A 14-person AI agency in Denver delivered a customer service chatbot to a regional insurance carrier last year. The chatbot worked well — until it hallucinated a coverage guarantee to a policyholder who then filed a claim expecting $340,000 in benefits. The carrier turned to the agency, pointing to the terms of service that accompanied the product. The problem: those terms were a generic SaaS template the agency had pulled from a legal docs website and barely customized. There was no language addressing AI-generated output accuracy, no disclaimers around hallucination risk, and no liability cap specific to AI behavior. The agency spent $87,000 in legal fees negotiating a settlement that could have been avoided with properly drafted AI terms of service.
This is not an edge case anymore. Every AI product your agency delivers — chatbots, recommendation engines, content generators, classification systems, predictive models — produces outputs that carry real-world consequences. Your terms of service are the document that defines who is responsible when those outputs go wrong, what the client can and cannot do with the system, and how disputes get resolved.
Most agencies treat terms of service as an afterthought. They copy-paste language from a non-AI product, slap it on the deliverable, and hope for the best. That approach is a ticking time bomb in the age of AI. Here is how to draft AI terms of service that actually protect your agency and set clear expectations with clients.
Why Standard SaaS Terms Fail for AI Products
Traditional SaaS terms of service were written for deterministic software. When a user clicks a button, the software produces a predictable, repeatable result. The terms assume that the product behaves consistently and that bugs can be identified and fixed.
AI products break every one of those assumptions.
Non-deterministic outputs — The same input can produce different outputs depending on the model state, temperature settings, context window, and a dozen other factors. Your terms need to acknowledge that outputs are probabilistic, not guaranteed.
Continuous learning and drift — If the AI system incorporates feedback or retrains on new data, its behavior changes over time. Standard terms that warrant "product functionality as described in documentation" become meaningless when the product's behavior evolves.
Data dependency — AI product quality depends heavily on the quality and relevance of input data. If a client feeds garbage data into a classification system, the outputs will be garbage. Your terms need to address data quality responsibilities clearly.
Third-party model dependencies — Most agency-built AI products rely on third-party foundation models (OpenAI, Anthropic, Google, open-source models). These models get updated, deprecated, or modified by their providers. Your terms need to address what happens when underlying model changes affect product behavior.
Black box decision-making — Some AI systems produce outputs through processes that are difficult or impossible to fully explain. Your terms need to address transparency obligations and limitations.
The Core Sections Every AI Terms of Service Needs
1. Service Description with AI-Specific Qualifiers
Start your terms with a clear description of what the AI product does — and what it does not do. This is where most agencies fail. They describe the product in aspirational marketing language rather than precise legal language.
What to include:
- A factual description of the AI system's function
- Explicit statement that the system produces probabilistic outputs, not guaranteed results
- Clear boundaries on what the system is designed to handle versus what falls outside its scope
- Statement that the system is a decision-support tool, not an autonomous decision-maker (if applicable)
- Identification of the underlying AI models or technologies used, at a category level
Example language direction: "The Service utilizes machine learning models to generate [descriptions/classifications/recommendations] based on inputs provided by the Client. Outputs are probabilistic in nature and may vary across identical inputs. The Service is designed to assist human decision-making and is not intended to replace professional judgment in [specific domain]."
2. Output Accuracy and Reliability Disclaimers
This section is the single most important piece of your AI terms of service. It addresses the fundamental truth about AI: it can be wrong.
What to cover:
- No accuracy guarantee — State clearly that outputs are generated by machine learning models and are not guaranteed to be accurate, complete, or error-free
- Hallucination risk — For generative AI products, explicitly acknowledge that the system may produce outputs that appear plausible but are factually incorrect
- Human review requirement — State whether and when human review of AI outputs is required before acting on them
- Decision responsibility — Clarify that the client bears responsibility for decisions made based on AI outputs
- Domain limitations — Identify specific domains where the AI output should not be relied upon (medical advice, legal counsel, financial recommendations)
3. Data Rights and Responsibilities
AI products consume data, generate data, and often learn from data. Your terms need to address all three flows.
