SOC 2

SOC 2 for AI Vendors: Enterprise Buyer’s Security Checklist

Enterprise adoption of artificial intelligence is growing rapidly. Businesses are using large language models (LLMs) to improve customer support, automate workflows, analyse documents, write software and support business decisions. However, as AI tools become more deeply connected to business operations, security and compliance risks are becoming harder to ignore.

An AI vendor may offer impressive features, fast responses and competitive pricing, but these benefits mean little if the organisation cannot trust how its information is handled.

This is why SOC 2 for AI vendors has become an important consideration for enterprise technology buyers.

SOC 2 reports provide independent assurance about the controls a service organisation has in place. They can help buyers understand how a vendor manages security, protects information and operates its systems.

However, a SOC 2 report does not automatically mean that an AI product is safe for every use case. It does not guarantee that confidential information will never be exposed, that AI outputs will always be correct or that an AI agent cannot perform an unwanted action.

Enterprise buyers need to look beyond the existence of a SOC 2 report and ask detailed questions about data handling, model providers, access controls, AI governance and operational risks.

This guide explains what SOC 2 means for AI vendors, how Type I and Type II reports differ, and which questions businesses should ask before adopting an LLM tool.

What Is SOC 2 for AI Vendors?

SOC 2 stands for System and Organisation Controls 2. It is an assurance reporting framework developed by the American Institute of Certified Public Accountants (AICPA).

A SOC 2 examination is performed by an independent CPA firm and evaluates controls against relevant Trust Services Criteria.

These criteria cover security, availability, processing integrity, confidentiality and privacy.

Why Does SOC 2 Matter for AI Vendors?

AI tools often process information that businesses would not share publicly.

This may include customer records, employee information, financial documents, software code, business plans and internal communications.

When an organisation adopts an external LLM tool, it may be trusting that vendor and its service providers with important information.

A SOC 2 report can provide evidence about how the vendor manages relevant risks and whether its controls have been designed and, for Type II reports, operated effectively.

For enterprise buyers, this makes SOC 2 a useful part of vendor due diligence.

However, it should be treated as evidence to review, not as a guarantee of complete security.

SOC 2 Type I vs Type II: What Is the Difference?

One of the first questions enterprise buyers should ask is whether the vendor has a SOC 2 Type I or Type II report.

Although both can provide useful information, they answer different questions.

SOC 2 Type I Report

A Type I report examines the design of relevant controls at a specific point in time.

It helps buyers understand whether the controls were suitably designed to meet the applicable criteria as of the report date.

For example, an AI vendor may have policies and systems for managing employee access, protecting customer information and responding to security incidents.

A Type I report assesses the design of those controls, but it does not provide the same evidence about their operation over time as a Type II report.

SOC 2 Type II Report

A Type II report examines both the design and operating effectiveness of relevant controls over a defined review period.

This gives buyers more information about whether the controls operated as intended during that period.

For enterprise procurement, a Type II report is often more useful because it provides evidence from ongoing operations rather than a single date.

However, the value of the report still depends on its scope, review period, findings and the systems covered.

Which Report Should Enterprise Buyers Prefer?

For established AI vendors handling sensitive business information, a current SOC 2 Type II report is generally a stronger starting point.

However, buyers should not reject a vendor solely because it has a Type I report. The decision should depend on the risks involved, the vendor’s maturity and the other evidence available.

A Type II report also does not replace the need to review AI-specific security controls.

CTA

What Does SOC 2 Actually Cover?

SOC 2 examinations use the AICPA Trust Services Criteria.

The Security category is required for SOC 2 examinations, while the other categories are included when relevant to the engagement.

Security

Security focuses on protecting systems and information against unauthorised access and other security risks.

For an AI vendor, relevant controls may include identity management, employee access reviews, system monitoring, vulnerability management and incident response.

Availability

Availability concerns whether systems are available for operation and use according to the vendor’s commitments.

This matters when AI applications support important business processes or customer-facing services.

Processing Integrity

Processing integrity focuses on whether system processing is complete, valid, accurate, timely and authorised in relation to the service’s commitments.

However, this does not mean a SOC 2 report guarantees that an LLM will generate factually correct answers.

AI output quality requires separate testing and monitoring.

Confidentiality

Confidentiality concerns the protection of information designated as confidential.

For AI vendors, this may be relevant to customer documents, internal business information, source code and other sensitive material.

Privacy

Privacy addresses the collection, use, retention, disclosure and disposal of personal information according to the applicable criteria.

