Commercial Analysis Report: Contract Clauses for AI Vendor and SaaS Agreements
Before signing, a venture backed startup should treat the indemnification clause and the data use clause as the two provisions that decide whether an artificial intelligence vendor relationship transfers risk onto the company or keeps that risk with the party that built and trained the model. Vendor paper almost always caps the vendor’s own liability at the fees the startup pays, commonly at twelve months of fees under current market practice, while leaving a gap between what the vendor’s own upstream model provider will cover and what the vendor promises the startup in return. That gap lands on the startup by default unless the company negotiates it closed before signature. Every dollar of exposure inside that gap, and every right the vendor claims to reuse the startup’s own customer data, prompts, and outputs to train or improve its models, is negotiable, and most vendors concede the point once a company actually redlines the draft rather than accepting it as presented.
A. Executive Summary
A startup that accepts an AI vendor’s standard indemnification and data use terms without negotiation absorbs liability that the Federal Trade Commission’s September 2024 enforcement actions show regulators now target. The commercial question facing a venture backed company’s founders and general counsel before signing an artificial intelligence vendor or software as a service agreement is not whether to sign, since the company usually has no realistic alternative once the product decision is made, but which of a short list of clauses actually shifts risk back onto the vendor rather than leaving it with the company that bought the tool. Two provisions carry most of the exposure. The indemnification clause decides who pays when a rights holder, a regulator, or a customer sues over something the AI system produced or leaked. The data use clause decides whether the vendor can retain, review, and train future models on the company’s own customer data, source code, and trade secrets once the relationship ends.
Market data collected across thousands of software as a service contracts in 2026 puts ninety six percent of agreements at a liability cap equal to one times the fees paid, records unlimited liability carve outs rising from about one percent of contracts in 2025 to over ten percent in 2026, and records clauses barring the vendor from training on customer data rising from under one percent of contracts in 2024 to fourteen percent in 2026. What remains unsettled is how far upstream indemnities from foundation model providers such as OpenAI and Anthropic actually flow through a reseller’s own contract, and how the European Union’s Artificial Intelligence Act deployer obligations, which take effect December 2, 2027 for the first category of regulated systems, will layer onto data processing terms already in place.
B. Detailed Findings
Intellectual property posture: data use rights, training restrictions, and output ownership
The single highest value clause in an AI vendor contract states, in plain terms, whether the vendor may use the startup’s own customer data, prompts, and generated outputs to train or improve any model, including a model the vendor later sells to other customers or to competitors of the startup. Left silent, a vendor’s general terms of use typically control by default, and as of 2026 those default terms still grant broad training rights in roughly three percent of contracts while an express prohibition on that use now appears in fourteen percent of contracts, up from under one percent in 2024. A company negotiating this clause should ask for six specific protections: an express opt out of any training use of its data, segregation of its data from other customers’ data, de-identification standards paired with a prohibition on re-identification, confirmation that the company owns its own inputs, a license rather than an assignment of outputs sufficient for the company’s lawful business use, and survival of all of these restrictions through any assignment or change of control of the vendor.
In redlining these agreements for venture backed clients since generative AI vendors became a standard line item in every technology stack, the training opt out is consistently the fastest concession in the entire markup. Most vendors return it within a single exchange, because the clause costs the vendor nothing until the vendor decides to build a model business on the enterprise data already sitting in its logs. The harder ask, and the one worth the extra week of negotiation, is survival through a change of control, since a small vendor with no model ambitions today can be acquired tomorrow by a company that wants exactly the training rights the original vendor gave up.
Confidential information and trade secret provisions deserve the same scrutiny. Once a company types its own proprietary process or algorithm into a prompt and sends it to a third party’s model, the secrecy that trade secret status depends on can be compromised unless the vendor’s confidentiality terms are drafted broadly enough to cover model interaction data and expressly bar the vendor from using that data outside the specific engagement.
Indemnification structure and where the coverage gap lands
Indemnification in an AI vendor contract operates across three layers, and the coverage gap that most often lands on a startup sits between the first two: the model provider’s promise to the vendor, and the vendor’s promise to the startup. OpenAI’s chief executive announced a copyright indemnity, branded Copyright Shield, at the company’s developer conference in October 2023, covering ChatGPT Enterprise and API customers against third party intellectual property infringement claims arising from authorized use, while excluding beta services. Microsoft made a parallel commitment on September 7, 2023, covering paying customers of its Copilot products, including Microsoft 365 Copilot, against copyright claims tied to the output those products generate. Anthropic separately defends authorized customers against copyright infringement claims and covers approved settlements and judgments. Each of these upstream indemnities is limited strictly to intellectual property infringement arising from compliant, authorized use.
