SOC 2

Using OpenAI or Anthropic APIs? How to Handle LLM Providers as Subprocessors in SOC 2

By LowerPlane Team
November 19, 2026
11 min read

Short answer

When customer data flows through OpenAI, Anthropic or any other LLM API, they are subservice organisations under SOC 2 (and subprocessors under GDPR). You need to list them in your subprocessor register, carve them out in Section 3 of your SOC 2 report, obtain a signed DPA with suitable data-processing clauses, and notify enterprise customers before adding or switching providers.

The 2027 Trust Services Criteria add two AI points of focus that tighten this specifically. Below is the full register field list, the due-diligence checklist auditors expect, Section 3 language you can lift, and the DPA clauses to look for.

What a complete subprocessor register entry looks like

Minimum viable: provider name and purpose. What a 2026 enterprise security reviewer expects: ten fields. Below is the template for a single LLM provider entry. Put each field in your public Trust Centre and in your internal GRC system.

FieldExample value
ProviderOpenAI, Inc.
ServiceGPT-4o and GPT-5 Chat Completions API
PurposeGenerate natural-language responses to user queries in the product
Data categoriesUser-submitted prompts, uploaded documents, conversation history
Data locationUnited States (enterprise tier with EU data residency available)
Retention30 days (enterprise tier) — see DPA §4.3
Training on inputsNo — contractually excluded per enterprise MSA §7.2
AttestationSOC 2 Type 2 (annual, dated YYYY-MM-DD) — see Trust Centre
Breach notificationWithin 72 hours per DPA §9
Added to registerYYYY-MM-DD; customer notice sent YYYY-MM-DD

OpenAI example shown. The same ten fields apply to Anthropic, Google Gemini, Mistral, Cohere, Azure OpenAI Service, AWS Bedrock, and any vector-database-plus-model stack.

Due-diligence checklist

Collect these before onboarding the provider. Attach each artefact to the register entry; auditors sample from this list during Type 2 testing.

SOC 2 Type 2 report

Confirms the LLM provider operates the controls they claim. Request the full report under NDA, not just the public assurance page.

Signed DPA / MSA addendum

Required for GDPR Article 28 flow-down and for the SOC 2 subservice organisation carve-out language in your Section 3.

Zero-retention or short-retention commitment

Reduces data exposure window. Document the specific contract clause, not the marketing page.

No-training commitment

Prevents your customers' data becoming training signal for a future model version.

Data residency options

Needed for EU, UK and India customers with data-localisation requirements. OpenAI and Anthropic both offer EU residency on enterprise tiers.

Breach notification timeline

Must be short enough to meet your own customer-facing SLA (GDPR 72h, most enterprise MSAs 24–72h).

Change notification process

How the provider tells you about new sub-subprocessors. Needed for your own flow-down to customers.

Section 3 language

Section 3 of a SOC 2 Type 2 report is where management describes the system. For LLM providers you need three paragraphs: carve-out disclosure, a complementary user entity control (CUEC), and the data-flow description. Reusable boilerplate below — adapt the specifics but keep the structure.

Carve-out disclosure

"The Company uses OpenAI, Inc. as a subservice organisation to provide large language model inference for the natural-language response feature. OpenAI is carved out of this report; its controls have been evaluated separately by its service auditor (SOC 2 Type 2 report available under NDA)."

CUEC — customer responsibility

"User entities are responsible for classifying the data they submit to the Company's API and for ensuring that any data unsuitable for processing by third-party LLM providers is excluded from such submissions."

Data flow description

"User-submitted prompts and attachments are transmitted over TLS 1.3 to the OpenAI API endpoint, processed in OpenAI's US infrastructure, and responses returned to the Company. Retention at OpenAI is limited to 30 days per enterprise MSA §4.3."

What the 2027 criteria change

The AICPA exposure draft (28 Oct 2026) adds seven AI points of focus to the Trust Services Criteria. Two directly affect how LLM subprocessors are handled:

  • CC9 (Risk Mitigation) AI point of focus. Vendor management must cover AI-specific risks: model drift, provider deprecation, training-data provenance for the provider's models, and the provider's own AI risk governance. Expect auditors to ask for an AI-specific vendor risk register row per provider.
  • C1 / P4 (Confidentiality and Privacy) AI point of focus. Data flowing to AI providers must be classified before transmission and documented in the data flow map. "We send user prompts to OpenAI" is no longer sufficient — you need a classification table that says which data categories can go to which provider.

The new criteria are effective for engagements beginning on or after 31 December 2027, with early adoption permitted. If your Type 2 observation window spans 2027–2028, build to the new points of focus now so no rework on renewal.

One rule

Treat LLM provider changes like you treat cloud-region changes. New provider, new region, new fine-tuning service — all go through the same vendor-onboarding pipeline: due diligence pack, DPA, register entry, customer notice, Section 3 update. Teams that add providers ad-hoc for new features end up with gaps that surface at the worst possible moment — during a customer's audit or your own Type 2 testing.

Frequently asked questions

Are OpenAI and Anthropic subprocessors under SOC 2?

Yes, when customer data flows through their APIs. SOC 2 does not use the GDPR term 'subprocessor' directly — it uses 'subservice organisation' — but in a 2026 enterprise questionnaire the two mean the same thing. Any third party that processes customer-controlled data on your behalf belongs on the subprocessor list and in your Section 3 description as a carved-out subservice organisation.

Does sending a prompt to an API count as data processing?

If the prompt contains customer-provided content (text, documents, chat history, PII, confidential business data), yes. Prompts from app telemetry or documentation strings that your team wrote do not count. The practical test: could the customer object to that data leaving your boundary? If yes, the LLM provider is a subprocessor for that flow.

What does the auditor want to see?

Five artefacts. (1) An updated subprocessor list with the LLM provider named, purpose stated, and data categories listed. (2) A signed DPA (or MSA with data-processing addendum) that covers use of LLM APIs. (3) Evidence of due diligence — the provider's SOC 2 report or equivalent. (4) A change-notification mechanism for when you add, remove or switch providers. (5) Evidence that enterprise customers were notified of the subprocessor before you shipped the feature.

What is the right DPA language?

Three clauses to add, if not already present. First, a definition of 'LLM subprocessor' or 'AI subprocessor' that covers model providers, fine-tuning services and vector-database-plus-model stacks. Second, a notice provision — typically 30 days before a new LLM subprocessor goes live, with the customer's right to object. Third, an audit flow-down — your DPA with the LLM provider must give you contractual rights to request their SOC 2 report and breach notifications. OpenAI and Anthropic both provide enterprise DPAs with suitable language; use theirs.

How does the 2027 SOC 2 revision change this?

The 2027 Trust Services Criteria add seven AI-focused points of focus. Two of them explicitly cover AI subprocessor governance: one under CC9 (risk mitigation — vendor management for AI providers) and one under C1 / P4 (data governance — what customer data enters a model and where it goes). Engagements beginning on or after 31 December 2027 use the new criteria. If you are heading into a Type 2 observation window in 2027, build to the new points of focus now so no rework.

What about zero-retention or no-training commitments?

OpenAI's enterprise tier, Anthropic's Claude for Work, and most other serious providers commit in writing to not train on API inputs and to short retention windows (30 days on the OpenAI enterprise tier, zero retention available on Anthropic). Document these commitments in the subprocessor register. Auditors are now asking for the specific contract clause reference, not just a vendor marketing page.

Keep your subprocessor register audit-ready

LowerPlane tracks LLM subprocessors, auto-generates Section 3 language, and sends customer notices on provider changes.

Book a demo