Buyers should still review the vendor’s privacy terms and their own legal obligations, including requirements under UK GDPR where applicable.

Why SOC 2 Alone Is Not Enough for LLM Security

Traditional software vendors and AI vendors share many security requirements.

Both need strong access controls, secure infrastructure, incident response processes and appropriate protection of customer information.

However, LLM tools introduce additional risks because they process natural language, generate outputs and may connect with other applications.

LLMs Can Process Sensitive Information in Unexpected Ways

Employees may submit confidential documents, customer records or internal business information to AI tools.

The organisation needs to understand where that information goes, how long it is retained and whether other providers can access it.

A SOC 2 report may cover some related controls, but buyers must check whether the relevant AI processing systems are actually included in the report’s scope.

AI Outputs Can Be Incorrect

LLMs can generate information that appears convincing but is inaccurate.

This creates risks when AI outputs are used in legal, financial, customer-facing or other important business processes.

SOC 2 does not replace the need for output testing, human review and clear rules about when AI-generated information can be trusted.

AI Agents Can Take Actions

Some AI tools can connect to business systems and perform tasks automatically.

An AI agent might create a support ticket, update a customer record or access internal files.

These capabilities require additional controls over permissions, tool access and human approval.

A vendor’s SOC 2 report should not be treated as proof that every possible agent action is safe.

12 Security Questions Enterprise Buyers Should Ask AI Vendors

Before adopting an LLM tool, enterprise buyers should review the vendor’s security evidence and ask questions about how the product operates.

The following questions can form part of a practical vendor assessment.

1. Do You Have a Current SOC 2 Type II Report?

Ask the vendor to provide its latest SOC 2 report under an appropriate confidentiality agreement.

Review the reporting period, the independent auditor’s opinion and the systems included in the examination.

What Should Buyers Check?

Confirm that the report covers the service your organisation intends to use.

A vendor may have a SOC 2 report for one platform or business unit while offering other products that fall outside its scope.

Also review any exceptions identified during testing and the vendor’s response to them.

A report should be assessed based on its findings, not simply the fact that it exists.

2. Does the SOC 2 Report Cover the Actual AI Product?

This is one of the most important questions when evaluating SOC 2 for AI vendors.

Some technology companies operate several products, services and infrastructure environments.

Their SOC 2 examination may not cover every system.

Why Product Scope Matters

An AI vendor may provide a chatbot, API service, agent platform and data processing tools.

If the report covers only part of the organisation, buyers need to understand which systems are included.

Ask the vendor to identify the specific product, infrastructure and service boundaries described in the report.

This helps prevent organisations from assuming that controls tested for one service apply to every AI capability the vendor offers.

3. Is Customer Data Used to Train AI Models?

Enterprise buyers should ask whether the vendor uses customer inputs, outputs, uploaded files or other information to train or improve AI models.

The answer should be clear and supported by contractual terms.

Questions About Model Training

Ask whether customer information is used for training by default, whether training can be disabled and whether these settings apply to every model provider used by the service.

Also ask whether the vendor uses customer information for evaluation, abuse monitoring or other purposes separate from model training.

A statement that data is not used for training does not necessarily mean it is never stored or processed.

4. Where Is Our Data Stored and Processed?

Data location can affect privacy, regulatory and contractual obligations.

Ask where the vendor stores customer information and where AI requests are processed.

Data Residency and International Transfers

Some AI tools send requests to external model providers operating in different regions.

Enterprise buyers should understand which countries may be involved and what safeguards apply to international transfers of personal data.

For organisations subject to UK GDPR, these questions can be particularly important.

Buyers should also ask whether the vendor offers regional processing or data residency options.

5. Which Third-Party Model Providers and Subprocessors Do You Use?

Many AI vendors do not operate every model or infrastructure component themselves.

They may depend on external model providers, cloud platforms, analytics services and other subprocessors.

Understand the AI Supply Chain

Ask for a current list of relevant subprocessors and their responsibilities.

Find out which providers receive customer prompts, files, outputs or metadata.

Also ask how the vendor assesses third-party security and informs customers about changes to its subprocessors.

This matters because a vendor’s own controls do not automatically cover every risk introduced by external providers.

6. How Long Do You Retain Prompts, Outputs and Logs?

Data retention is an important part of AI security.

Prompts and outputs may contain confidential information even when the original application was not designed to store sensitive records.

Review Retention and Deletion Controls

Ask how long the vendor retains prompts, outputs, uploaded files, logs and other related information.

