Short answer: a client contract that forbids public AI rarely forbids all automation. Read the clause for what it actually restricts: third-party AI tools, sending data out of a region, vendors keeping or training on data, or any AI use without consent. Work on your own non-client data, automation that uses no AI, and AI on infrastructure you control are often still open, depending on the wording. For anything else, send the client a short written proposal and get approval before you start.
This post is for BPO and KPO operators in India serving clients in the US, UK and EU. It describes the common shapes of AI clauses in general terms, what each one usually leaves open, and how to ask for approval. It is not legal advice. Your contract's exact words, and your client's reading of them, decide what you can do.
Why clients added AI clauses
Your client usually isn't being difficult. They have their own obligations, and an AI tool your team uses can become part of their supply chain. Two examples:
- GDPR. If your client is the controller and you are its processor, Article 28(2) says a processor "shall not engage another processor without prior specific or general written authorisation of the controller." An AI service that receives client data can be another processor. Article 28(3)(a) also says the processor processes personal data "only on documented instructions from the controller."
- HIPAA. A business associate may disclose protected health information to a subcontractor only after getting satisfactory assurances that the subcontractor will safeguard it (45 CFR 164.502(e)(1)(ii)). An AI vendor handling that data is in the same position.
On top of the law, clients have promises to their own customers, their own security teams, and their own worries about data ending up in a model they don't control. A clause is the simplest way for them to keep that risk out. Your job is to show them a version of AI use that keeps the risk out too.
The five common clause shapes
Contracts word this differently, but AI clauses tend to take one or more of these shapes. The descriptions below are generic, not quotes from real contracts.
1. A ban on third-party AI tools
The clause says client data must not be entered into, processed by, or shared with any third-party generative AI or machine learning service. Sometimes it names tools. Sometimes it covers "any external AI".
What is often still open: AI that is not third-party. A model running on servers you own, or in a private cloud account you control, may fall outside a clause written against outside services. Read the definition carefully. Some clauses ban "generative AI" outright, wherever it runs, and then this route is closed without consent.
2. AI use only with prior written consent
The clause allows AI, but only after the client approves it in writing, often per tool or per use case. This is the most workable shape. It's the same logic as sub-processor approval: the client wants to know who touches the data before it happens.
What is often still open: anything the client approves. The work is in the proposal, covered below.
3. Data residency
The clause says client data must stay in a named country or region, or must not leave the delivery centre's network. This affects AI because many AI services process requests in whichever region has capacity. Microsoft's documentation for its Azure-hosted models, for example, says that for "Global" deployments, "prompts and responses may be processed in any geography where the relevant model sold by Azure is deployed." A standard regional deployment keeps processing within the chosen geography.
What is often still open: AI that runs inside the required region, on infrastructure where you can show where every request goes.
4. No retention and no training
The clause says no vendor may store client data beyond the task, or use it to train or improve a model. Many enterprise AI terms promise no training. Microsoft's page for its Azure-hosted models, for example, says prompts and completions are not used to train, retrain or improve the base models. Retention is harder. Microsoft's same page says prompts and completions can be stored for human abuse review unless a customer is approved for modified abuse monitoring, and that this option will not be offered for models made available from 1 November 2026 onwards, from a date still to be announced. Read the vendor's retention terms, not only the training terms.
What is often still open: a vendor whose terms match the clause, with evidence, or a private model where you set retention yourself.
5. Audit and logging rights
The clause lets the client audit how their data is processed, sometimes with a right to logs. This doesn't ban AI. It means you must be able to show, after the fact, who sent which data to which model and what came back.
What is often still open: any approved AI use that is logged per request and per client, with logs you can hand over.
Clause by clause: what you can still do

| Clause shape | What you can usually still do | Check before you start |
|---|---|---|
| Ban on third-party AI | What you can usually still do: Non-AI automation; AI on your own non-client data; possibly a private model you host | Check before you start: Whether the ban covers all AI or only outside services |
| Consent required | What you can usually still do: Any use the client approves in writing | Check before you start: Who can approve, and whether approval is per tool or per use case |
| Data residency | What you can usually still do: AI that processes inside the required region | Check before you start: The provider's deployment type and processing location |
| No retention, no training | What you can usually still do: Vendors whose terms match; private models where you set retention | Check before you start: Abuse-monitoring storage, logs and backups, not just training |
| Audit and logging rights | What you can usually still do: Approved AI use with per-request logs | Check before you start: That logs record user, client, data type, model and time |
| Several of the above at once | What you can usually still do: Usually a private model in the right region, with consent | Check before you start: Each condition separately, in writing |
Five routes that are often still open
- Work on data that isn't the client's. Internal training material, quality frameworks written from scratch, job descriptions, your own HR and finance processes. Be careful with SOPs and call scripts: if they contain client information, they may count as client data under the contract.
- Automation without AI. Rules engines, scripts, robotic process automation, macros, templates and workflow tools. Copy-paste between systems, for example, can often be automated without any model. Check that the contract's definition of AI doesn't also catch these tools.
- AI on de-identified data, if the contract allows it. Some clauses only cover personal data or confidential client data. If so, AI on properly de-identified or synthetic data may be allowed. Many contracts define client data to include anything derived from it, so check, and remember that free text such as call transcripts is hard to de-identify reliably.
- Approved enterprise AI with client consent. A business tier of a major AI service, configured to meet the client's residency and retention terms, and approved in writing.
