Document Automation
Product features
By practice area
By company size
Contract Review
PRODUCT FEATURES
Resources
LearN
Case Studies
Pricing
Gavel Exec reviews and redlines contracts inside Microsoft Word, checked against your own playbook. Start with 25 free queries, no credit card required.
An AI data processing addendum is the set of terms a customer adds when a vendor's product sends customer content to an artificial intelligence (AI) model. The customer's core positions: no training without written opt-in, named model providers, defined retention, customer ownership of outputs, confidential prompts, AI incident notice and an indemnity for infringing outputs.
Last reviewed November 16, 2026. This playbook is general information for lawyers, not legal advice for a specific matter. Gavel publishes it and makes Gavel Exec.
AI features usually arrive under the vendor's existing paper. The DPA ties the vendor to the customer's instructions for personal data under Article 28 of the General Data Protection Regulation (GDPR), but says little about confidential business content, output ownership or the model provider two steps away. The DPA playbook covers the processor terms; this page covers the AI layer.
Common Paper's 2026 software as a service (SaaS) contract benchmark of 16,140 signed agreements found AI language in 53% of cloud service agreements and an AI-specific edit in about one in five. Lisa R. Lifshitz, in the American Bar Association (ABA) publication Business Law Today (October 2026), warns that "toggling off the user interface may not actually stop the vendor from using your data for training."
The addendum should prevail over the vendor's online AI terms and acceptable use policy. Where it overlaps the DPA, the term more protective of the customer applies.
Positions are written for the customer, which supplies the prompts and documents. The vendor sells the product. The model provider runs the large language model (LLM) the vendor calls, which may be the vendor itself or a third party such as OpenAI or Anthropic.
A vendor should expect to settle near the fallback column. Upstream, the vendor is the customer and should run this playbook against its model provider, since GDPR Article 28(4) leaves a processor to "remain fully liable to the controller" for its subprocessors.
The positions assume a mid-sized customer whose documents go to a third-party model. Tighten them for privileged, health or financial data.
| Issue | Preferred position | Acceptable fallback | Escalate or walk away if | Why it matters |
|---|---|---|---|---|
| Scope and precedence | Covers all AI features, current and future, and all Customer Content; prevails over online AI terms and the acceptable use policy; where it overlaps the DPA, the term more protective of Customer applies. | Covers listed features; new AI features stay off until Customer turns them on. | AI features on by default, or AI terms changed by web posting. | Turning a feature off in the interface may not stop training use (Lifshitz). |
| Training and fine-tuning | No training, fine-tuning or distillation on Customer Content in any form, including de-identified or aggregated; exceptions only by signed opt-in naming dataset, model and purpose. | Content-free telemetry (counts, latency, error codes) may be used to run and improve the service. | Opt-out model, or training on de-identified content. | Common Paper AI Addendum Section 1.3 default; bans in 14% of cloud agreements in 2026. |
| Feedback and ratings | Ratings and corrections are Customer Content under the same ban. | Ratings kept apart from the text they rate. | Perpetual license to anything a user submits. | A feedback license can reach the prompt and output being rated. |
| Model providers | Schedule A names each provider, model, region and retention setting; 30 days' notice of changes; restrictions flowed down; termination with pro rata refund. | 15 days' notice of a new provider; Customer may disable the feature. | Vendor will not name providers, or provider terms allow training. | GDPR Article 28(2) and 28(4). |
| Retention and zero data retention | Zero data retention (ZDR) with each provider, defined in the addendum, exceptions listed with maximum periods. | Provider keeps content up to 30 days for abuse monitoring, no human review except a flagged event. | Indefinite retention, or retention to improve the services. | OpenAI logs up to 30 days by default; Anthropic may keep flagged content up to 2 years under ZDR. |
| Embeddings and fine-tuned weights | Deleted within 30 days after termination, with certificate. | Within 90 days, unused meanwhile. | Vendor keeps model improvements built from Customer Content. | GDPR Article 28(3)(g); derived artifacts can carry the content. |
| Ownership of inputs and outputs | Customer owns both; Vendor assigns any rights in outputs; Vendor license limited to serving Customer. | Same, with no warranty that outputs are unique or protectable. | Ownership conditioned on acceptable use compliance, or a vendor reuse license. | Common Paper Section 2.1; Copyright Office report of January 29, 2025. |
| Confidentiality of prompts | Prompts, files, outputs and embeddings are Confidential Information; staff access only on Customer request or a documented security investigation, logged. | Abuse review by named trust and safety staff, logs on request. | Prompts treated as usage data. | Privilege; ABA Formal Opinion 512. |