Find out whether customers can configure retention periods and request deletion.

Buyers should also understand how deletion applies to backups and third-party processors.

Retention settings may differ between the main AI application, monitoring systems and model providers.

7. How Do You Protect Sensitive Information?

Ask what measures the vendor uses to protect customer information during transmission, processing and storage.

These may include encryption, access restrictions, network security and data protection controls.

Look Beyond Encryption Claims

Encryption is important, but it does not solve every security problem.

An authorised user with excessive permissions may still access sensitive information.

Buyers should therefore review how the vendor limits employee access, monitors administrative activity and protects confidential records.

For AI applications, also ask whether sensitive information can be identified or restricted before it is sent to a model.

8. How Do You Manage User Access and Permissions?

Enterprise AI tools should support appropriate access management.

Ask whether the platform supports single sign-on, multi-factor authentication, role-based access controls and user lifecycle management.

Access Controls for AI Agents

If the tool includes AI agents, ask how agent permissions are managed.

An agent should not automatically receive access to every system or document available to the user who created it.

Buyers should understand whether permissions can be restricted by task, application or business role.

They should also ask which actions require human approval.

9. How Do You Protect Against Prompt Injection?

Prompt injection is a security risk in which malicious or untrusted instructions attempt to influence an AI system’s behaviour.

For example, an attacker might place instructions inside a document or webpage that an AI assistant is expected to read.

The attacker may try to make the AI ignore its original task, reveal information or perform an unwanted action.

Ask About Practical Defences

Enterprise buyers should ask how the vendor tests for prompt injection and limits the actions an AI system can take.

Useful safeguards may include restricted tool permissions, separation of trusted instructions from untrusted content, output checks and human approval for sensitive actions.

No single safeguard can guarantee that every prompt injection attempt will be prevented.

Organisations should evaluate the vendor’s controls against the risks of their intended use case.

10. How Do You Detect and Respond to Security Incidents?

Security incidents can occur even when a vendor has established controls.

Buyers should understand how the vendor detects suspicious activity, investigates incidents and informs affected customers.

Review Incident Notification Commitments

Ask whether the vendor has a documented incident response process.

Review its contractual commitments concerning notification times, customer communication and incident support.

Also ask how the vendor handles incidents involving external AI model providers or subprocessors.

Clear responsibilities are especially important when several companies participate in delivering the AI service.

Quick link: Shadow AI Detection Tools Compared: Browser, Endpoint, Network and Email

11. Can We Monitor AI Usage and Access Activity?

Visibility is essential for enterprise AI governance. Buyers should understand what information the vendor makes available about users, applications, model requests and administrative actions.

Look for Useful Audit Records

Ask whether the platform records authentication activity, permission changes, model usage and other relevant events.

Check whether logs can be exported to existing security monitoring systems.

Organisations should also review how logs are protected and how long they are retained.

For AI applications, monitoring can support security investigations, operational reviews and cost management.

However, logging should be designed to avoid unnecessarily storing sensitive prompt content.

12. How Do You Manage AI Costs, Usage Limits and Unexpected Activity?

Security and compliance are central to vendor assessments, but enterprise buyers should also understand how AI usage is controlled.

LLM tools may charge based on tokens, requests, users, subscriptions or other measures.

Automated agents can also generate repeated requests that increase spending.

Ask About Financial Controls

Find out whether the vendor supports budgets, spending alerts, usage limits and reporting by team or application.

Ask whether administrators can identify unusual consumption and restrict access when needed.

A platform may have strong security controls while still offering limited visibility into AI costs.

For organisations scaling AI adoption, financial governance should be part of the procurement process.

SOC 2 for AI Vendors: Enterprise Evaluation Checklist

The 12 questions above can help procurement, security and technology teams build a consistent review process. However, enterprise buyers should also collect evidence rather than relying only on written answers.

The following checklist can help teams compare AI vendors before approving a purchase.

Evaluation AreaEvidence to RequestWhy It Matters
SOC 2 assuranceCurrent Type II report and auditor’s opinionConfirms the scope and results of an independent examination
Product coverageDescription of systems included in the reportShows whether the intended AI service is covered
Data protectionData processing agreement and security documentationExplains how business information is protected
Model trainingWritten customer data usage termsClarifies whether information may be used for model training
Data locationProcessing regions and transfer safeguardsSupports privacy and regulatory reviews
Third-party providersSubprocessor list and model provider detailsHelps assess external dependencies
Access managementSSO, MFA and role-based access documentationSupports control over users and permissions
AI securityPrompt injection testing and agent security informationHelps assess risks specific to AI applications
Incident responseSecurity incident procedures and notification termsClarifies responsibilities when problems occur
Usage governanceAudit logs, budgets and monitoring featuresSupports ongoing oversight of AI operations

A completed checklist does not replace a detailed security assessment. However, it helps buyers identify missing information and compare vendors using the same standards.

