Most of the beta agreement discussion on this site looks at the startup’s side of the table: the startup runs the beta, and customers test the product.
But AI startups are usually testers too.
A founder gets early access to a new model from a foundation model provider, or a “preview” feature in a cloud AI platform. The new model is faster, cheaper, or better at the one task the product depends on. The team builds on it, ships it to customers, and moves on.
What nobody reads closely is the preview terms that came with it.
SaaS AI Lawyer Andrew S. Bosin, a two-time SaaS founder, sees the same problem repeatedly: a startup’s customer contract promises uptime, data protection, and indemnity for a feature that sits on a preview model whose provider has promised none of those things.
This article explains what preview terms usually say, where they leave the startup exposed, and how to write the startup’s own customer terms so they don’t promise more than its provider does. For a comparison of the major providers’ general commercial terms, see Andrew’s guide to OpenAI vs. Anthropic API terms for SaaS founders.
Quick Answer
When an AI startup builds on a preview, beta, or early-access model, it should assume the provider’s terms:
- Disclaim all warranties and provide the model “as is”
- Exclude the preview from any SLA and often from technical support
- Exclude the preview from the provider’s indemnity, including IP indemnity
- Allow the provider to change or discontinue the preview, sometimes without notice
- Cap the provider’s liability at a low amount, or at what the startup paid
- May not apply the provider’s data processing terms, and sometimes say not to use personal data at all
- Treat the startup’s test observations as the provider’s confidential information, or give the provider ownership of them
The startup’s job is to make sure its own customer contracts, privacy commitments, and marketing don’t promise more than those terms support.
What the Major Providers’ Preview Terms Actually Say
Preview terms vary by provider and change over time, so always read the current version. But the published terms of the major cloud and model providers show how consistent the pattern is.
Google Cloud: Pre-GA Offerings
Google Cloud’s Service Specific Terms cover “Pre-GA Offerings,” which include preview AI features on its platform. They state that Pre-GA Offerings:
- Are provided “as is” without warranties
- “May be changed, suspended or discontinued at any time without prior notice”
- “Are not covered by any SLA or Google indemnity”
- Are generally not covered by technical support, and the data location commitments don’t apply
Two terms stand out for AI startups:
- Data processing: unless Google says otherwise in writing, no data processing terms (including its Cloud Data Processing Addendum) apply, and customers “should not use Pre-GA Offerings to process personal data or other data subject to legal or regulatory compliance requirements.”
- Liability: Google’s liability for Pre-GA Offerings is limited to the lesser of the agreement’s general cap or $25,000.
AWS: Betas and Previews
AWS’s Service Terms cover features labeled “beta,” “preview,” “pre-release,” or “experimental.” Under those terms:
- Service Level Agreements don’t apply
- AWS may suspend or terminate beta access at any time, and customer content may not carry over to the generally available version
- AWS owns and may use the customer’s Test Observations, meaning its feedback and performance information about the beta
- Test Observations and information about the beta are AWS Confidential Information
- AWS’s liability is limited to what the customer actually paid for the beta service in the preceding 12 months
OpenAI: Beta Services
OpenAI’s Service Terms state that Beta Services are offered “as-is” for testing and evaluation, are “excluded from any indemnification obligations OpenAI may have to you,” and carry no guarantee that they’ll become generally available or be uninterrupted or error-free.
The Pattern
The details differ, but the direction doesn’t. Preview terms move nearly all the risk of an unfinished model onto the customer, which here is the AI startup. That’s reasonable for a preview. It becomes a problem only when the startup passes the preview on to its own customers under terms written for a stable product.
Where the Gaps Show Up
1. The Indemnity Gap
Many startups rely on their provider’s IP indemnity to support the indemnity they give customers for AI outputs. If the underlying model is a preview, that support is usually gone.
The result is a startup that has promised customers protection against third-party infringement claims for outputs, with nothing behind it from upstream. Andrew’s guide to indemnification for AI-generated outputs covers how output indemnities are usually structured.