Input data:
- Who owns the data the client provides to the system
- What your agency can do with that data (process, store, use for model improvement)
- Data quality requirements and who bears responsibility for input data quality
- Data format and preparation requirements
Output data:
- Who owns the outputs generated by the AI system
- Whether outputs can be used for any purpose or are restricted
- Whether the client has exclusive rights to outputs or whether similar outputs may be generated for other clients
Training data:
- Whether client data will be used to train or fine-tune models
- Whether the client can opt out of training data usage
- How training data is anonymized or aggregated
- Data retention periods for training data
4. Intellectual Property Provisions
IP in AI is genuinely complex, and your terms need to navigate several distinct IP questions.
Model IP — Your agency or the third-party model provider typically retains ownership of the underlying AI models. State this clearly.
Custom training and fine-tuning IP — If the client paid for custom model training or fine-tuning, who owns the resulting model weights? This is a critical negotiation point. Many agencies retain model ownership while granting the client a license to use the trained model.
Prompt and configuration IP — System prompts, retrieval configurations, and pipeline architectures that your agency develops represent significant intellectual property. Your terms should protect these.
Output IP — Under current law in most jurisdictions, AI-generated content has uncertain copyright status. Your terms should address this reality rather than making claims about copyright protection that may not hold up.
5. Third-Party Dependencies and Model Changes
Your AI product almost certainly depends on third-party services — cloud infrastructure, foundation model APIs, embedding services, vector databases. Your terms need to address these dependencies.
What to cover:
- Identification that the product relies on third-party AI services
- Disclaimer that third-party service changes may affect product behavior
- Your agency's obligation (or lack thereof) to maintain compatibility when third-party services change
- Pass-through of third-party terms where applicable
- Right to substitute equivalent third-party services
6. Usage Restrictions and Acceptable Use
Define what the client can and cannot do with the AI product. This section protects both your agency and the client.
Common restrictions:
- Prohibited use cases (generating illegal content, circumventing safety measures, processing data in violation of privacy laws)
- Volume limitations and rate limits
- Restrictions on reverse engineering the AI models
- Restrictions on using the product to build competing AI services
- Geographic restrictions based on AI regulations in different jurisdictions
7. Liability Limitations Specific to AI
Your general liability limitation clause needs AI-specific provisions. Standard limitation of liability language may not adequately address AI-specific risks.
Key provisions:
- Consequential damages exclusion — Exclude liability for damages arising from reliance on AI-generated outputs
- Aggregate liability cap — Set a reasonable cap (typically tied to fees paid)
- Carve-outs — Identify what is not subject to the liability cap (data breaches, IP infringement, willful misconduct)
- Indemnification — Define mutual indemnification obligations, including the client's obligation to indemnify for misuse of AI outputs
- Force majeure for model changes — Consider whether third-party model deprecation or significant behavioral changes constitute force majeure
8. Compliance and Regulatory Provisions
AI regulation is evolving rapidly. Your terms need to address compliance obligations without over-promising.
What to include:
- Statement of current regulatory compliance (EU AI Act classification, if applicable)
- Allocation of compliance responsibilities between agency and client
- Right to modify the product to maintain regulatory compliance
- Client's obligation to use the product in compliance with applicable laws in their jurisdiction
- Cooperation obligations for regulatory inquiries or audits
Drafting Process and Best Practices
Start with a Risk Assessment
Before drafting, catalog the specific risks associated with the AI product you are delivering. A content generation tool has different risk profiles than a medical image classifier. Your terms should be calibrated to the actual risks, not copied from a generic template.
Risk categories to assess:
- Output accuracy risk (what happens when the AI is wrong?)
- Data privacy risk (what data does the system process and what are the privacy implications?)
- Regulatory risk (what regulations apply to this specific use case?)
- Dependency risk (what third-party services could fail or change?)
- Misuse risk (how could the client use the product in ways you did not intend?)
Layer Your Terms
Consider a three-layer approach to AI terms of service:
Layer 1: Master Service Agreement — Covers the overall agency-client relationship, general liability, payment terms, and dispute resolution.
Layer 2: AI Product Terms — Covers AI-specific provisions that apply across all AI products your agency delivers. This includes output accuracy disclaimers, data rights frameworks, and AI-specific liability provisions.