Common Mistakes When Reviewing SOC 2 for AI Vendors

Even experienced procurement teams can make mistakes when reviewing AI security.

Understanding these mistakes can help organisations avoid approving tools without sufficient evidence.

Mistake 1: Treating SOC 2 as a Certification

SOC 2 is an attestation reporting framework, not a product certification.

A vendor may describe itself as SOC 2 compliant, but enterprise buyers should ask for the actual independent report and review its findings.

The report provides more useful information than a badge on a website.

Mistake 2: Assuming Every AI Feature Is Covered

A vendor may introduce new models, agents or integrations after its SOC 2 review period.

These changes may create risks that were not part of the examination.

Buyers should ask how the vendor assesses new AI features and whether important changes have occurred since the report was issued.

Mistake 3: Ignoring Third-Party Model Providers

An AI application may rely on several external services.

Even when the main vendor has a SOC 2 Type II report, buyers still need to understand the controls and responsibilities of relevant third parties.

Reviewing the wider AI supply chain is essential when sensitive business information is involved.

Mistake 4: Focusing Only on Security and Not Operations

A secure AI platform can still create operational problems through unreliable outputs, poor access management or uncontrolled spending.

Enterprise buyers should evaluate the complete service, including reliability, monitoring, support and financial controls.

Mistake 5: Treating Vendor Approval as a One-Time Task

AI products change frequently. A vendor may introduce new models, update its data handling terms or add integrations that affect security requirements.

Organisations should review important vendors regularly and reassess them when material changes occur.

How SOC 2 Fits Into Wider AI Governance

SOC 2 provides useful evidence about a vendor’s controls, but enterprise AI governance involves additional responsibilities.

Organisations need to understand not only whether a vendor operates appropriate controls, but also whether their own employees and applications use AI safely.

SOC 2 and ISO 27001

ISO 27001 is an international standard for information security management systems.

It focuses on establishing, maintaining and improving a structured approach to information security.

SOC 2 and ISO 27001 can complement each other, but they are not identical.

A SOC 2 report provides an independent examination of controls against the selected Trust Services Criteria, while ISO 27001 certification assesses an organisation’s information security management system against the standard’s requirements.

Neither automatically proves that an AI model is accurate or safe for every use case.

SOC 2 and ISO 42001

ISO/IEC 42001 focuses specifically on AI management systems.

It provides requirements for organisations establishing and improving processes to manage AI responsibly.

For enterprise buyers, ISO 42001 may provide additional evidence about how an AI vendor manages AI-related responsibilities and risks.

However, the scope of any certification should still be checked.

SOC 2 and the NIST AI Risk Management Framework

The NIST AI Risk Management Framework provides a voluntary approach to managing AI risks.

Its Govern, Map, Measure and Manage functions can help organisations structure AI risk assessments.

For enterprise buyers, this means considering the intended use of an AI system, the risks it creates, how those risks will be measured and which controls are needed.

SOC 2 evidence can support this work, but it does not replace an AI-specific risk assessment.

Why Continuous AI Monitoring Matters After Vendor Approval

Vendor due diligence is an important first step, but it does not provide complete visibility into how AI is used after deployment.

An organisation may approve an AI vendor after reviewing its SOC 2 report, security documentation and contractual terms.

However, employees and developers may later use the platform in ways that create new risks.

AI Usage Can Grow Beyond Initial Expectations

A tool originally approved for a small internal team may eventually support several departments.

As adoption increases, more information may be processed and more applications may connect to the service.

This can change the organisation’s risk profile.

Regular monitoring helps teams understand whether usage remains within approved boundaries.

AI Spending Can Become Difficult to Explain

LLM usage may grow through increased requests, longer prompts, larger models and automated workflows.

Without suitable monitoring, organisations may struggle to identify which teams or applications are responsible for rising costs.

Security assurance and financial visibility address different problems, but both matter for responsible AI adoption.

AI Governance Requires Operational Evidence

Written policies are useful, but organisations also need evidence that controls are working.

