Short answer: enterprise customers now add five groups of AI clauses to a SaaS master services agreement (MSA): no training on their data, ownership of outputs, notice of AI subprocessors, fast incident notice, and proof through logs and audit. You can sign them if your engineering can show each one. No training, output ownership and subprocessor notice are already standard in the public terms of the big model providers, so the real work is not the wording. It is a tenant-level AI switch, a provider register, retention and deletion that you control, logs, and an incident runbook. Add a new AI provider to your subprocessor list before you launch the feature, give notice through a channel customers can subscribe to, and renewals are less likely to stall.
This post is for SaaS vendors selling AI features to enterprise buyers. It covers the vendor side of the clause. Our post on client contracts that forbid public AI covers the other side, for service providers who receive such clauses. This is engineering guidance, not legal advice. Have your counsel review the wording of any clause before you agree to it.
Why are enterprise customers adding AI clauses to MSAs?
An AI feature changes the vendor's risk in three ways. Your customer's data may reach a model provider they never chose. A model may keep or learn from that data. And the output can be wrong, with a person or a regulator relying on it. A clause is the cheapest way for the buyer to push that risk back to you.
Law adds pressure on top. Under India's DPDP Act, a Data Fiduciary stays responsible for processing done by its Data Processor (Section 8(1)), and it can engage a processor only under a valid contract (Section 8(2)). Under the GDPR, a processor may not engage another processor without the controller's written authorisation, and must pass the same obligations down the chain (Article 28(2) and (4)). Your customer's contract with you is where those duties land. Our post on the DPDP processor contract covers the Indian side.
The EU AI Act adds role labels. A provider develops an AI system and puts it on the market under its own name. A deployer uses it under its own authority. An Indian SaaS vendor selling an AI feature to EU customers is usually a provider, and the EU customer is the deployer. The Act applies to third-country providers whose output is used in the Union (Article 2(1)). Buyers copy these labels into their MSAs.
Which AI clauses do enterprise buyers add, and what engineering supports each?

Clause wording varies by buyer. The groups below recur. The right-hand column is what lets you say yes and prove it.
| Clause group | What the buyer usually asks | Engineering that supports it | Evidence to hold |
|---|---|---|---|
| No training | What the buyer usually asks: The vendor and its model providers will not train on customer inputs or outputs | Engineering that supports it: Provider settings and contract terms that block training; no customer data in fine-tuning sets or shared evaluation sets | Evidence to hold: Provider terms for each service, account configuration screenshots, a data-flow diagram |
| Output ownership | What the buyer usually asks: The customer owns the outputs | Engineering that supports it: Outputs stored per tenant, exportable and deletable | Evidence to hold: Your terms, tenant data model |
| Subprocessor notice | What the buyer usually asks: A list of AI providers, advance notice of changes, a right to object | Engineering that supports it: One provider register behind the website list, DPA annex and questionnaire answers; a tenant-level switch to turn an AI provider off | Evidence to hold: Published list, change log, notification record |
| Location, retention, deletion | What the buyer usually asks: Where data is processed, how long it is kept, deletion on termination | Engineering that supports it: Retention settings at each provider; a deletion job that reaches provider-side stores; region pinning where offered | Evidence to hold: Retention table per provider, deletion job logs |
| Incident notice | What the buyer usually asks: Notice of an AI-related incident within a stated time | Engineering that supports it: Detection, per-request event logs, a runbook with a named owner | Evidence to hold: Runbook, test record, log samples |
| Oversight and accuracy | What the buyer usually asks: Output may be wrong; a human reviews regulated decisions | Engineering that supports it: Human review step and an override in the product; clear labelling of AI output | Evidence to hold: Product screens, workflow description, review logs |
| Audit and logs | What the buyer usually asks: The right to see how their data was processed | Engineering that supports it: Per-tenant, per-request logs of which model received what, with retention you control | Evidence to hold: Log schema, sample export, access-control record |
Liability caps, indemnities and warranties about accuracy are legal questions. Engineering gives counsel the facts, such as what the system logs and what a provider promises. Counsel decides the wording.
