SOC 2 and AI: Which AI Tools Have SOC 2 Type II, and What It Proves
Team Sonomos · Sonomos
The Sonomos research team writes about AI privacy, data protection, and how to use generative AI safely at work.
Short answer, checked 18 November 2026: Most major AI vendors have a SOC 2 Type II report for their business products. OpenAI's covers ChatGPT Enterprise, ChatGPT Edu and its API; Anthropic's covers the Claude API, Claude Team and Claude Enterprise; Google lists SOC 1, 2 and 3 for Gemini in Workspace; GitHub includes Copilot Business and Enterprise in its SOC reports; Cursor, Perplexity, Cognition (the owner of Windsurf) and the legal AI company Harvey list one on their trust pages. Microsoft's Copilot privacy page names ISO 27001, ISO 42001 and HIPAA but not SOC 2, so check its Service Trust Portal for the report you need.
A SOC 2 Type II report proves less than "SOC 2 compliant AI" suggests. It says an independent CPA firm tested the vendor's own controls over a past period, for the products named in its system description. It does not cover a consumer account, it does not say whether the vendor trains on your prompts, and it does nothing for your own SOC 2 audit unless you manage the tool as a vendor and keep the evidence.
This guide lists who has a report and for what, explains what the report does and does not prove, and maps employee AI use to the Trust Services Criteria your auditor will test.
Which AI tools have SOC 2 Type II
Each row is taken from the vendor's own trust portal or documentation on 18 November 2026. Reports themselves sit behind an access request and usually an NDA.
| Vendor | Products named | SOC 2 Type II | Also listed | Source |
|---|---|---|---|---|
| OpenAI | ChatGPT Enterprise and Edu; API | Yes | SOC 3 report; ISO 27001 | OpenAI Trust Portal |
| Anthropic | Claude API, Claude Enterprise, Claude Team | Yes | ISO 27001, ISO 42001; HIPAA for API and Enterprise | Anthropic Trust Center |
| Gemini in Workspace apps; Gemini app for Workspace users | SOC 1, 2 and 3 (type not stated on the hub) | Excludes Gemini Notebook and Gemini in Chrome | Workspace AI privacy hub | |
| GitHub | Copilot Business and Enterprise | In GitHub's SOC 2 reports | ISO 27001 scope | GitHub changelog |
| Microsoft | Microsoft Copilot | Not named on the Copilot privacy page | ISO 27001, ISO 42001, HIPAA | Copilot privacy page |
| Cursor | Cursor | Yes | ISO 27001, ISO 42001 | Cursor security |
| Perplexity | Perplexity | Yes | ISO 27001; HIPAA gap assessment | Perplexity Trust Center |
| Cognition | Devin; Devin Desktop (formerly Windsurf) | Yes, since September 2024 | FedRAMP High for Devin Desktop | Cognition Trust Center |
| Harvey | Harvey legal AI platform | Yes | ISO 27001, 27701, 42001 | Harvey Trust Center |
Three readings of the table matter more than any single row. First, every report is scoped to business products: Anthropic's scope table does not list its consumer plans, and OpenAI's portal is for its enterprise services and API. Second, the GitHub entry dates from a December 2024 notice that Copilot Business and Enterprise were "gaining coverage of control operating effectiveness" for the period April to September 2024; ask for the current report and bridge letter. Third, for legal teams searching for a SOC 2 legal AI tool, Harvey's trust center shows a Type 2 report alongside ISO 27001 and ISO 42001; our comparison of Harvey, Westlaw and Casetext covers the confidentiality terms that the report does not.
What a SOC 2 Type II report proves, and what it does not
SOC 2 reports are attestations under the AICPA's Trust Services Criteria, which cover security, availability, processing integrity, confidentiality and privacy. Security is always in scope; the others are chosen by the vendor. Microsoft's SOC 2 explainer gives a clear account of the mechanics: a Type 2 report describes the system and assesses whether controls "were operating effectively over a specified time period," while "Type 1 audits don't look back over a period of performance."
What a Type II report proves:
- an independent CPA firm tested the vendor's controls, for the systems in its description, across the audit period;
- exceptions the auditor found are written up, usually with management's response, which Microsoft notes is a normal outcome.
What it does not prove:
- That your account is covered. The system description sets the boundary. A personal ChatGPT Plus or Claude Pro account sits outside reports written for enterprise products.
- That controls work today. The report looks backward. Bridge letters fill the gap until the next one, and Microsoft describes them as "self-attestations," not auditor examinations.
- That the vendor will not train on or keep your data. Training defaults and retention periods are set by contract and settings, which our audit of AI training defaults tracks by plan.
- That you are covered. Every report ends with responsibilities the customer must carry out, which Microsoft calls user entity responsibilities and auditors often call complementary user entity controls. If you skip them, the vendor's report does not protect your audit.
- That the tool is "SOC 2 certified." There is no SOC 2 certificate. There is a report, an opinion and a period.
How employee AI use maps to the Trust Services Criteria
Your own SOC 2 report tests your controls, and AI use touches several common criteria and the confidentiality criteria.
CC9.2, vendors and business partners. Every AI tool that receives confidential information is a vendor. It needs an entry in the inventory, a risk assessment, the vendor's report or equivalent evidence, and a review of the user entity responsibilities. A tool nobody assessed, discovered through log review or an employee interview, is the shadow AI gap this criterion exposes.