| Accuracy and human review | Each feature performs materially as documented; documentation states intended uses and limitations; human step for significant decisions. | Duty to fix reproducible failures within 30 days or allow termination of the feature. | Final decisions about individuals with no human step. | GDPR Article 22(1); Common Paper Section 3.1. |
| Model changes | 30 days' notice of a new provider, major version or region, with release notes. | Version updates from a listed provider at the time of change. | Model swaps with no notice. | Lifshitz: require advance written notice of model changes. |
| Audit and transparency | Annual documentation pack (providers, regions, retention, evaluations), System and Organization Controls (SOC) 2 Type II report, questionnaire answered in 30 days. | Pack and SOC 2 report; one questionnaire a year. | Refusal to disclose providers, regions or retention. | GDPR Article 28(3)(h); 11 California Code of Regulations (CCR) 7051(a)(7). |
| AI incident notice | 48 hours after awareness of disclosure through AI features or training misuse. | 72 hours; personal data breaches stay on the DPA clock. | Notice only for personal data, or tied to the vendor's harm view. | GDPR Article 4(12); Open Worldwide Application Security Project (OWASP) LLM01, LLM02, LLM07, LLM08. |
| Output intellectual property (IP) indemnity | Vendor defends claims that outputs used as documented infringe; narrow exclusions; separate higher cap. | Conditioned on filters staying on; capped at greater of 2 times annual fees and a fixed sum. | No output coverage, or only a model provider pass-through. | Standard IP indemnity often excludes outputs (Lifshitz). |
| Regulatory cooperation | Written classification of each feature under AI Act Annex III; instructions for use; logs kept six months; help with assessments; automated decisionmaking technology (ADMT) opt-outs honored. | Classification and reasonable cooperation at agreed rates. | Feature sold for hiring or credit with no classification. | AI Act Articles 25 and 26; California ADMT rules from January 1, 2027; 11 CCR 7051(a)(6). |
The Common Paper AI Addendum, version 1.0, Section 1.3, states: "Provider may not use Customer's Inputs or Outputs to Train any Model," unless the cover page names training data and purposes.
The risk is scope. A ban on "Customer Data" may miss ratings, user corrections, embeddings in a shared index and outputs, and a DPA clause reaches personal data only. Lifshitz calls de-identification offered as a compromise "far less protective than it appears."
Preferred language:
Vendor will not use Customer Content to train, fine-tune, distill or otherwise improve any AI model, or add Customer Content or embeddings derived from it to any index or dataset available to anyone other than Customer. "Customer Content" means all inputs, prompts, files, outputs, feedback, ratings and corrections submitted to or generated by the Services for Customer, in any form, including after de-identification or aggregation. Any exception requires Customer's signed opt-in naming the dataset, model and purpose, revocable on thirty (30) days' notice.
Fallback language:
Vendor may use operational telemetry limited to counts, latency, error codes and feature selections, containing no text, files, outputs or embeddings, to operate and improve the Services. Any other training use requires Customer's written opt-in.
What vendors push back with: ratings improve quality. Offer ratings separated from the text they rate.
The model provider's terms bind the vendor and give the customer no direct rights. OpenAI's application programming interface (API) data documentation says API data "is not used to train or improve OpenAI models" unless the customer opts in, but abuse monitoring logs are "retained for up to 30 days." Zero data retention (ZDR) removes content from those logs and is "subject to prior approval by OpenAI"; file, vector store and fine-tuning endpoints are listed as ineligible.
Under ZDR, Anthropic's API documentation says it "does not store customer prompts or responses at rest after the API response is returned," yet for flagged content it "may retain inputs and outputs for up to 2 years." A vendor promising zero retention without those exceptions is promising more than it has. Under GDPR Article 28(2) a model provider that receives personal data is a subprocessor; list the model, region and retention setting too, because they decide where prompts go and how long they stay.
Preferred language:
Schedule A lists each Model Provider, model, hosting region and retention setting for Customer Content. Vendor will maintain zero data retention with each Model Provider, meaning no storage of Customer Content after a response is returned, except content flagged by automated safety systems, content held under legal compulsion, and files or indexes Customer chooses to store for a feature, each of which Schedule A will describe with its maximum retention period. Vendor will give thirty (30) days' written notice before adding a Model Provider or changing a model, region or retention setting, and Customer may terminate the affected feature for a pro rata refund.
Fallback language:
A Model Provider may retain Customer Content for up to thirty (30) days for abuse monitoring, with no training use and no human review except to investigate a specific flagged event. Vendor will give fifteen (15) days' notice before routing Customer Content to a new Model Provider.