What do major model providers already promise in their public terms?
Your promise to a customer cannot be wider than the promise you hold from your model provider. Read the terms for the exact service you use. As of 20 November 2026, the public terms of the main providers cover the common clauses like this:
| Source | No training | Output ownership | Deletion and subprocessors |
|---|---|---|---|
| OpenAI Services Agreement (updated 1 December 2025, effective 1 January 2026) | No training: Section 4.2: OpenAI will not use Customer Content to develop or improve the Services, unless the customer explicitly agrees | Output ownership: Section 4.1: the customer keeps its Input and owns all Output | Deletion and subprocessors: Section 11.3: Customer Content deleted within 30 days of termination |
| OpenAI Data Processing Addendum (same dates) | No training: Not the main subject | Output ownership: Not the main subject | Deletion and subprocessors: Section 2.9: subprocessor changes notified, 30 days to object, termination if unresolved. Section 2.7: breach notice "without undue delay", with no hour figure |
| Anthropic Commercial Terms (effective 17 June 2025) | No training: "Anthropic may not train models on Customer Content from Services" | Output ownership: The customer keeps its Inputs and owns its Outputs | Deletion and subprocessors: Check the current DPA for the service you use |
| Common Paper AI Addendum v1.0 (CC BY 4.0) | No training: Section 1.3: the provider may not use Customer's Inputs or Outputs to train any model, unless the cover page allows it | Output ownership: Section 2.1: the customer owns all Output | Deletion and subprocessors: No subprocessor or incident-notice clause, so these come from the DPA |
No training, output ownership and subprocessor notice are now standard in major vendors' public terms. Breach notice in these DPAs is "without undue delay", so tighter hour limits are negotiated. These are template terms. Your signed contract and plan tier decide what actually applies, so check them before you flow a promise down. Our post on answering whether you train on customer data goes through the training question and the proof for it.
How do you add an AI provider to your subprocessor list without stalling renewals?
Renewals stall when a customer learns about an AI provider from a release note, not from you. The fix is a process that makes the notice routine. Do these in order:
- Decide whether the provider is a subprocessor. If a model provider processes customer personal data on your behalf, treat it as one. A provider that only sees your own internal data is not. Write the test down, and apply it the same way every time.
- Build one provider register. For each AI provider, record the service name, what it does, which data categories it receives, where it processes them, the retention setting, and the contract it sits under. The website list, the DPA annex and the questionnaire answers all read from this register, so they never disagree.
- List the provider before the feature ships. Add the provider to the public list and send notice before general release. Advance notice lets a customer raise concerns before launch, not after it.
- Give notice through a channel customers can subscribe to. A change log with an email or feed subscription, often on your trust centre, works. See our AI trust centre guide.
- Set a notice period you can keep. OpenAI's public DPA gives customers 30 days to object to a subprocessor change. That is one benchmark. Pick a period for your own MSA that your release process can meet, and write the same period into every contract.
- Prepare for an objection. Build a tenant-level switch that turns an AI feature off, or routes it to another provider, for one customer. Without it, an objection becomes a contract dispute. With it, an objection is a settings change.
- Keep a change log. Record what changed, when notice went out and who was notified. Auditors and procurement teams ask for it.
In the MSA, prefer notice with a right to object over a right to approve each change. Approval rights give every customer a veto on your roadmap. Counsel should draft this, and your largest customers may still ask for more. A security questionnaire will often ask the same questions, so the register saves you time there too.
How tight should AI incident notice be in an MSA?
Buyers ask for a number of hours. Do not promise a number you cannot meet. Notice runs from when you become aware, and you can only become aware of what you log and monitor.