Layer 3: Product-Specific Schedule — Covers provisions unique to the specific AI product being delivered, including use case restrictions, performance benchmarks, and domain-specific compliance requirements.
This layered approach lets you maintain a consistent AI terms framework while customizing for each engagement.
Use Plain Language Where Possible
Legal terms do not need to be incomprehensible. Clients who understand the terms are less likely to dispute them later. Use clear headers, bullet points, and plain language explanations alongside the formal legal provisions. Some agencies include a "plain language summary" at the top of each section that explains the legal provision in everyday language.
Address the AI Explanation Gap
Many clients do not understand how AI works. Your terms should include a brief, plain-language explanation of how the AI product generates outputs, what factors affect output quality, and why outputs may vary. This context helps clients understand the disclaimers and limitations rather than viewing them as the agency trying to avoid responsibility.
Build in Amendment Mechanisms
AI technology and regulation change rapidly. Your terms need mechanisms for updating without requiring a complete renegotiation every time a regulation changes or a model provider updates their service.
Practical approaches:
- Allow unilateral updates to technical specifications with reasonable notice
- Require mutual agreement for changes that materially affect client rights or service functionality
- Include automatic compliance updates that take effect when new regulations apply
- Set periodic review dates to assess whether terms remain appropriate
Common Mistakes Agencies Make
Mistake 1: Using consumer-facing AI terms for B2B products. Consumer AI terms (like those for ChatGPT or Claude) are designed for individual users. They do not address enterprise deployment scenarios, custom integrations, or the specific risks of AI products built for business operations.
Mistake 2: Over-promising accuracy. Any language that guarantees specific accuracy levels, error rates, or performance benchmarks becomes a liability when the AI underperforms. Use "commercially reasonable efforts" language rather than guarantees.
Mistake 3: Ignoring jurisdiction-specific requirements. The EU AI Act, state-level AI regulations in the US, and emerging frameworks in other jurisdictions create different requirements. If your client operates in multiple jurisdictions, your terms need to address this.
Mistake 4: Not addressing model sunset scenarios. What happens when the underlying AI model is deprecated by its provider? Your terms need a clear process for migration, substitution, or termination.
Mistake 5: Failing to separate data processing terms. In many jurisdictions, AI data processing requires a separate Data Processing Agreement that addresses GDPR, CCPA, or other privacy regulation requirements. Your AI terms of service should reference but not try to replace a proper DPA.
Mistake 6: Silent on monitoring and support. AI products require ongoing monitoring for drift, degradation, and behavioral changes. Your terms should clearly define what monitoring is included, what is additional, and who is responsible for what.
Working with Legal Counsel
Drafting AI terms of service is not a DIY project. You need legal counsel who understands both AI technology and contract law. Here is how to work effectively with lawyers on this.
Prepare before the engagement:
- Document the specific AI products you deliver and how they work
- Catalog the third-party services and models you depend on
- List the known risks and your current mitigation approaches
- Gather examples of AI terms from comparable companies
- Identify the jurisdictions where your clients operate
Guide the drafting process:
- Provide technical context so the lawyer understands what AI outputs actually are
- Review drafts for technical accuracy (lawyers may mischaracterize how AI works)
- Push back on overly conservative language that would make the product unmarketable
- Ensure the terms are practically enforceable, not just legally sound
Budget appropriately:
- Expect to spend $5,000 to $15,000 for a comprehensive AI terms of service package with an experienced technology attorney
- Budget for annual reviews and updates as regulations evolve
- Consider whether a law firm with AI specialization is worth the premium over a general technology attorney
Your Next Step
Pull up the terms of service currently attached to your most recent AI product delivery. Read them through the lens of this article. Count how many of the eight core sections are missing or inadequately addressed. If the answer is more than two, you have a terms of service gap that creates real liability exposure.
Start with the risk assessment. Catalog the specific risks of your most common AI product. Then engage legal counsel to draft or revise your AI terms of service using the framework outlined here. The goal is not a perfect legal document on day one — it is a terms framework that addresses AI-specific risks and can evolve as your products and the regulatory landscape evolve.
The agency in Denver wishes they had spent $10,000 on proper AI terms of service before they spent $87,000 cleaning up a mess that proper terms would have prevented. Do not be that agency.