C1.1 and C1.2, confidentiality. You must identify and maintain confidential information and dispose of it as your policy says. Customer data, source code and financial projections pasted into a consumer account leave the processors you have approved and the retention you control.
CC6.1, CC6.6 and CC6.7, access and transmission. These cover logical access, protection at the system boundary, and restricting the transmission of information to authorized parties. An unmanaged AI site is an unapproved outbound path.
CC2.2, internal communication. Staff must know their responsibilities, which in 2026 means an AI use policy they have acknowledged. Our AI use policy template and policy generator cover the sections auditors look for.
CC8.1, change management. Code written with an AI assistant goes through the same review, testing and approval as any other change. The criteria care about the change, not how it was drafted.
How to satisfy SOC 2 when staff put customer data into AI
The controls below are the ones auditors can test, and each produces evidence across a Type II period.
- Approve a short list of AI tools on business plans whose SOC 2 scope names the product, and record the decision in the vendor inventory.
- Sign the contract terms that match your commitments to customers: no training on your data, defined retention, deletion on request and breach notification, through a data processing agreement where personal data is involved.
- Configure the tools to those terms, and keep screenshots or exports of the settings each quarter.
- Write the rule down: which data classes may go to which tool, and what to do after an accidental paste.
- Control the path. Data loss prevention at the network or on the device, or masking sensitive values before a prompt is sent, gives you evidence the policy is enforced; our comparison with cloud DLP sets out the trade-offs.
- Log incidents and near misses, including AI-related ones, and show they were handled.
If you sell to regulated customers, expect their questionnaires to ask about each of these. Our pages for technology companies and legal teams cover the questions specific to each.
What changes with the proposed 2027 criteria
On 28 October 2026 the AICPA proposed updated SOC 2 Trust Services Criteria. The draft "incorporates expanded considerations for emerging technologies, including AI," with new points of focus on "AI governance, model monitoring, prompt manipulation, data poisoning." Comments are open until 15 January 2027, and if approved the criteria would apply to engagements beginning on or after 31 December 2027, with early adoption permitted.
Until then the 2017 criteria apply, and none of the current criteria names AI. That makes the mapping above an interpretation that your auditor may read differently, so agree it with them before fieldwork rather than during it.
Frequently asked questions
What does SOC 2 compliant mean for an AI tool?
Usually that the vendor has a SOC 2 report in which an independent CPA firm tested its controls against the Trust Services Criteria. There is no SOC 2 certificate. Check which products the report covers, whether it is Type I or Type II, the dates of the period, and the responsibilities it leaves to you.
Does ChatGPT have a SOC 2 Type II report?
Yes, for business use. OpenAI's trust portal lists a SOC 2 Type 2 report covering ChatGPT Enterprise, ChatGPT Edu and its API, as checked on 18 November 2026. A personal Free, Plus or Pro account is not what that report describes.
Is my personal Claude Pro or ChatGPT Plus account covered by the vendor's SOC 2 report?
Not for your audit's purposes. Anthropic's trust center lists SOC 2 Type II for the Claude API, Team and Enterprise only, and OpenAI's portal covers its enterprise services and API. Work done in a personal account sits outside your vendor contracts, whatever the vendor's own controls are.
Does using a SOC 2 Type II AI vendor make us SOC 2 compliant?
No. It helps you meet the vendor-management criterion, CC9.2, but your own report tests your controls: the policy, the approved-tool list, the configuration, the enforcement and the incident records. You also have to carry out the user entity responsibilities listed at the end of the vendor's report.
Treat AI tools as vendors, and keep the evidence
SOC 2 does not ask you to stop using AI. It asks you to know which tools hold your confidential information, to choose vendors whose reports cover the plans you use, to contract and configure them to match what you promised customers, and to show over a full period that the rules held. For the parallel exercise under ISO, see our guide to ISO 27001 and AI tools.
Locke, made by Sonomos, runs on your own computer and replaces names, account numbers, health details and other sensitive values with placeholders before a prompt reaches the AI tools it screens, so the provider never receives them. For the tools it covers, that gives you a control on the transmission path at the point where staff type, though coverage depends on the app, platform and settings, and it is not a SOC 2 report or a substitute for one. You can read how Locke works.
Sonomos publishes this blog and makes Locke, which detects and masks sensitive information on your device before it reaches an AI model. How Locke works
This article is general information about privacy and AI, not legal advice, and reading it does not create a solicitor–client or attorney–client relationship. Regulatory requirements turn on your specific facts and jurisdiction and change over time. Consult qualified counsel before relying on anything here, and check the publication and last-reviewed dates above. We may revise posts, but we do not undertake to keep any post current as the law changes.
Reference pages for this post
Related Articles
Best Privacy-Focused AI in 2026: Chatbots That Keep Your Data Private
The most private AI runs locally, via Ollama or LM Studio. Lumo, Brave Leo and Venice say they keep no chat logs; Duck.ai hides your IP. Compared, Nov 2026.
Can My Employer See My AI Chats? ChatGPT, Claude, Gemini and Copilot
Yes, if your employer controls the account, the computer or the network. Work accounts on ChatGPT, Claude, Gemini and Copilot let it read or keep chats. A side-by-side table of what each vendor documents, as of November 2026.
Can My Employer See My Copilot Chats? Work vs Personal Accounts
Copilot chats on a work account sit in your mailbox, where your employer can search, hold and export them with Purview. A personal account sits outside the tenant, but a work laptop can still capture it. What Microsoft documents, as of November 2026.