What vendors push back with: model releases are unpredictable. Accept version updates from a listed provider at the time of change, and keep notice for new providers, regions and retention settings.
Common Paper's Section 2.1 says the customer "owns all Output" and the provider assigns its rights, "if any." Those two words matter: the U.S. Copyright Office's January 29, 2025 report concluded that "prompts alone do not provide sufficient human control" (Sidley summary). Lifshitz adds that ownership "is frequently conditioned on ongoing compliance with the vendor's acceptable use policy," which turns a policy breach into a title defect.
Lifshitz also notes that "Almost all standard AI agreements state that outputs are made available on an 'as-is,' 'as-available' basis." The workable ask is a documentation warranty with known limitations disclosed. Where outputs feed decisions about people, GDPR Article 22(1) gives a right "not to be subject to a decision based solely on automated processing," so keep a human step the customer controls.
Preferred language:
As between the parties, Customer owns all Inputs and Outputs, and Vendor assigns to Customer any right it has in Outputs. Vendor may use Inputs and Outputs only to provide the Services to Customer. Customer's rights in Outputs do not depend on compliance with any acceptable use policy. Vendor warrants that each AI feature performs materially as described in its documentation, which states its intended uses and known limitations.
Fallback language:
Vendor assigns to Customer any right it has in Outputs, without warranty that Outputs are unique or protectable. If Vendor does not correct a reproducible failure of an AI feature to perform as documented within thirty (30) days, Customer may terminate that feature for a pro rata refund.
What vendors push back with: model behavior varies between runs. Tie the warranty to documented uses and reproducible failures, which the vendor can test.
Lifshitz notes that "The vendor's standard intellectual property indemnity often does not extend to AI-generated outputs."
OpenAI's service terms exclude output that "was modified, transformed, or used in combination with products or services not provided by or on behalf of OpenAI," and under Microsoft's Customer Copyright Commitment the customer "must have implemented all mitigations required." A vendor's product combines the model with its own software, so the combination exclusion may reach many claims. Ask the vendor for its own output indemnity.
Preferred language:
Vendor will defend Customer against any third-party claim that an Output, or Customer's use of it in accordance with the documentation, infringes or misappropriates intellectual property rights, and will pay resulting damages, settlements and costs, except to the extent the claim arises from Customer's Inputs, Customer's modification of the Output, or Customer's use of an Output it knew to infringe. This Section is subject to the increased cap in Section [X] in place of the general cap.
Fallback language:
The output indemnity applies only where Customer kept Vendor's content filters enabled and is capped at the greater of two (2) times the fees paid in the prior twelve (12) months and [$1,000,000].
What vendors push back with: an offer to pass through the provider's indemnity. Read its exclusions first.
Breach clauses written for databases miss how AI features leak. The OWASP Top 10 for LLM Applications 2025 lists Prompt Injection (LLM01), Sensitive Information Disclosure (LLM02), System Prompt Leakage (LLM07) and Vector and Embedding Weaknesses (LLM08). An injected prompt that reveals another customer's file is an incident whether or not personal data is involved.
GDPR Article 4(12) covers "unauthorised disclosure of, or access to, personal data," and the DPA clock handles that. Define an AI Incident for the rest: disclosure through prompt injection, system prompt exposure, a shared index or a model provider event, plus training or retention contrary to the addendum. The notice should say whether any content entered a model's training data. Vendors may fairly exclude unsuccessful injection attempts; refuse any trigger based on the vendor's own view of harm.
The AI Act applies to deployers "located within the Union" and to providers and deployers elsewhere "where the output produced by the AI system is used in the Union" (Article 2(1)). The customer is usually the deployer and the vendor the provider. Regulation (EU) 2026/1744, the Digital Omnibus on AI, dated July 8, 2026, entered into force July 27, 2026. Under amended Article 113:
Article 4 now requires measures "to support the development of AI literacy" of staff. Article 26 requires high-risk deployers (for Annex III systems, from December 2, 2027) to assign oversight to people with "the necessary competence, training and authority" and to keep logs under their control for a period "of at least six months." Under Article 25(1), a deployer that changes a system's intended purpose so it becomes high-risk is treated as a provider, so state each feature's intended purpose.
Article 28(10) treats a processor that sets its own purposes as a controller. The European Data Protection Board's Opinion 28/2024 says whether a model trained on personal data is anonymous "should be assessed on a case by case basis" (Hunton summary), so a vendor's statement that its model holds no personal data needs support.