This may include access reviews, application inventories, usage records, spending reports and incident investigations.

Monitoring supported AI environments can help teams identify unexpected activity and make better operational decisions.

CTA

How WrangleAI Supports Enterprise AI Governance

Enterprise buyers should evaluate SOC 2 for AI vendors before adopting LLM tools, but vendor assurance is only one part of effective AI governance.

Once an organisation approves its AI providers, it still needs visibility into how those services are being used.

WrangleAI helps organisations improve AI usage visibility, cost management and governance across supported AI environments.

Its capabilities include AI consumption monitoring, Optimised AI Keys, spending alerts, usage controls and tools designed to help businesses manage AI resources more effectively.

Bringing Visibility to AI Operations

As businesses adopt several AI models and providers, usage information can become spread across different accounts, applications and teams.

WrangleAI helps organisations bring greater structure to AI consumption and understand where resources are being used.

This supports engineering, finance and governance teams that need clearer information about AI activity and spending.

Supporting Better Financial Governance

AI adoption can create unexpected costs when applications generate large numbers of requests or use expensive models unnecessarily.

WrangleAI’s cost management capabilities help organisations monitor spending and apply suitable controls to supported workloads.

These controls can complement security and compliance reviews by addressing the financial and operational side of AI governance.

Why Vendor Assurance and Usage Monitoring Work Together

A SOC 2 report helps buyers evaluate a vendor’s controls within the report’s scope.

AI usage monitoring helps organisations understand activity within their own connected environments.

Both are valuable, but they serve different purposes.

WrangleAI should therefore be considered as part of a wider AI governance strategy alongside vendor due diligence, security controls, privacy reviews and employee policies.

Organisations should independently verify the security and compliance evidence of every platform they consider, including WrangleAI.

Conclusion: SOC 2 Is the Starting Point, Not the Final Answer

SOC 2 for AI vendors is an important part of enterprise AI procurement. A current SOC 2 Type II report can provide useful independent evidence about a vendor’s controls, but buyers must review its scope, findings and limitations.

Before adopting an LLM tool, organisations should ask how customer information is processed, whether data is used for model training, which third parties are involved and how access is managed.

They should also evaluate AI-specific risks such as prompt injection, agent permissions, unreliable outputs and unexpected spending.

The most effective approach combines vendor assurance with practical security reviews, clear policies and continuous monitoring.

WrangleAI helps organisations strengthen AI usage and cost governance by providing greater visibility and control over supported AI environments. For businesses adopting multiple models and applications, this can support a more accountable and cost-aware AI strategy.

Enterprise AI adoption should not depend on trusting a compliance badge alone. It should be supported by evidence, clear responsibilities and controls that continue working as AI usage grows.

FAQs

Is SOC 2 required for AI vendors?

SOC 2 is not a universal legal requirement for AI vendors. However, many enterprise buyers request SOC 2 reports as part of their security and procurement assessments. The requirement often depends on the buyer’s internal policies, industry and the sensitivity of information being processed.

What is the difference between SOC 2 Type I and Type II?

A SOC 2 Type I report evaluates the design of relevant controls at a specific point in time. A Type II report also examines whether those controls operated effectively over a defined period. Type II reports generally provide stronger evidence about ongoing control performance.

Does SOC 2 mean an AI vendor is secure?

No. A SOC 2 report provides independent assurance about the controls and systems included in its scope. It does not guarantee that a vendor will never experience a security incident or that its AI tools are safe for every use case.

Does SOC 2 cover AI model training?

A SOC 2 examination may include controls relevant to data handling, depending on its scope. However, it does not automatically guarantee that customer data is excluded from AI model training. Buyers should review the vendor’s written data usage terms and contractual commitments.

What should enterprises ask LLM vendors before adoption?

Enterprise buyers should ask about SOC 2 reports, data retention, model training, subprocessors, data location, access controls, prompt injection protection, incident response, audit logs and AI spending controls. The answers should be supported by appropriate evidence.

Is SOC 2 the same as ISO 27001?

No. SOC 2 is an assurance reporting framework based on the AICPA Trust Services Criteria. ISO 27001 is an international standard for information security management systems. They can complement each other but have different requirements and assessment approaches.

How can WrangleAI support AI governance after vendor approval?

WrangleAI helps organisations monitor AI usage, understand spending and apply consumption controls across supported environments. These capabilities can complement vendor security reviews by providing greater operational and financial visibility after AI tools are adopted.

Scroll to Top