The vendor’s downstream promise to its own customer, the startup, often sounds broader than what the vendor actually receives from its upstream provider. The common drafting failure is a vendor granting its customer an apparently uncapped indemnity for any claim arising from the AI service or its output, while the vendor’s own upstream coverage is limited to intellectual property infringement alone. Claims for inaccurate output, defamation, a violation of the vendor’s acceptable use policy, or a beta feature malfunction fall into the gap between those two promises and default to the vendor. If the vendor cannot pay, because it is thinly capitalized or the claim exceeds its own liability cap, the startup’s customer has only the startup left to pursue, even though the startup never trained the model or wrote the underlying code.
A company negotiating this clause should confirm that its own downstream indemnity to its customers is no broader than the coverage it receives from its vendor, and should set the indemnification cap independently of, and higher than, the general limitation of liability, since a defined increased claim allowance is the only mechanism that protects against a claim large enough to matter. By 2026, close to seven percent of software as a service contracts included some form of increased claim allowance above the general liability cap, up from about four percent the year before, and privacy or security breach carve outs specifically appeared in nearly four percent of agreements.
Regulatory exposure shaping the contract
Three regulatory regimes independent of the contract itself dictate terms a startup should insist on regardless of what the vendor offers as a default: the Federal Trade Commission’s enforcement authority over deceptive claims, the European Union’s Artificial Intelligence Act obligations for deployers of high risk systems, and state consumer privacy statutes governing what a service provider may do with personal information.
On September 25, 2024, the Federal Trade Commission announced five coordinated enforcement actions under its unfair and deceptive practices authority against companies that overstated what their AI products could do, including a $193,000 settlement against a company marketing an AI product as a substitute for a lawyer and federal court actions alleging tens of millions of dollars in consumer losses tied to unsubstantiated AI earnings claims. A startup that repeats a vendor’s overstated capability claim to its own customers inherits the same exposure the Commission used against direct sellers in 2024. The contract should include a specific representation that the vendor’s performance claims are substantiated, paired with a right to terminate and a refund if the system does not perform as marketed.
For a startup selling into the European Union, Article 26 of the European Union’s Artificial Intelligence Act imposes direct deployer obligations effective December 2, 2027 for systems listed in the Act’s Annex III and August 2, 2028 for certain other regulated systems. Those obligations include using the system only according to the provider’s instructions, assigning human oversight to personnel with the competence and authority to exercise it, monitoring operation and reporting serious incidents to the provider without undue delay, and retaining system logs for at least six months. Every one of those duties depends on information only the vendor can supply, so the contract needs a clause requiring the vendor to deliver current instructions for use, incident notice, and log access sufficient to satisfy Article 26 before the startup deploys the system in the European market.
Where the vendor processes the startup’s customer data as a service provider or contractor under the California Consumer Privacy Act as amended by the California Privacy Rights Act, codified beginning at California Civil Code section 1798.140, the contract must prohibit the vendor from retaining, using, or disclosing that data for any purpose outside the specific business purpose stated in the contract, must define that business purpose narrowly enough to exclude model training by default, and must give the startup an audit right to confirm compliance. A data processing addendum negotiated before an AI feature existed rarely anticipated that the vendor’s newest model update would want to learn from the same data.
Risk factors the company must manage after signature
Four risks recur across AI vendor and software as a service agreements once the indemnification and data clauses are settled. A liability cap set too low relative to the company’s actual exposure leaves the company self insuring a risk it thought it had transferred. An indemnity gap between the vendor’s upstream coverage and its downstream promise leaves specific categories of claim, particularly accuracy and defamation claims, with no source of payment at all. A data use right that outlives the contract term lets a vendor keep training on the company’s data long after the relationship ends unless the contract expressly requires deletion and certification of deletion at termination. Dependence on a vendor’s own subcontracted foundation model provider means the company’s data flows through a party it never signed a contract with directly, so the contract should require the vendor to disclose which underlying model providers it uses and to flow down the same data use and deletion protections the company negotiated with the vendor itself.
Want more industry insights? I publish daily analysis on the intersection of deep tech and venture law.
C. Open Questions
1. Whether a reseller’s own indemnity, once negotiated down from the model provider’s original scope, actually protects the startup for claims the model provider itself would have covered directly, since resellers rarely disclose the full text of their upstream agreement with a foundation model provider. This matters because a startup’s general counsel cannot verify the coverage it is relying on without seeing a document the vendor is unlikely to produce.