Under the California Consumer Privacy Act (CCPA) regulations, 11 CCR 7050(a)(3) allows use "to build or improve the quality of the services" if the provider does not use the data to serve another person, and 7051(a)(5) requires the contract to bar use "outside the direct business relationship." A model improved on one customer's data and sold to all sits close to that line. The approved ADMT regulations apply to significant decisions from January 1, 2027 (section 7200(b)), require later opt-outs to be passed to service providers (section 7221(n)(2)) and require service providers to help (section 7051(a)(6)).
Customer size matters less than the content sent and the decisions outputs feed.
| Buyer or vendor type | What changes | Why |
|---|---|---|
| Vendor that trains its own model | The ban must reach the vendor's base models and evaluation sets, and the IP indemnity should cover the model as well as outputs. | AI Act Article 53(1)(c) and (d): providers of general-purpose AI models need a copyright policy and a public training content summary. |
| Vendor that wraps a third-party model | The negotiation moves upstream: ask for Schedule A, ZDR approval and the provider's indemnity exclusions before marking up. | OpenAI ZDR is "subject to prior approval"; its indemnity excludes combined output. |
| Law firm or legal department | No vendor human review of prompts unless requested, a privilege non-waiver clause, and client consent before matter content enters a self-learning tool. | ABA Formal Opinion 512 (July 29, 2024), as summarized: client data may be used for self-learning. |
| Healthcare provider or health tech | Protected health information (PHI) also needs a Health Insurance Portability and Accountability Act (HIPAA) business associate agreement (BAA); training on PHI is limited to what the BAA permits, and model providers sign as subcontractors. | 45 C.F.R. 164.504(e)(2)(ii)(A) and (D). |
| Financial services and lending | Contractual safeguards and periodic vendor assessment; for credit scoring, reasons specific enough for adverse action notices. | 16 C.F.R. 314.4(f); Regulation B, 12 C.F.R. 1002.9(b)(2); AI Act Annex III point 5(b). |
| Hiring tools | For New York City roles, vendor supports a bias audit within one year before use and candidate notices. | New York City Local Law 144; AI Act Annex III point 4, from December 2, 2027. |
| Customer in the EU, or output used there | Add AI Act cooperation: classification, instructions for use, log access and AI literacy materials. | AI Act Article 2(1), Article 4 and Article 26. |
Gavel Exec is AI contract review and drafting software that works in Microsoft Word. It generates playbook rules from a checklist, a spreadsheet playbook or signed agreements (.docx, PDF or .csv), so the fourteen rows above can be uploaded as a starting rule set. See how playbooks in Gavel Exec are built and shared.
With a vendor's AI terms open in Word, Gavel Exec applies the rules as priority-ranked tracked changes with reasoning for each, and lists missing terms. In the web app, benchmarking labels each provision on-market, off-market or missing. Gavel states that it has Zero Data Retention agreements with the AI model providers it uses, including OpenAI, and that customer data, prompts and document content are not used to train AI models. Pricing is $160 per user per month or $1,740 per user per year, with 25 free queries per user and no credit card.
What is an AI data processing addendum?
It is a contract addendum that sets the terms for a vendor's AI features: training on customer content, model providers, retention, output ownership and AI incidents. It sits alongside the data processing agreement, which governs personal data.
Do I need an AI addendum if the vendor already signed a DPA?
Usually yes, if the product sends your content to an AI model. A DPA covers personal data, while confidential business documents, outputs and model provider retention often fall outside it.
What should a no-training clause cover?
It should bar training, fine-tuning and other model improvement on all customer content, including prompts, files, outputs, feedback and embeddings, in de-identified or aggregated form as well. Any exception should be a signed opt-in naming the dataset, model and purpose.
Who owns AI outputs under a vendor contract?
The contract decides as between the parties, and the customer position is that the customer owns inputs and outputs. Copyright may not protect purely AI-generated material, because the U.S. Copyright Office concluded in January 2025 that prompts alone do not give sufficient human control.
Does the EU AI Act apply to companies that buy AI tools?
It can. It applies to deployers in the EU and to deployers elsewhere when the output is used in the EU, and its AI literacy provision, narrowed in July 2026, has applied since February 2, 2025. High-risk duties for Annex III uses such as hiring and credit scoring apply from December 2, 2027.
What does zero data retention mean in an AI vendor contract?
It usually means the model provider does not store prompts or outputs after returning a response. Provider terms carry exceptions, such as flagged content, so the addendum should list each exception with a maximum retention period.
Related reading: the DPA playbook for controllers and processors, how Dorna Moini uses AI to review data protection agreements, and all Gavel contract playbooks.
AI contract review and redlining in Microsoft Word, checked against your own playbook, with the reasoning behind each change.