+150 XP

Why AI regulation now sets the terms for SaaS contracts

A procurement manager at a mid-size insurer opens a vendor security questionnaire in 2026 and finds a new section: "Does your product embed a general-purpose AI model? If so, name the provider, the risk classification of each AI feature under the EU AI Act, and attach your latest model documentation." Three years ago that question did not exist. Today it can stall a six-figure deal for weeks.

AI regulation has stopped being a compliance afterthought for SaaS (Software as a Service) vendors. It now appears line by line inside Data Processing Agreements (DPAs) and vendor risk questionnaires, and it decides who bears liability when an embedded model gets something wrong.

The regulatory backdrop, briefly

Two regimes matter most right now.

The EU AI Act entered into force in August 2024 and applies in stages. The bans on unacceptable-risk practices took effect in February 2025. Obligations for general-purpose AI (GPAI) model providers followed in August 2025. The Act sorts AI systems into risk tiers:

  • Unacceptable risk: banned outright (social scoring, untargeted scraping of facial images, emotion recognition at work).
  • High risk: strict obligations such as conformity assessments, technical documentation, risk management and human oversight. This covers AI used in hiring, credit scoring, access to essential services and medical devices.
  • Limited risk: transparency duties. Chatbots must tell users they are talking to an AI, and synthetic content must be labelled.
  • Minimal risk: largely unregulated (spam filters, AI in video games, most internal productivity tools).

The Act also sets rules for GPAI models, the foundation models behind products like GPT-5, Claude or Gemini. Models deemed to carry "systemic risk" (training compute above 10^25 FLOPs is the default trigger) face extra duties. Most major providers have signed the Commission's GPAI Code of Practice, published in July 2025. Source: EU AI Act overview, European Commission.

The high-risk timeline is moving. The original date for most high-risk obligations was August 2026. In November 2025 the Commission proposed a "Digital Omnibus" package that would tie those obligations to the availability of standards and support tools. Under that proposal, the latest dates would be December 2027 for use cases like hiring and credit, and August 2028 for AI embedded in regulated products. Check the current status before quoting a date in a contract. Enterprise buyers are not waiting for it, and they ask about classification now.

In the US, there is still no comprehensive federal AI law. State rules fill the gap. Colorado's AI Act covers high-risk automated decision systems and was postponed to 30 June 2026. California added SB 53 in 2025, a transparency law for frontier model developers, alongside its rules on automated decision-making. The FTC (Federal Trade Commission) still pursues deceptive AI claims. A December 2025 executive order directed federal agencies to challenge state AI laws seen as burdensome, which adds legal uncertainty without removing any state obligation yet. For a vendor selling nationally, the practical result is the same: disclosures may differ state by state, and they may change.

How this lands inside actual contracts

Enterprise buyers, especially in banking, insurance and healthcare, now push AI-specific clauses into DPAs and Master Service Agreements. Common additions in 2026:

  • Model provenance disclosure: name the underlying model provider (OpenAI, Anthropic, Google, Mistral, or an open-weight model such as Llama that the vendor hosts itself).
  • Risk tier attestation: a written statement of where each AI feature falls under the EU AI Act. US-only customers ask for it too, because many of them operate globally.
  • Rights to request documentation: model cards, evaluation results, known limitations, and for high-risk features, the instructions for use the Act requires.
  • Data flow restrictions, such as a ban on using customer data to train the foundation model, or a confirmed opt-out.
  • Liability allocation: if the AI feature produces a discriminatory hiring recommendation or a hallucinated compliance answer, does the SaaS vendor pay, or the upstream model provider?

That last point is where most vendors get caught flat-footed.

What a vendor actually inherits by embedding a third-party model

If your SaaS product calls the OpenAI or Anthropic API to power a "smart summarize" feature, you cannot say "that's their model, not our problem."

The EU AI Act uses three roles, and the labels matter in contracts:

  • The GPAI model provider (OpenAI, Anthropic, Google) is responsible for the base model: technical documentation, a training data summary, copyright policy, and systemic-risk duties where they apply.
  • The AI system provider is the company that builds the model into a product and places it on the market under its own name. That is usually you, the SaaS vendor.
  • The deployer is the organisation using the system in its operations. That is usually your customer, the insurer or the HR department.

