DEV Community

Cover image for The AI data processing agreement: nine clauses to check before you sign
Azrty
Azrty

Posted on Originally published at azrty.com

The AI data processing agreement: nine clauses to check before you sign

Your procurement team has forwarded a standard data processing agreement from an AI vendor, attached to a services agreement presented as non-negotiable. It reads like the GDPR-style paperwork you have seen before, but it is missing the clauses that matter most for AI: what happens to your prompts, whether they train the vendor's next model, and where the data actually goes. Here is the checklist we use, clause by clause, so you know exactly what to demand before you sign.

Whether you are a CTO signing off a copilot rollout or a business leader accepting a vendor's click-through terms, the data processing agreement (DPA) is where the real risk allocation happens. The legal question of whether a UAE company may put customer data into a tool like ChatGPT is answered in our earlier article. This one is about the contract itself: the document your regulator, your auditors and your customers will ask for after something goes wrong.

Why a standard SaaS DPA is the wrong starting point

Most vendors hand you a DPA written for productivity software: store a file, send an email, run a report. AI workloads break that model in four ways. Your inputs are unstructured and unpredictable. The vendor's infrastructure may use your content for model improvement unless you stop it. The processing chain is long: cloud host, moderation contractor, support desk, evaluation vendor. And the output may itself become a decision about a person.

Point Standard SaaS DPA What an AI DPA must add
Purpose of processing "Provide the services" Prompts, outputs, embeddings, fine-tuning files and logs named separately
Model training Silent, or "improve the services" Express prohibition on training and fine-tuning on your data
Sub-processors Generic affiliates list Named list with country and purpose for each, plus notice and objection
Transfers "Wherever the vendor operates" Countries named, transfer mechanism named under PDPL Articles 22 and 23
Retention "Per our retention policy" (a URL) Fixed periods, deletion on exit, treatment of backups
Breach notice "Without undue delay" A hard cap in hours, because your clock to the regulator is 72 hours
Audit "On reasonable request" Annual evidence, plus audit rights after an incident
Rights requests "Reasonable assistance" Response times that fit your own 30-day clock
Automated decisions Missing entirely Human review and objection handling (PDPL Article 18)

What the PDPL obliges you to put in the contract

The federal Personal Data Protection Law, Federal Decree-Law 45 of 2021, in force since 2 January 2022, is unusually explicit about contracts. If you sit in DIFC or ADGM, the federal law does not apply to you, but the DIFC Commissioner's controller-processor guidance is the best available template and covers the same ground.

Four provisions drive almost everything below:

  • Article 7(5): you must appoint only processors with "sufficient guarantees" to implement the required technical and organisational measures. If you sign without checking, the failure is yours.
  • Article 8(1): processing must follow your instructions and the contract, which must specify "the scope, subject, purpose, nature and type of Personal Data, and the category of the Data Subject". A one-line purpose clause does not satisfy this.
  • Articles 8(3), 8(4), 8(7) and 8(8): the processor works to a specified purpose and period, then erases the data or hands it back; keeps a record including "any data related to the cross-border movement" of the data; and proves its compliance to you or to the UAE Data Office on request.
  • Article 8(10): where more than one processor is involved, their roles and obligations must be defined in writing, "otherwise they shall be deemed jointly responsible". This is why the sub-processor chain matters.

On top of that: Article 9 governs breach reporting, Article 20 sets the security baseline (encryption, pseudonymisation, resilience, restore capability, testing), and Articles 22 and 23 govern cross-border transfers. PwC's data privacy and responsible AI handbook for the UAE summarises the minimum contract content as scope, nature and purpose of processing, the types of personal data and categories of data subjects, the processor's minimum obligations, the controller's rights, and the processor's duty to erase or hand over the data at the end.

The nine clauses

1. Documented instructions and processing scope

Demand a Schedule 1 that names the AI-specific data types: prompts, completions, generated outputs, embeddings, fine-tuning or training files, evaluation datasets, feedback labels and API logs. It must also name the categories of data subject (your customers, your staff, third parties mentioned in free text) and state that processing is continuous for the term. OpenAI's Data Processing Addendum, effective 1 January 2026, does this in its Schedule 1 and states plainly that it acts as data processor and processes "only in accordance with Customer Instructions".

