Skip to main content
    Jacobs Counsel LLC logo
    Schedule a Call
    All deal anatomies
    Tech, AI & SaaS · Deal Anatomy

    SaaS MSA for AI Startups: Key Contract Terms

    Review an AI startup's SaaS MSA for customer data, training rights, output ownership, indemnity, liability limits and exit terms before signing.

    What it is

    A SaaS master services agreement sets the relationship between a software provider and its customer. For an AI product, it should work with the order form, data processing addendum and actual model-provider terms to address inputs, outputs, training, permitted uses and responsibility when something goes wrong.

    The points below are negotiation questions, not a universal market standard or a ready-to-sign contract. Terms depend on the product, customer, data, bargaining position and applicable law.

    Parties

    • Vendor
      Contracting party. Provides the AI or SaaS service.
    • Customer
      Contracting party. Uses the service on the agreed terms.
    • Model providers and hosting
      Dependencies in the stack, not automatically parties to this agreement. Their own terms may control training, retention and permitted uses.
    • End users
      Other participants who access the service through the customer. Whether they sign anything depends on the deal structure.

    Clause by clause

    12 clauses · what it does, what to negotiate, what to flag.

    01

    Scope and order forms

    Purpose

    Identifies the service actually being bought and where deal terms live.

    What to negotiate

    Describe the service, environments, users and affiliates, and state which document controls if the MSA, order form and policies conflict.

    Red flag

    A scope that does not match the product demonstrated, or an order form that silently changes MSA terms without saying so.

    02

    Customer data

    Purpose

    Defines what the customer's data covers.

    What to negotiate

    Say whether inputs, prompts, outputs, uploaded files, configuration and telemetry are treated as customer data, and how each may be used.

    Red flag

    Categories left undefined, so prompts or outputs fall outside the protections the customer thinks it has.

    03

    Output rights

    Purpose

    Allocates whatever rights exist in what the system generates.

    What to negotiate

    Allocate use rights between the parties and state any restrictions. No contract can create copyright in material that is not copyrightable.

    Red flag

    A promise that every generated output carries exclusive copyright without assessing human authorship and third-party rights.

    04

    Training and improvement

    Purpose

    Decides whether customer material can improve the service or a model.

    What to negotiate

    State whether training is off by default or opt-in, how de-identification or aggregation works, and whether model-provider terms are flowed down.

    Red flag

    Broad service-improvement language that leaves training permissions unclear or inconsistent with the provider terms behind the product.

    05

    IP indemnity

    Purpose

    Allocates third-party intellectual property claims.

    What to negotiate

    Identify what is covered, what is excluded, who controls defense and settlement, and how the indemnity interacts with the liability cap.

    Red flag

    An indemnity whose exclusions remove the risks that actually apply to the product in use.

    06

    Output errors and reliance

    Purpose

    Addresses inaccurate, incomplete or unsuitable output.

    What to negotiate

    Describe review and verification obligations, human oversight where relevant, documentation commitments and the remedy for defects.

    Red flag

    A guarantee of accurate output, or reliance obligations that do not match how the customer plans to use the product.

    07

    Permitted and prohibited uses

    Purpose

    Sets the boundaries of the customer's use.

    What to negotiate

    List prohibited uses specifically, connect them to enforcement steps, and confirm the customer's real use case is permitted.

    Red flag

    Open-ended restrictions decided at the vendor's discretion, or restrictions that conflict with the customer's disclosed use.

    08

    Model providers and other vendors

    Purpose

    Handles the rest of the stack behind the service.

    What to negotiate

    Identify material providers, change-notice rights and flow-down obligations. A model provider is not automatically a party to this agreement.

    Red flag

    MSA promises that cannot be met because the underlying provider terms say something different.

    09

    Privacy and security

    Purpose

    Covers personal data and security commitments.

    What to negotiate

    Attach a data processing addendum where applicable, define roles, set incident-notice timing against the law that applies, and describe security controls and audits. There is no universal 72-hour notice rule between commercial parties.

    Red flag

    Obligations the security program cannot perform, or missing terms required for the actual processing roles and applicable law.

    10

    Liability

    Purpose

    Allocates financial exposure.

    What to negotiate

    Negotiate the cap, any separate caps, exclusions and carve-outs against the deal's actual risk. There is no universal twelve-month-fee cap or standard uncapped carve-out.

    Red flag

    A cap or exclusion set by template that leaves the priced risk with the party least able to control it.

    11

    Renewal and termination

    Purpose

    Sets how the relationship continues or ends.

    What to negotiate

    Define the term, renewal mechanics, notice periods, cure rights, suspension rights and the fee consequences of ending early.

    Red flag

    Renewal or notice mechanics nobody has calendared, or termination rights that are unusable in practice.

    12

    Data return and deletion

    Purpose

    Covers what happens to the data at exit.

    What to negotiate

    Define export formats, the export window, deletion timing, backup handling and any retention required by law.

    Red flag

    Best-efforts deletion language with no timeline, or export rights that expire before a migration can finish.

    How to negotiate it

    The order to work through these clauses for max leverage.

    1. 1
      Map the actual service and data

      Write down what the product does, who uses it, what data goes in, what comes out and which of it is sensitive or regulated.

    2. 2
      Read the complete contract stack

      Review the MSA, order form, DPA, published policies and the underlying model-provider terms together, and specify which document controls on conflict.

    3. 3
      Resolve data and output permissions

      Settle how inputs, prompts and outputs may be used, whether any training is permitted and what rights each side has in the output.

    4. 4
      Price the risk allocation

      Look at indemnities, caps, exclusions, remedies, fees, insurance and what happens if the service fails, and treat them as one commercial package.

    5. 5
      Confirm delivery and exit

      Check security, support, incident handling, renewal, export and deletion with the team that will actually perform them.

    6. 6
      Record decisions before signing

      Close open issues in the document, confirm signatories and confirm which version is the approved one.

    Red flag checklist

    • The written scope does not match the service and data you actually reviewed.
    • Documents in the stack conflict and no order of precedence is stated.
    • Data, training and output permissions are left unclear.
    • Indemnities, caps and exclusions do not match where the risk sits.
    • Security, incident and support commitments were never checked with the performing team.
    • Open issues are left to a side email instead of the signed version.

    Frequently asked

    Does an AI startup need a different MSA?+

    It needs an MSA that fits its service. AI features often require express treatment of prompts, outputs, training, model providers and reliance on results. The rest of the contract still needs to match the business.

    Who owns AI outputs?+

    A contract can allocate the parties' rights, but it cannot create copyright where the law provides none. The U.S. Copyright Office distinguishes human-authored expression from material generated entirely by AI. Review the human contribution, third-party rights and provider terms.

    Can a vendor train on customer prompts?+

    The answer depends on the contract, applicable law and actual service settings. Address training and improvement explicitly, including the role of model providers; do not infer a permission or prohibition from a generic confidentiality clause.

    What liability cap should a SaaS MSA use?+

    There is no single cap that fits every deal. Evaluate fees, likely loss, insurance, available remedies and the duties each party controls, then specify any higher caps or exclusions clearly.

    Related deep-dives

    Sources

    Updated September 16, 2026. General information about contract terms—not legal advice on your specific deal.

    Have one of these on your desk right now?

    Fixed-fee review, scoped before any work begins.

    Discuss a SaaS contract review

    More deal anatomies