Many vendors assume they are "just deployers". When the feature ships inside your product, you are normally the provider of that AI system, and for a high-risk feature the heaviest obligations land on you. In practice that means:

  1. Classification. You assess whether your specific use case falls into a high-risk category. The base model in the abstract does not decide this. A general writing assistant is limited risk, and the same model wired into resume screening is high risk.
  2. Conformity work for high-risk features: risk management, data governance, technical documentation, logging, accuracy and robustness testing, then a conformity assessment and CE marking before release.
  3. Transparency. Users must know they are dealing with AI, and generated content has to be marked.
  4. Design for human oversight, so your customer, as deployer, can actually review and override outputs.
  5. Data governance: knowing what customer data reaches the model provider, how long it is kept, and whether it could be used for training.

Your customers then carry deployer duties of their own, such as assigning trained staff to oversee the system and keeping logs. They will expect your product and documentation to make that possible, and that is why these clauses appear in their redlines.

A simple way to see the exposure

A quick internal checklist a product and legal team can adapt:

Feature: "AI resume screener" in an HR SaaS product

1. Underlying model: third-party LLM via API
2. Our role (EU AI Act): provider of the AI system
3. Use case risk tier: HIGH (employment decision)
4. Obligations triggered:
   - Technical documentation + conformity assessment: REQUIRED
   - Human oversight built into the workflow: REQUIRED
   - Bias/accuracy testing on outputs: REQUIRED
   - Instructions for use for the customer (deployer): REQUIRED
5. Contract clauses needed:
   - upstream: model provider commitments on data and documentation
   - downstream: customer duties as deployer, liability split

This does not run. It is the structured risk triage to do before shipping any AI feature, and before signing the DPA that a customer's procurement team sends back redlined.

Why this matters even for "boring" SaaS

It is tempting to think this only applies to hiring or credit. But ordinary features can reach high-risk territory or trigger transparency duties: a support bot that ends up giving financial guidance, a scheduling tool that decides who gets access to a public service, a moderation model that shapes what users see. Classification has to be done feature by feature. A product-level label will not hold up.

Knowledge check

1. A SaaS vendor embeds a third-party foundation model into its hiring-screening product. Under the EU AI Act's risk-tiering logic, why would this likely trigger stricter obligations than the same model embedded in a customer support chatbot?

2. Why has AI-specific language begun appearing inside vendor risk questionnaires and DPAs, rather than being handled separately through general compliance channels?

3. What does the concept of 'systemic risk' obligations for General-Purpose AI (GPAI) models under the EU AI Act primarily depend on?

MULTIPLE CHOICE

4. Select ALL correct answers about the EU AI Act's risk-tier system.

Select all the correct answers.

MULTIPLE CHOICE

5. Select ALL correct answers about why the US and EU regulatory landscapes create different challenges for SaaS vendors selling AI-embedded products.

Select all the correct answers.

What good governance looks like in practice

Vendors that move smoothly through enterprise deals in 2026 tend to do a few things early.

They keep a living AI inventory: every AI feature, which model powers it, what data it touches, their role under the Act, and its risk tier. That document gets attached to security questionnaires, so nobody rebuilds the answer from scratch each time.

They read their upstream terms before promising anything downstream. If your model provider's API terms allow 30 days of abuse-monitoring retention, you cannot promise a customer zero retention. Your customer contract cannot promise more than your upstream contract allows.

And they build a human-in-the-loop path for anything touching employment, credit, healthcare or safety, even when the feature feels low stakes today. Adding oversight before launch is far cheaper than retrofitting it after an incident or a reclassification.

The NIST AI Risk Management Framework (US, voluntary but widely referenced) is a good complement to the EU's binding rules if you want one internal governance process that satisfies both.

Key Takeaways

  • AI regulation now shows up as contract language. DPAs and security questionnaires ask about model provenance, risk classification and data flows.
  • The EU AI Act's risk tiers apply to use cases. The same LLM can be limited risk in one feature and high risk in another.
  • A SaaS vendor that builds a third-party model into its product is usually the provider of that AI system under the Act, not just a deployer. For high-risk features, that brings documentation and conformity duties.
  • EU high-risk deadlines may slip to late 2027 or 2028 under the Digital Omnibus proposal, while GPAI rules have applied since August 2025. In the US, state laws (Colorado, California) face a federal push for preemption, so track dates closely.
  • An up-to-date AI inventory and upstream terms you have actually read get you through enterprise procurement faster than answering each questionnaire ad hoc.