Which evidence matters if the product calls a third-party model?
For an application that calls a hosted LLM, ask the auditor to trace one customer request through the in-scope service: what enters the prompt, which context data is retrieved, what reaches the provider, and what your application does with the response. Agree which controls your company operates and how the provider is treated in the system description.
Review the provider’s actual SOC 2 report for coverage of the service you use, exceptions, and customer controls. Its report does not cover your implementation, and it often covers only some plans: see which AI tools have SOC 2 reports, by plan. Our SOC 2 preparation guide for AI companies explains the boundary and evidence questions in more detail.
Which evidence matters if you develop or modify the model?
If you train or fine-tune a model for the in-scope service, ask the firm how it will sample the model and data lifecycle: promotion to production, access to training and evaluation data, and incident handling when a model change affects customers. Agree which evaluations support the selected criteria and your service commitments before the observation period starts.
This page does not have a public, firm-by-firm file of those samples. The shortlist tells you which qualification the firm has. The proposal has to show the sample.
Who can issue an extra AI-governance output?
Name the output the buyer asked for before you pick the firm. Confirm the CPA issuer for the SOC 2 report and the certification body for the ISO 42001 certificate. A firm that offers readiness or referrals may not be the one that issues the certificate. If the buyer requires an accredited certificate, check that ISO 42001 is included in the issuing body’s accreditation scope.
Thoropass covers the SOC 2 examination and markets ISO 42001 certification audits through Thoropass Certification LLC. It has not announced an accreditation body for ISO 42001, so confirm the accreditation body named on any certificate. Accredited bodies such as Sensiba LLP or Schellman are on the full list of accredited ISO 42001 bodies in our directory, which names each one’s accreditor and certifying entity. Surveillance and recertification add cost after the first audit, so budget ISO 42001 certification over three years. Budget and first-audit timing for a young company are on the startup auditor page. How AI changes the examination itself is in the SOC 2 guide for AI companies.
If the product is an AI agent or connects to MCP servers, read what auditors test when AI agents and MCP servers are in scope. AIUC-1 and the firms that can audit it: a separate AI-agent certification that only AIUC issues, with the firms in our directory that AIUC has accredited and those that offer the joint SOC 2 + AIUC-1 engagement.
ISO’s certification guidance states that accreditation is not compulsory. Confirm which certificate and accreditation your buyer accepts before signing.
Should an AI company choose SOC 2, ISO 42001, or both?
Choose SOC 2 first when US enterprise security review is the immediate blocker. Add ISO 42001 when procurement needs a formal AI management system, especially for regulated or European buyers. Coordinated engagements can reuse evidence. The proposal should name the issuer and scope for each output, including any certificate accreditation your buyer requires.
SOC 2 covers the security controls around your product and data. ISO 42001 addresses the management system governing how AI risk is identified, owned, measured, and reviewed. If you are still deciding what belongs in the audit boundary, our SOC 2 implementation guide for AI companies maps model workflows and AI infrastructure to the Trust Service Criteria.
What should AI procurement teams see in the audit scope?
A procurement-ready scope names the AI services customers actually rely on, the customer and training data they touch, and the controls governing model changes, access, monitoring, and incidents. It should also make the boundary between your application and external LLM providers explicit, because a provider's SOC 2 report does not cover your implementation.
Before comparing quotes, give each firm the same workflow and buyer requirements. Ask: which AI features and data stores are in scope, which categories are selected, what output commitments will be tested, and how external providers are treated? Request the proposed evidence sample, report type, observation period, deliverables, and fee inclusions in writing so you can compare the engagements.
These questions use the AICPA’s 2017 Trust Services Criteria with 2022 points of focus. Baker Tilly’s December 2025 AI controls overview explains why Processing Integrity depends on scope and the limits of what SOC 2 says about AI outputs.
When is an AI-specialist SOC 2 auditor worth the premium?
Pay for AI specialization when model behavior changes customer risk, regulated data enters prompts or training pipelines, or buyers ask detailed AI-governance questions. A generalist can still fit a conventional SaaS product with a narrow AI feature, provided the auditor can clearly scope provider responsibilities, data flows, and change management.
Test the distinction before signing. Give each finalist one real workflow and ask what evidence they would sample, what stays outside the SOC 2 boundary, and which output they can issue: the SOC 2 report, an ISO 42001 certificate, or both as separate deliverables. Concrete answers reveal more than an AI industry badge.
Can one audit engagement satisfy enterprise AI buyers?
Agree the outputs the buyer asked for, usually a SOC 2 Type 2 report and sometimes a separate ISO 42001 certificate. A firm may coordinate both services through separately scoped engagements and named issuers. Confirm the report type, observation period, issuing entities, final deliverables, and buyer acceptance in the proposal.
Also ask how exceptions will be explained to customers and whether the firm will support procurement follow-up after issuance. The most useful auditor is not merely familiar with AI terminology; it can produce evidence and language that your security team can defend in a real buyer review.