AI twist: your "documented instructions" should include the configuration choices you make in the product, not just the signed contract. Retention settings, abuse-monitoring options and opt-in toggles are instructions. Record them.

2. No training on your data, and no silent human review

The contract, not the marketing page, must say that the vendor will not use your inputs, outputs or feedback to train, fine-tune, evaluate or improve any model, and will not allow humans to read your content except to investigate abuse or a support case you raised. Where the vendor wants improvement rights, it must be a separate written opt-in you can refuse without losing the service.

OpenAI states it does not train on data from ChatGPT Enterprise, Business, Edu or the API by default, with opt-in only via the API dashboard. Anthropic goes further in its Commercial Terms of Service, effective 17 June 2025: "Anthropic may not train models on Customer Content." Both are worth citing in your review, but write the rule into your contract anyway: product pages change, contracts do not.

Watch for weasel words such as "aggregated or de-identified data" with no definition. The PDPL defines anonymisation strictly: data must not be linkable to a data subject. Ask how the vendor achieves that before you accept the carve-out.

3. Cross-border transfer mechanism (Articles 22 and 23)

Article 22 allows transfers to jurisdictions the UAE Data Office recognises as having an adequate level of protection. No such whitelist has been published, so in practice you rely on Article 23(1)(a): a contract or agreement obliging the recipient to adopt the measures, controls and requirements of the PDPL, with enforceability provisions. PwC's handbook also lists the safeguards the Executive Regulations contemplate for transfers: standard contractual clauses, binding internal rules, codes of conduct and accreditation certificates.

The clause must therefore name, in writing: the importing legal entity, every country where the data will be processed or stored, the transfer mechanism you are relying on (Article 23(1)(a) contractual safeguards, plus SCCs where the vendor offers them), and the same obligations on onward transfers to sub-processors. Note that a vendor's EU or UK transfer module does nothing for your PDPL position: OpenAI's DPA section 4 covers EEA, Swiss and UK data specifically, not UAE data. You must build your own PDPL route and document the risk assessment for each destination.

4. Sub-processors: list, notice, objection and flow-down

Under Article 8(10), every link in the chain needs a written contract defining roles, or everyone is jointly responsible. The DIFC guidance is blunt: a processor may not engage a sub-processor without prior written authorisation, and the processor remains "fully liable" for the sub-processor's failings.

What good looks like is OpenAI's clause 2.9: general authorisation for a published sub-processor list, notice of changes by blog post or email, a 30-day objection window, commercially reasonable alternatives, and termination with a refund if the objection is unresolved. Its sub-processor list names each entity, the product it touches, the processing location and the purpose: Microsoft for cloud infrastructure (including the UAE), TaskUs in the Philippines for moderation-related support, Intercom and Salesforce for customer support, and so on. Demand that level of detail as an annex, subscribe to change notices, and assign a named owner for the objection clock.

5. Retention, return and erasure

Articles 8(3) and 8(4) are your anchors: processing is for a specified purpose and period, then the data is erased or handed over. A clause that says deletion happens "in accordance with our retention policy" is not a term; it is a link the vendor can edit.

Ask for: a stated retention period for prompts and outputs during the term (zero or 30 days is common for enterprise APIs), a deletion deadline after expiry, a return option in machine-readable form, and honest treatment of backups. Anthropic's DPA commits to deletion within 30 days of termination, subject to narrow exceptions for law, disputes and harmful use. The DIFC guidance accepts that backup data may survive briefly if it is immediately put beyond use and destroyed on the next deletion cycle; get that stated in writing.

Anthropic's API and data retention documentation is a good reference point: zero data retention (ZDR) is available per organisation on request, but models it designates as "Covered Models" require 30-day retention and cannot be used under ZDR. Check which tier you are actually buying.