ISO/IEC 42001 sets no reporting deadline. Its clause 10.2 requires you to control and correct a nonconformity, find its causes, review whether the action worked and keep documented evidence. Your legal clocks come from elsewhere:
| Source | Clock | Applies to |
|---|---|---|
| CERT-In directions | Clock: 6 hours | Applies to: Reportable cyber incidents in India |
| DPDP Rule 7 (from 13 May 2027) | Clock: Notify affected Data Principals and the Board without delay, with a detailed Board report within 72 hours | Applies to: Personal data breaches by a Data Fiduciary |
| EU AI Act, Article 73 | Clock: 15 days from awareness; 2 days for a widespread infringement or critical-infrastructure disruption; 10 days for a death | Applies to: Providers of high-risk AI systems reporting serious incidents to the market surveillance authority |
| Customer MSA | Clock: Whatever you agree | Applies to: Your contractual duty to the customer |
India has no AI-specific incident reporting rule as of November 2026. A customer's AI incident clause is therefore a contract term, not a legal duty, unless the event is also a personal data breach or a reportable cyber incident. Write definitions that separate three kinds: a security incident, a personal data breach, and an AI-quality incident such as a harmful or wrong output. Then set a time for each. Our post on breach reporting under DPDP and CERT-In covers the first two, and our guide to AI incident response covers the third.
Engineering can back a short time if three things exist. First, per-request event logs, which ISO 42001 control A.6.2.8 names as AI system recording of event logs. Second, monitoring of AI system operation (A.6.2.6). Third, a route for customers to report problems (A.8.3 and A.8.4 cover external reporting and communication of incidents). Without these, a six-hour promise is a promise to guess.
Do EU AI Act clauses apply to an Indian SaaS vendor?
They can. The Act reaches providers outside the EU where the output is used in the Union. Its dates have moved, but the obligations remain. The AI Omnibus (Regulation (EU) 2026/1744, OJ 24 July 2026) moved the high-risk deadlines: Annex III high-risk systems, such as employment and education uses, now apply from 2 December 2027, and Annex I high-risk systems from 2 August 2028. The transparency rules in Article 50 and the AI Office's powers over general-purpose models apply from 2 August 2026, with some Article 50 marking duties for existing systems from 2 December 2026. Do not tell a buyer that the high-risk rules apply from August 2026.
Buyers may ask you to confirm your role as provider, to cooperate with their own obligations, and to warn them if a feature falls into a high-risk use. They may also ask about AI literacy. The Omnibus softened the AI literacy duty, but a customer's contract may still ask for staff training. For fines, the Act as adopted in 2024 sets up to €35 million or 7% of worldwide turnover for prohibited practices, and up to €15 million or 3% for most other obligations, whichever is higher, or lower for SMEs. Check the current text with counsel. Our guide on the EU AI Act for Indian SaaS goes further.
Which AI clauses can you usually accept, negotiate or send to counsel?

| Group | Clauses | Why |
|---|---|---|
| Usually workable if you can show it | Clauses: No training; output ownership; subprocessor list and notice; deletion on termination; AI output labelled as AI | Why: Common in public provider terms, and each maps to a setting or a record you control |
| Negotiate with engineering input | Clauses: Hour-based incident notice; region pinning; audit and log access; a right to object to a provider; fixed retention periods | Why: They depend on what you can detect, where your providers process data, and what your logging captures |
| Take to counsel | Clauses: Liability caps; indemnities; warranties on accuracy; regulatory compliance warranties; sole-source or approval rights over providers | Why: They shift risk and need legal judgment. Engineering supplies the facts |
A pre-approved set of positions speeds this up. Write a short internal document: for each clause group, what you accept as written, what you accept with edits, and who must approve anything else. Sales can then answer without a long legal loop. This is the same idea as the evidence pack in enterprise deals stuck in AI security review.
What should you build before you sign these clauses?
Build in this order. Each item removes a negotiation, not only a risk.
- A provider register, as above, kept as structured data and not as a slide.
- A tenant-level AI control. Per customer, the ability to turn AI features on or off, choose a provider and set retention.
- Provider configuration you can prove. Zero-retention or no-training settings where offered, with a dated record of the setting.
- Per-tenant event logs. Which model received what, when, for which tenant, with a retention period you set.
- A deletion job that reaches providers. It must cover provider-side stores, backups on a documented schedule and derived data such as embeddings.