2. How the European Union’s Artificial Intelligence Act deployer obligations under Article 26, effective December 2, 2027, will be enforced against a United States based startup that sells into the European market through a vendor rather than directly, since the Act’s incident reporting and log retention duties assume the deployer controls operational data the startup may not actually hold.
3. Whether liability caps will keep expanding, given that unlimited liability carve outs rose from about one percent of contracts in 2025 to over ten percent in 2026, or whether that trend reverses once technology errors and omissions insurers begin pricing AI specific claims and push vendors back toward capped exposure.
4. Whether a training restriction and data deletion obligation negotiated with a small AI vendor survives an acquisition of that vendor by a company with a foundation model business, since the acquiring company is the buyer most likely to want the training rights the original vendor gave up.
5. How state privacy statutes built around the service provider and contractor categories, including the California Consumer Privacy Act as amended, will treat a vendor that fine tunes a shared model on data collected from many customers at once, where no single customer’s data is separately identifiable in the output but the training input came from identifiable customer records.
6. Whether the Federal Trade Commission’s 2024 enforcement pattern against direct to consumer AI marketing claims will extend to business to business vendor sales representations that a startup relied on in deciding to sign, which would give startups a stronger basis to walk away from a vendor’s unsubstantiated performance promises after the fact.
D. Source List
1. Federal Trade Commission. Press release announcing five enforcement actions under the FTC’s unfair and deceptive practices authority against companies overstating AI product capabilities, September 25, 2024. Primary regulatory source for the accuracy representation analysis.
2. European Union, Artificial Intelligence Act, Article 26, obligations of deployers of high risk AI systems. Primary statutory source for deployer obligations and effective dates used in the regulatory exposure analysis.
3. California Civil Code section 1798.140 and following, the California Consumer Privacy Act as amended by the California Privacy Rights Act. Primary statutory basis for service provider and contractor data use restrictions, verified through secondary commentary.
4. A 2026 benchmark study of thousands of software as a service contracts. Market data on liability caps, indemnification carve outs, and AI training use restrictions, used to calibrate negotiating positions in the market benchmarks analysis.
5. A law firm client alert detailing OpenAI’s and Microsoft’s copyright indemnity announcements, published November 2023. Used for the three layer indemnification analysis.
6. A law firm client alert describing the three layer indemnity chain among model provider, vendor, and customer and the coverage gap problem, published 2026. Used as the core structure for the indemnification analysis.
7. A law firm client alert recommending data use, training restriction, and ownership clause language for AI vendor agreements. Used for the intellectual property posture analysis.
8. A law firm client alert surveying standard AI vendor clause categories. Used to cross reference the clause priority list.
9. An institutional startup legal playbook addressing generative AI contract terms. Used to confirm standard market practice on ownership and licensing provisions.
10. An institutional startup legal playbook on commercial contract fundamentals for startups. Used for general negotiation posture guidance.
11. A commercial legal technology publication surveying which SaaS clauses carry the most risk. Used to prioritize the clause list in the detailed findings.
E. Bibliography
Andrew S. Bosin LLC. “Indemnification for AI Generated Outputs: Who Pays? (2026).” Accessed September 2026.
California Civil Code. Section 1798.140 et seq., as amended by the California Privacy Rights Act.
Common Paper. “2026 SaaS Contract Benchmark Report.” 2026.
Cooley LLP. “Generative AI: Important Considerations When Navigating Contract Terms.” Cooley GO. Accessed September 2026.
Cooley LLP. “Six Things Startups Need to Know About Commercial Contracts.” Cooley GO. Accessed September 2026.
European Union. Artificial Intelligence Act, Article 26, “Obligations of Deployers of High Risk AI Systems.” 2024.
Federal Trade Commission. “FTC Announces Crackdown on Deceptive AI Claims and Schemes.” Press release. September 25, 2024.
GC AI. “A SaaS Agreement Has up to 50 Clauses. Six Carry the Risk.” Accessed September 2026.
Gouchev Law. “10 Critical Clauses for AI Vendor Contracts.” Accessed September 2026.
Proskauer Rose LLP. “OpenAI’s ‘Copyright Shield’ Broadens User IP Indemnities for AI Created Content.” New Media and Technology Law Blog. November 10, 2023.
Ward and Smith, P.A. “Contract Architecture: Core AI Clauses for Vendor Agreements.” Accessed September 2026.
Interested in analysis about the intersection of venture capital, tech, artificial intelligence, public policy and the law? Check out my Substack channel. https://theinnovationattorney.substack.com/