6. Security measures (Article 20)

Article 20 requires encryption, pseudonymisation where appropriate, resilience, restore capability and regular testing of controls. Your DPA should incorporate a technical and organisational measures (TOMs) annex that you can read, not a promise to "use industry-standard security". Ask for the current ISO/IEC 27001 certificate and SOC 2 Type II report, confirm encryption in transit and at rest, check tenant isolation for retrieval or fine-tuning data, and require confidentiality commitments from anyone with access, matching DIFC clause 24(5)(b)(iii).

7. Breach notification (Article 9)

Article 9(3) requires a processor to notify you "as soon as it becomes aware" of a breach. You then carry the clock: PwC's handbook states that controllers must notify the UAE Data Office within 72 hours, and notify affected data subjects immediately where their privacy or confidentiality is prejudiced. Article 9(1) also lists what the notification must contain: nature and form of the breach, approximate number of records, DPO details, expected effects and corrective measures.

"Without undue delay" from the vendor is not enough to protect a 72-hour clock. Negotiate "without undue delay and in any event within 24 hours of becoming aware", with content that maps to Article 9(1), plus assistance with your notification to the regulator.

8. Audit and evidence rights (Article 8(8))

Article 8(8) says the processor must provide all means to prove its compliance when you or the Bureau ask. In practice you will rarely get an on-site audit at a hyperscaler, so the contract must give you the next best thing: annual delivery of audit reports (SOC 2 Type II, ISO 27001, penetration test summaries), a right to ask questions and get written answers, and an expanded audit right triggered by an incident or a regulator's request. OpenAI's clause 2.8 offers exactly this shape: evidence once a year at most, with a summary of audit reports substituting for an inspection. Push for a named number of days to produce reports after request.

9. Data subject rights, automated decisions and DPIA support

You have 30 days to answer a data subject request under the PDPL regime described in the PwC handbook, and Articles 13 to 18 give individuals rights over access, correction, erasure and restriction. Article 18 is the AI clause that almost no SaaS DPA contains: individuals can object to decisions from automated processing that have legal or similarly significant effects, and you must provide human review on request. If the AI assistant only drafts replies, this is light; if it approves refunds or triages claims, it is central.

The contract must therefore promise: assistance with rights requests within a short fixed period (5 business days is a workable ask), no direct responses to your data subjects without your authorisation, support for your data protection impact assessment (Article 21 requires one for high-risk new technology before processing starts), and the information you need for your Article 8(7) record. That record should look something like this:

# Entry for your Article 8(7) processing record
vendor: "Example AI Platform Inc."
role: processor
contract: "MSA 2026-014 + DPA v3, effective 2026-01-01"
purpose: "AI customer support assistant for WhatsApp and web chat"
data_categories: ["name", "mobile", "email", "free-text conversation content"]
data_subjects: ["customers", "staff mentioned in conversation"]
processing_locations: ["UAE", "United States", "Philippines"]
transfer_mechanism: "PDPL Art. 23(1)(a) contractual safeguards + SCCs where applicable"
sub_processors:
  - {name: "cloud provider", location: "UAE", purpose: "inference hosting"}
  - {name: "moderation contractor", location: "Philippines", purpose: "content moderation"}
retention: {prompts_outputs_days: 30, deletion_after_termination_days: 30}
security: ["ISO/IEC 27001:2022", "SOC 2 Type II (2025)"]
breach_notice_hours: 24
Enter fullscreen mode Exit fullscreen mode

What the major AI vendors already give you

Use this as a baseline when you review their paper. It is not a substitute for reading it.

Clause OpenAI DPA (effective 1 January 2026) Anthropic DPA and Commercial Terms (2025) What to do
Training on your content Not used for training by default on API and Enterprise data, opt-in only Commercial Terms section B: "may not train models on Customer Content" Still write the prohibition into your contract
Deletion after you leave Return or delete at your instruction after the term (clause 2.11) Delete within 30 days of termination, with narrow exceptions (part H) Add a deadline and a certificate of deletion
Retention during the term ZDR available on some API paths (see sub-processor list footnote) ZDR per organisation; designated models require 30-day retention Confirm the tier and the model before you rely on ZDR
Sub-processor changes Published list, notice, 30-day objection (clause 2.9) List-based authorisation Subscribe to notices; keep an objection owner
Contracting entity for a UAE customer OpenAI OpCo, LLC (non-EEA) Anthropic, PBC (non-EEA) Check it matches the entity in your order form

Worked example: an AI support assistant in a Dubai services firm

Scenario: a Dubai services company with 250 staff runs 50,000 customer conversations a month through WhatsApp Business and web chat. It is buying an enterprise LLM API tier for AED 220,000 a year (hypothetical figures for this example). The conversations contain names, mobile numbers, email addresses and free text that occasionally mentions health or family matters, so sensitive personal data may appear.

