: Counsel and your bank partner should sign off
Customer contract says India-only processingUsual fit: Self-hosted, the Bedrock India profile for the models it lists, or the OpenAI API India region for gpt-6-lunaWatch for: Confirm the exact model, features and logs stay in India
Steady heavy volume and a platform teamUsual fit: Self-hosted open-weight, possibly beside an API for hard tasksWatch for: Evaluation, upgrades and on-call

Many firms use two or three together: Copilot or ChatGPT Enterprise for staff, a cloud model API for the product, and a small self-hosted model for sensitive jobs. That is fine if one record shows which tool handles which data.

What should you ask any vendor before you sign?

  • Where are prompts and outputs stored, and where is the model run? Ask for both in the contract, not the website.
  • Is training on our data off by default, and can it change without our consent?
  • How long are prompts, files and logs kept, including for abuse monitoring? Who can read them?
  • How do we delete one person's data, and how do we prove it was deleted?
  • Which sub-processors touch the data, and how will you tell us before you add one?
  • What do you tell us if there is a breach, and how fast? Rule 7 expects you to tell the Board and affected people without delay.
  • Which regions and models are available to our account on the date we sign, in writing? Region lists change.
  • What are the exit terms: export, deletion and the date by which our data is gone?

Keep the answers, the plan or tier you bought, the settings you chose and the date you checked in one record per tool. That record is your evidence for a customer questionnaire, and your AI security questionnaire answers will draw on it.

Sources: Digital Personal Data Protection Act, 2023, Sections 8 and 16. Digital Personal Data Protection Rules, 2025, Rules 6 and 7. RBI circular RBI/2017-18/153, 6 April 2018, and RBI FAQ on storage of payment system data. OpenAI Help Centre: data residency and inference residency for ChatGPT; OpenAI Enterprise privacy (updated 8 January 2026); OpenAI developer docs: data residency and your data. Microsoft Learn: data, privacy and security for Foundry Models sold by Azure (change log 17 November 2026); data, privacy and security for Microsoft Copilot (updated 4 November 2026). Anthropic: data residency documentation and privacy centre article on data retention (updated 1 July 2026). AWS: Amazon Bedrock expands Claude model availability to India; Bedrock data retention documentation. Vendor details checked 20 November 2026 and liable to change. Last reviewed 20 November 2026.

FAQ

Is ChatGPT Enterprise DPDP compliant?

No product is compliant on its own. The Act puts the duties on you as the Data Fiduciary. ChatGPT Enterprise is not trained on your data by default and offers India storage at rest, but as of 20 November 2026 its inference residency covers only the EEA and Switzerland, the US and the UAE. You still need a data processing agreement, retention settings, staff rules and a deletion path. The core DPDP duties start on 13 May 2027.

Does the DPDP Act require AI to run on servers in India?

No. The Act does not require data to stay in India. Section 16(1) lets the government restrict transfers to notified countries, and none are notified as of 20 November 2026. Sector rules can be stricter, for example the RBI rule that payment system data be stored only in India. Hosting in India avoids a possible future restriction but is not a legal requirement as of 20 November 2026.

Is Azure OpenAI better than ChatGPT Enterprise for data privacy?

They suit different jobs. Azure OpenAI is a model service you call from your own app, inside your Azure account. Microsoft says prompts and completions are not available to OpenAI and are not used to train foundation models without your permission. Standard deployments process data in the geography you choose, but Global and DataZone deployments may process elsewhere. ChatGPT Enterprise is a finished workspace for staff. Compare them on location, retention and admin controls for your use case.

Where does Microsoft 365 Copilot process data for Indian companies?

Microsoft's privacy page, updated 4 November 2026, says customers outside the EU may have queries processed in the US, EU or other regions. Microsoft has announced in-country processing for India, but you should confirm what applies to your tenant in writing. Copilot also shows each user only content they have permission to see, so permission clean-up matters before rollout.

When should a company self-host an LLM instead of buying ChatGPT Enterprise or Copilot?

Self-host when a hard rule or contract says prompts must be processed in India and no managed option meets it, when you need a pinned model version, or when steady heavy volume makes rented GPUs cheaper than per-token prices. You also need people to run it. If you have no platform engineer who can operate GPUs, a managed option is usually the safer choice.

Not sure which AI setup fits your data? We can map it with you.