- A private model on infrastructure you control. An open model running on your own servers or in your own private cloud account, in the required region, with no outside connections. It may sit outside a third-party AI ban. Under a consent clause it still needs approval, but it's often an easier approval to get.
How to ask the client for approval
Clients say no to vague requests and yes to specific ones. Send a short written proposal, two pages at most, that their security and privacy teams can review without a meeting. Include:
- The use case. One task, described plainly. For example: drafting replies to routine billing emails for agents to review.
- The benefit to the client. Faster handling, more consistent quality, fewer errors. Avoid promising numbers you haven't measured.
- The data. Which fields reach the model, which are removed or masked first, and which never go near it.
- The data flow. A simple diagram from source system to model to output, with every hop and location marked.
- The model and where it runs. Model name, who hosts it, the region, and whether any request can leave your network.
- Retention. How long prompts, outputs and logs are kept, where, and how they are deleted. Confirm that no data is used for training.
- Access and logging. Who can use the tool, how they sign in, and what is logged per request. Offer the client log access under their audit clause.
- Human review. Who checks each output before it reaches the client's customer or system, and what happens when the AI is wrong.
- A pilot and an exit. A small, time-boxed pilot on one process, the measures you'll report, and how you'll switch it off and delete everything if the client asks.
Ask for the approval in writing, scoped to that use case, and keep it with the contract. If the client later adds a new process, send a new proposal. Approval for email drafting is not approval for call analysis.
A note on India's DPDP Act
For client data about people outside India, India's Digital Personal Data Protection Act, 2023 gives Indian service providers a specific exemption. Section 17(1)(d) says most of the Act's duties do not apply where "personal data of Data Principals not within the territory of India is processed pursuant to any contract entered into with any person outside the territory of India by any person based in India."
Three limits matter:
- Some duties remain. The exemption excludes Section 8(1) and 8(5). So responsibility for compliance for processing done by you or on your behalf, and the duty to take reasonable security safeguards to prevent a personal data breach, still apply.
- It covers people outside India only. Your own staff, and any client data about people in India, are fully in scope. AI tools that monitor or score your agents process Indian employees' personal data.
- Timing. Section 17 starts with the Act's main duties on 13 May 2027.
The exemption also doesn't touch your client's own laws or your contract. GDPR, HIPAA and the AI clause still bind you through the client. Our guide to DPDP and AI tools covers the Indian side in more detail. Ask your lawyers how the exemption applies to your contracts.
Close the gap before the client finds it

The biggest risk usually isn't a planned AI project. It's agents pasting client emails into a free chatbot on their phones, because it's faster. A clause you're breaking without knowing is worse than one you've negotiated. A shadow AI audit shows which tools your teams already use, and an AI governance policy gives them an approved alternative.
If your contracts point towards private models, Mono's Own Your AI service runs open models in a private, isolated section of your own cloud account, or on servers you own. It can run on a network with no internet connection at all. Personal details are removed before the AI sees a request, outgoing traffic is blocked, and every request is logged. That gives you the data flow, location and logs a client proposal needs. Private cloud setups often take two to four weeks; your own servers or an offline network take up to eight. If you're also weighing cost, our post on LLM bill shock explains when private models pay off and when they don't. For India-hosted options and what the DPDP Act does and does not require, see private LLM deployment in India. If you're not sure where to start, see AI adoption.
This article is general information, not legal advice. How a clause applies to your work depends on its exact wording and your client's agreement.
Sources: Digital Personal Data Protection Act, 2023, Sections 8 and 17 (Gazette of India, 11 August 2023). Digital Personal Data Protection Rules, 2025, Rule 1 (G.S.R. 846(E), 13 November 2025), and the MeitY notification of the same date bringing provisions of the Act into force. GDPR, Article 28 (Regulation (EU) 2016/679). HIPAA Privacy Rule, 45 CFR 164.502(e). Microsoft Learn, "Data, privacy, and security for Foundry Models sold by Azure" (checked November 2026).
FAQ
Can a BPO use generative AI on client data?
Only if the client contract allows it. Many contracts ban third-party AI, require written consent, or limit where data can be processed and kept. Read the AI, confidentiality, sub-processor and data location clauses together. If the contract is silent or requires consent, send the client a specific written proposal and get approval first.
Our client says no ChatGPT on their data. What can we still automate?
Often a lot. Automation that uses no AI, such as scripts, rules and workflow tools, is usually unaffected. AI on your own internal data is usually fine. Depending on the wording, a private model running on infrastructure you control, or an approved enterprise AI service, may be possible with the client's written consent.
Does a private AI model get around a client's AI ban?
Not automatically. If the clause bans third-party or external AI services, a model you host yourself may fall outside it. If it bans generative AI wherever it runs, or requires consent for any AI, you still need approval. A private model in the right region usually makes that approval easier to get.
Does India's DPDP Act apply to client data of US, UK or EU customers?
Mostly not. Section 17(1)(d) exempts processing of data about people outside India done in India under a contract with a person outside India, from most of the Act's duties. Sections 8(1) and 8(5), responsibility and reasonable security safeguards, still apply. The exemption doesn't cover your staff or people in India, and starts on 13 May 2027.
What should a proposal to a client for AI use include?
One specific use case, the data fields the model sees and which are masked, a data flow diagram, the model name and where it runs, retention and deletion, a no-training statement, access control and per-request logging, who reviews each output, and a small pilot with a clear way to switch it off.
