SOC 2 and employee AI tool use: what an auditor can test today

Updated

No SOC 2 criterion names AI. Staff use of AI tools still reaches the audit, because it touches criteria you already have: logical access (sign-ins and OAuth grants), monitoring, vendor management and, if it is in scope, Confidentiality. An auditor can test five things today: a written rule on what data may go into which tools, an approved tool list that went through vendor review, a scheduled check of OAuth grants and browser extensions, training that covers the rule, and a risk assessment that mentions AI. The two sources here are an audit firm and a vendor, not the AICPA, so confirm with your CPA firm.

Why it matters when no criterion says "AI"

Linford & Co's article is by Tim Nackos, a CPA and partner at the firm, dated 18 March 2026. The firm also sells ISO 42001 audits, so it has a reason to raise the topic. Its argument is about how SOC 2 is built: "The criteria-based nature of SOC 2 means that if AI is introducing risk to your environment, you are expected to address it", whether or not it appears on a checklist (Linford). It defines shadow AI as "the use of AI-powered tools by company personnel without IT approval" and lists chat assistants, code completion tools, writing assistants, meeting transcription services, image generators and AI browser extensions as examples.

Sonomos published its article on 8 May 2026 under a team byline. It sells a tool that masks sensitive data before it reaches an AI model, so read it as a seller's view (Sonomos).

Which criteria it touches

AreaLinford (audit firm)Sonomos (vendor)
Logical accessCC6: staff who connect AI tools with company credentials, SSO tokens or OAuth give an outside service access that skipped provisioning and access reviewCC6.6 and CC6.7: AI use opens a new outbound path for data, and a transmission to a consumer tool may not be protected
MonitoringCC7.1: data sent to AI services that monitoring does not capture weakens detectionAsks whether DLP, CASB or browser controls exist
VendorsCC9.1: self-adopted tools are vendors onboarded outside vendor management; may be a subservice organizationCC9.2: a tool that receives confidential data without inventory, assessment or contract is a gap; may be a sub-processor
CommunicationTraining that is specific to AICC2.2: staff must be told their responsibilities
ConfidentialityEach prompt sends data to a third party, which is a problem when the Confidentiality criteria are in scopeCC6.1 and C1.1; says Security and Confidentiality are the most directly involved

The two do not cite the same criterion for vendors (CC9.1 against CC9.2), and only Linford raises Processing Integrity, for AI output nobody verified. Neither difference changes what you do. Vendor review itself is covered in SOC 2 vendor management, and whether Confidentiality belongs in your report in the Confidentiality guide.

Five things an auditor can test

Sonomos lists the questions it says auditors ask, starting with "Does the organization have a policy addressing employee use of generative AI tools?", then which tools are approved, whether AI vendors are in the vendor risk program, what the vendors commit to in contract, and whether there were AI incidents in the audit period (Sonomos). Linford's governance program has five parts: approve a set of tools, update the risk assessment, add technical controls, train the workforce, and keep monitoring (Linford). Put together, a small team can be tested on:

  1. A written rule. Which data classes may go into which tools, and who decides. The free AI acceptable use policy template is one way to write it; staff sign it like any other policy (see policy acknowledgment).
  2. An approved tool list that went through vendor review. One row per tool in the vendor inventory, with the plan, the data it may receive and the training, retention and sub-processor terms you read. The vendor review template holds the record.
  3. A scheduled check of OAuth grants and browser extensions. Linford names both. Add the connected-apps list to the periodic access review; the access review tool covers user accounts, and the OAuth list comes from your identity provider's admin console.
  4. Training that covers the rule. Linford wants it specific: why pasting customer data into a prompt is a disclosure, why output needs checking, and which approved tools exist. See security awareness training.
  5. A risk assessment that mentions AI. Linford lists disclosure through prompts, unverified output and unsanctioned vendors. Add them as rows in the risk register.

For evidence, Sonomos suggests screenshots or logs of AI tools configured per policy (training and data sharing turned off), vendor risk assessment records, staff acknowledgment of the AI rule, and incident logs that include near-misses (Sonomos). It also says auditors "want to see that violations are detected and addressed", not that none ever happen. A paste into the wrong tool that was reported, assessed and closed is evidence the control works.

What the sources do not prove

  • How often auditors ask. Linford writes that "auditors are asking about them now" and that shadow AI produces real findings, and Sonomos lists auditor questions, but neither gives a count, a survey or a named audit. Ask your own auditor what they will request.
  • Whether a tool without a SOC 2 report blocks anything. Neither source says an AI tool needs its own report. What you record is the review you did and the terms you relied on.
  • What any AI vendor's terms say today. Sonomos's line that enterprise tiers carry protective terms and "consumer tiers do not" is a summary by a seller. Read the terms for your plan and date your note.
  • That heavy tooling is expected of a ten-person company. Both lists include CASB and DLP. Neither says a small team fails without them. Your control should describe what you really run.