2. The SLA Gap
If the startup’s MSA promises 99.9% uptime across the whole service, and one feature depends on a preview model with no SLA, the startup is guaranteeing availability its provider won’t guarantee. When the preview has an outage, the startup owes service credits and the provider owes nothing.
3. The Data Processing Gap
This is the gap with the most regulatory weight.
Enterprise customers sign a DPA with the startup that lists the model provider as a subprocessor, on the assumption that the provider’s own data processing terms cover the data. If the preview is outside those terms, or the provider has said personal data shouldn’t be used in the preview at all, the startup’s DPA commitments may be untrue for any data that passes through that feature.
The startup should know, before a preview feature touches customer data:
- Whether the provider’s DPA covers the preview
- Whether the provider’s data retention and no-training commitments apply to it
- Whether the preview runs in the same regions as the generally available service
Andrew’s AI SaaS DPA guide explains what customers expect those subprocessor commitments to mean.
4. The Discontinuation Gap
Previews can be changed or shut down on short notice, or with none. Even generally available models get retired. Anthropic, for example, commits to at least 60 days’ notice before retiring publicly released models, a commitment that by its terms covers publicly released models, not necessarily previews.
A startup whose core feature depends on a preview has a single point of failure it doesn’t control. That matters for the product, for customer commitments, and for investor diligence.
5. The Confidentiality Gap
Early-access programs often come with confidentiality obligations about the preview itself: its existence, its capabilities, its benchmark performance, or the provider’s roadmap. Some treat the startup’s own test observations as the provider’s confidential information.
That can collide with the startup’s sales and marketing. A startup that announces “powered by [unreleased model]” or publishes benchmark comparisons during a preview may breach its early-access terms.
6. The Feedback and Test Observations Gap
When a provider owns or has free use of the startup’s test observations, the startup should think about what it’s actually sending:
- Detailed bug reports can reveal the startup’s product design and use cases
- Example prompts and outputs can reveal the startup’s system prompts or customer workflows
- Performance comparisons can reveal what the startup is building and for whom
Send what’s useful. Don’t send customer data or the startup’s own proprietary prompts as “examples” unless the terms and the customer’s contract allow it.
How to Write Your Own Customer Terms
The fix isn’t to avoid previews. Previews are often where a startup’s competitive advantage comes from. The fix is to label and contain them.
Label Preview Features Clearly
The startup’s customer terms should define “Preview Features” (features the startup designates as beta, preview, or early access, or that depend on a third-party model the provider has designated as pre-release) and treat them separately from the core service.
Carve Preview Features Out of the Standard Promises
For Preview Features, the customer terms should:
- Provide them “as is,” without warranties
- Exclude them from the SLA and from service credits
- Exclude them from the startup’s IP and output indemnity
- Allow the startup to change, suspend, or withdraw them, ideally with whatever notice the startup actually receives from its provider
- Apply a lower liability cap, or none beyond fees paid for the feature
Keep Customer Data Out Unless the Terms Allow It
If the provider says personal data shouldn’t be used in the preview, the startup’s customer terms should say the same thing, and the product should enforce it where possible. Consider making Preview Features opt-in for enterprise customers, so regulated data doesn’t flow through them by default.
Match the Notice You Give to the Notice You Get
Don’t promise customers 90 days’ notice before retiring a feature when the provider can retire the underlying preview tomorrow. Tie the startup’s notice commitment to the notice the startup receives.
Keep Marketing Consistent With the Contract
If the terms say “preview,” the website shouldn’t call the feature enterprise-grade or production-ready. The FTC has repeatedly said that AI claims have to be truthful and supported, and a feature the startup’s own provider calls experimental is hard to describe as anything else.
Sample Clauses
These are illustrative starting points, not ready-to-sign terms. The right language depends on the startup’s provider, product, and customers.
Preview Features: Customer Terms
“Preview Features” means any feature, model, or functionality of the Services that Company designates as “beta,” “preview,” “early access,” or “experimental,” or that relies on a third-party model or service designated by its provider as pre-release. Preview Features are provided “as is” and are excluded from the Service Level Agreement, from Company’s indemnification obligations, and from any warranty in this Agreement. Company may modify, suspend, or discontinue any Preview Feature at any time and will use reasonable efforts to give Customer notice at least as long as the notice Company receives from its provider.