  1. Map the flows first. The vendor's sub-processor list shows three entities touching content: the cloud host (processing in the UAE among other regions), a moderation contractor in the Philippines and a support tool in the United States. That is two Article 23 transfers before anyone writes a prompt.
  2. Fix the scope schedule. The vendor's Schedule 1 says "unstructured data" and "continuous". Ask for the four data categories above, the two data subject groups and an explicit note about possible sensitive data. Then paste the result into the YAML record above.
  3. Lock the training rule. The platform docs say no training by default for the API. Put the prohibition in the contract anyway, disable any opt-in toggles in the admin console, and record who holds the admin roles.
  4. Close the transfer gap. No UAE whitelist exists, so execute the Article 23(1)(a) contractual safeguards with the importing entity, attach SCCs where the vendor offers them, and run a written transfer risk assessment for the Philippines and US processing, as PwC's handbook recommends.
  5. Set the sub-processor clock. Subscribe to change notices. Assign the legal lead as the owner of the 30-day objection window and diary it. Request email notice rather than blog-only notice.
  6. Fix retention. Negotiate 30-day retention for prompts and outputs, deletion within 30 days after termination and a written statement on backups. If the compliance team needs zero retention, check whether the models in scope are ZDR-eligible before you commit.
  7. Collect security evidence. ISO/IEC 27001 certificate, current SOC 2 Type II report and the TOMs annex before the pilot moves to production.
  8. Cap the breach notice at 24 hours. Your own clock to the UAE Data Office is 72 hours from awareness, and you need time to assess records, effects and corrective action for the Article 9(1) report.
  9. Fit the rights clock. Require DSR assistance within 5 business days, because you have 30 days to answer. If the assistant is ever used to decide service outcomes rather than draft text, add Article 18 human review language and complete a data protection impact assessment before go-live.

Total effort to negotiate the missing items: roughly two legal reviews and one security questionnaire. The cost of getting it wrong is a regulator asking why your processor's Philippines-based moderation queue held customer conversations with no documented transfer basis.

Red flags that should stop you signing

  • "We may use Customer Data to improve our services", with no default-off switch and no opt-in process.
  • No named sub-processor list, or "our affiliates and trusted partners" with no locations.
  • Deletion "in accordance with our retention policy", or a 12-month post-termination deletion window with no justification.
  • Breach notice "without undue delay" and nothing more.
  • A transfer clause that says only "we comply with applicable data protection laws", with no mechanism, entity or country named.
  • No audit rights and no commitment to provide third-party reports.
  • Silence on model training, human review and automated decisions: the three AI-specific gaps a generic DPA never fills.
  • The vendor treats the DPA as non-negotiable and will not discuss even a schedule amendment. As the DIFC guidance puts it, if a party resists basic data protection clauses, ask whether they can be trusted with the data at all.

What to do next

Before the next AI contract reaches your desk: build the processing record template (the YAML in clause 9), name one owner for sub-processor change notices, and decide your standard position on the four deal-breakers (training, transfers, retention, breach hours). Then review the DPAs you already have against these nine clauses, because the tools already in use are the real exposure.

If you want support, our UAE-based team does this work as part of AI strategy and readiness: we map how data flows through your AI use cases and tell you plainly where the contractual gaps are and what to demand. Where a vendor's terms will never close the gap, the answer is often to keep the data on your own platform instead, which is what our AI infrastructure work covers. Either route, the goal is the same: when the regulator or the customer asks what your AI vendor may do with their data, the answer is in a signed document, not in a footnote on someone else's website.


Originally published on Azrty.

Top comments (0)