The customer side of the same question

Buyers ask about this too, usually as a questionnaire row on whether their data trains AI models. That answer is covered in "Is our data used to train AI models?", and the security questionnaire answers tool drafts the rest from your policies.

What the generated policies say

Policyseed sells a kit, so read these as one example, not a requirement. Three generated policies carry the rule: the Acceptable Use Policy sets it and the approval steps, the Vendor policy stops data reaching a tool before review (free and personal accounts included), and the Secure Software Development Policy covers AI-written code.

AI assistants, code generation tools and similar services are used only if they are on the approved tool list. Customer data, secrets, personal data and unreleased source code are not entered into any AI tool unless the Security Owner has confirmed in writing that its data handling terms permit it.

Acceptable Use Policy, item 4.6

Tool approval. Personnel who want to use a new SaaS product, browser extension, AI tool or piece of software submit a request to the Security Owner describing the tool, the data it will touch and the business need. The Security Owner reviews its data retention, training-use and sub-processor terms, responds within five business days, records the decision and permitted data classifications in the approved tool list, and arranges single sign-on through Okta where supported. The list is re-checked quarterly.

Acceptable Use Policy, item 5.3

No vendor receives Northwind Cloud Inc data, system access or a connection to Northwind Cloud Inc accounts until it is in the inventory, tiered by the Security Owner and cleared through the due diligence for its tier. Free, trial and personal-account services follow the same rule.

Vendor and Third-Party Risk Management Policy, item 4.2

Northwind receives an independent penetration test at least annually and after major architectural changes; findings are remediated under the Vulnerability and Patch Management Policy and Critical and High findings are retested before closure. Code generated by AI assistants or copied from external sources is reviewed to the same standard as human-written code, and secrets, customer data and unreleased proprietary code are never submitted to unapproved AI tools or external services.

Secure Software Development Policy, item 4.12

Read the full Acceptable Use Policy, Vendor and Third-Party Risk Management Policy and Secure Software Development Policy. An approved exception, such as a pilot of a tool still under review, goes in the policy exception register.

Frequently asked questions

Can employees use ChatGPT or Claude at a company going through SOC 2?
Yes, if your own rules allow it and you can show the rules work. Neither source fetched for this page says SOC 2 bans AI tools. Linford & Co, an audit firm, says the criteria-based design means a risk AI introduces has to be addressed even though no checklist names it. In practice that means an approved tool list, a rule on what data may go into each tool, and a vendor review of each tool before it is approved.
Do AI tools count as vendors under SOC 2?
Both sources treat them as vendors. Linford says employees who adopt AI tools on their own are onboarding vendors outside the vendor management process, and that a tool may be a vendor or a subservice organization depending on how it is used. Sonomos, which sells an AI privacy product, places AI vendors under CC9.2 and says an auditor may treat a tool as a sub-processor. Either way, put each tool in the vendor inventory with the data it receives and the terms you checked.
Is pasting customer data into an AI tool a SOC 2 breach?
The sources do not give a flat yes or no. Sonomos says it depends on how your security policy defines confidential information, whether the AI vendor is contractually bound not to use or disclose the data, whether your acceptable use rules authorized the transmission, and whether the auditor treats the tool as a sub-processor. It adds that pasting such data into a consumer tool is almost certainly unauthorized under most policies even when they never mention AI. Treat a paste into an unapproved tool as a security event and assess it under your incident response process.
Do we need a separate AI policy for SOC 2?
Neither source says a standalone document is required. Sonomos lists a policy addressing generative AI as the first thing an auditor asks about, and Linford's governance list starts with an approved set of tools. A rule inside the acceptable use policy can answer both. Policyseed's generated Acceptable Use Policy carries one AI rule and a tool approval procedure; the free AI acceptable use policy template expands them when a customer or auditor asks for more detail.
Are enterprise AI plans enough on their own?
No. Sonomos says the enterprise tiers of the large AI products provide contract terms that consumer tiers do not: use of your data limited to delivering the service, no model training on it without consent, retention limits and deletion duties, and breach notification. That is the vendor's claim summarized by a third party, so read the terms for the plan you actually buy and record what you found. The plan does not cover the browser extensions, meeting note takers and OAuth-connected assistants that staff add on their own.
How do we find AI tools nobody approved?
Linford lists five detection methods: a CASB or secure web gateway that classifies traffic to AI services, a periodic audit of OAuth grants, an approved list for browser extensions enforced through device management, DLP rules for sensitive data sent to known AI domains, and anonymous staff surveys. A small team without a CASB or DLP can still run the OAuth audit in its identity provider's admin console, restrict extensions on managed devices and ask people. Write down which of these you do, because the auditor tests what your control says.

Policyseed provides governance policy templates and AI tailoring. It is not legal advice and not a compliance guarantee. Management adopts the policies; the CPA firm performs the SOC 2 examination.