- An incident runbook with owner, triggers, a customer-notice template and a test record.
- A clause position sheet, as above.
Compliance platforms can automate evidence collection and control monitoring. As of 20 November 2026, Vanta lists ISO 42001, EU AI Act and NIST AI RMF support, varying by plan, and Drata lists ISO 42001 and NIST AI RMF. Neither publishes list prices, and neither builds tenant settings, audit logs or guardrails inside your product. That part is engineering work, as our platform versus consultant comparison sets out. Our post on ISO 42001 versus SOC 2 explains how certificates and reports fit alongside contract terms.
Where Mono helps
We are an engineering studio. We map your AI data flows, build the register, the tenant controls, the logs and the deletion jobs, and turn them into evidence a buyer's team can review. See AI governance for the service, and security questionnaires if a deal is waiting on one now. If a customer insists that no outside provider may touch their data, a model in your own cloud account removes the subprocessor question, and we cover that under Own Your AI. Mono is not a law firm and does not certify anyone. Legal wording goes to your counsel.
Sources: OpenAI Services Agreement (updated 1 December 2025, effective 1 January 2026), Sections 4.1, 4.2 and 11.3, openai.com/policies/services-agreement. OpenAI Data Processing Addendum, Sections 2.7 and 2.9, openai.com/policies/data-processing-addendum. Anthropic Commercial Terms of Service (effective 17 June 2025), anthropic.com/legal/commercial-terms. Common Paper AI Addendum v1.0 (CC BY 4.0), commonpaper.com/standards/ai-addendum/1.0. Digital Personal Data Protection Act, 2023, Section 8, and DPDP Rules, 2025, Rule 7. Regulation (EU) 2016/679 (GDPR), Article 28. Regulation (EU) 2024/1689 (AI Act), Articles 2, 3, 73 and 99, and the European Commission's AI Act page; AI Omnibus (Regulation (EU) 2026/1744, OJ 24 July 2026), with dates from the Commission and the Future of Privacy Forum. ISO/IEC 42001:2023, clause 10.2 and Annex A controls A.6.2.6, A.6.2.8, A.8.3 and A.8.4. CERT-In directions of 28 April 2022. Vanta and Drata public product pages, checked 20 November 2026. Last reviewed 20 November 2026.
FAQ
What AI clauses do enterprise customers add to SaaS MSAs?
The common groups are no training on customer inputs and outputs, customer ownership of outputs, a list of AI subprocessors with notice of changes, a time limit for incident notice, and audit or log rights. Some add human oversight for regulated decisions and a statement that output may be incorrect. Most have a matching setting or record your engineering can provide.
How do I add an AI provider to my subprocessor list without delaying renewals?
Keep one provider register, list the provider before the feature ships, and give notice through a channel customers can subscribe to. Set a notice period you can meet, and build a tenant-level switch so one customer's objection can be handled by changing a setting. OpenAI's public DPA gives customers 30 days to object, which is one benchmark.
Do OpenAI and Anthropic train on customer data under their business terms?
As of 20 November 2026, OpenAI's Services Agreement says it will not use Customer Content to develop or improve the Services unless the customer explicitly agrees. Anthropic's Commercial Terms say Anthropic may not train models on Customer Content from Services. Check the exact terms and plan for the service you use, because your signed contract decides what applies.
How quickly must a SaaS vendor notify customers of an AI incident?
No AI-specific rule in India sets a deadline as of November 2026, and ISO/IEC 42001 sets none either. The clock comes from your contract, plus CERT-In's 6 hours for reportable cyber incidents, DPDP Rule 7 for personal data breaches from 13 May 2027, and EU AI Act Article 73 for providers of high-risk systems. Promise only what your logging and monitoring can meet.
Does the EU AI Act apply to an Indian SaaS company?
It can. The Act covers providers outside the EU where the output of the AI system is used in the Union. An Indian SaaS vendor selling an AI feature to EU customers is usually a provider, and the EU customer is the deployer. The AI Omnibus moved the Annex III high-risk deadline to 2 December 2027, but obligations remain.