Data Restriction for Preview Features
Customer will not submit personal data, protected health information, payment card data, or other data subject to specific legal or regulatory requirements to any Preview Feature unless Company has confirmed in writing that the Preview Feature is covered by the Data Processing Addendum.
Liability for Preview Features
Notwithstanding anything to the contrary in this Agreement, Company’s total liability arising out of or relating to Preview Features will not exceed the fees paid by Customer specifically for Preview Features during the twelve (12) months preceding the claim.
Before You Build on a Preview: Checklist
- Read the provider’s current preview, beta, or pre-GA terms, not just its general commercial terms
- Confirm whether the provider’s SLA, indemnity, support, and data processing terms apply to the preview
- Check whether the provider allows personal or regulated data in the preview
- Note the provider’s right to change or discontinue the preview, and any notice commitment
- Note confidentiality obligations about the preview, benchmarks, and test observations
- Decide what the startup will and won’t send as feedback
- Define Preview Features in the startup’s customer terms and carve them out of the SLA, indemnity, and warranties
- Tie the startup’s discontinuation notice to the notice it receives
- Make Preview Features opt-in for enterprise customers, or keep regulated data out
- Align marketing claims with the preview status
- Plan a fallback: a generally available model the product can switch to
How This Connects to the Rest of Your Contract Stack
Preview terms touch almost every customer-facing document the startup has:
- The SaaS MSA needs a Preview Features carve-out
- The AI terms of service need consistent beta language for self-serve users
- The AI warranty disclaimers need to cover preview-specific risk
- The startup’s own beta test agreement should reflect any preview models the beta depends on
If the startup is running its own evaluation or beta for customers on top of a preview model, the AI model evaluation beta agreement guide covers the customer-facing side of that testing.
Frequently Asked Questions
Can an AI startup use a preview model in production?
Sometimes, but it should read the provider’s terms first. Some providers say previews are for evaluation, or that personal data shouldn’t be used in them. Even where production use is allowed, the startup carries most of the risk, so its customer terms should reflect that.
Does my model provider’s IP indemnity cover preview models?
Usually not. Major providers commonly exclude beta and preview offerings from their indemnity. Check the specific terms, and don’t extend an output indemnity to customers for preview features unless the startup is willing to carry that risk alone.
What happens if the provider discontinues the preview my product depends on?
The provider may be able to do so with little or no notice, and without liability. The startup should have a fallback model, and its customer terms should allow it to change or withdraw the feature.
Can we tell customers we’re using an unreleased model?
Only if the early-access terms allow it. Many preview programs restrict disclosure of the preview, its capabilities, or its benchmarks until launch.
Do we own the feedback we give the provider?
Often not. Some providers own the startup’s test observations or treat them as the provider’s confidential information. Avoid putting customer data or proprietary prompts into feedback.
Should we just avoid preview models?
Not necessarily. Early access can be a real competitive advantage. The goal is to use previews deliberately: labeled, contained, and matched by customer terms that don’t promise more than the provider does.
Get Your Customer Terms Aligned With Your Model Stack
Andrew S. Bosin is a New Jersey technology attorney and two-time SaaS founder who represents AI and SaaS startups nationwide on a flat-fee basis. He reviews model provider and early-access terms and drafts the startup’s customer agreements, DPAs, and beta terms to match, so the promises the startup makes to customers are backed by the terms it actually has upstream.
He personally drafts every agreement, gives every client his personal cell phone number, and responds quickly.
Andrew S. Bosin LLC
36 Highland Road, Glen Rock, NJ 07452
📞 201-446-9643
📧 andrewbosin@gmail.com
🌐 www.njbusiness-attorney.com
Call for a consultation. Flat-fee legal services for AI and SaaS startups, nationwide.
#AISaaSLawyer #AIStartupLawyer #BestAISaaSLawyer #AIContracts #EarlyAccess #BetaAgreement #FoundationModels #SaaSContracts
This post is provided for general informational purposes and does not constitute legal advice.