A SOC 2 report tests whether a service organisation's controls work, and competent people are one of those controls. Teams running AI agents should expect an auditor to sample who released agent actions, changed agent briefs or held agent credentials, then check each person was trained and authorised at the time. Evidence must be dated and linked to access.
What does SOC 2 ask of the people who run agents?
SOC 2 is an attestation report in which an independent auditor reports on a service organisation's controls relevant to security, availability, processing integrity, confidentiality and privacy. The organisation chooses which categories are in scope; security is always included. A report can cover the design of controls at a point in time, or their design and operating effectiveness over a period. Customers usually ask for the second.
The criteria behind the report reach people directly. In substance, they expect the organisation to show a commitment to competence, attracting, developing and retaining people able to carry out their responsibilities; to define and communicate responsibilities; to train staff on security; to control logical access; to manage changes; to monitor systems and respond to incidents; and to manage risks from vendors.
An auditor tests these by sampling. For a training control, that means taking a list of people in scope, picking a sample, and checking each person's record. Agents add new populations to sample from. The general method is set out in how to prepare training evidence for an AI audit.
Who does it apply to?
Service organisations whose systems handle customer data or perform services customers rely on: software and platform providers, outsourced operations firms, payment and data processors, and any business whose customers ask for assurance. If your service includes agents, or your team runs agents to deliver it, they are part of the system description and the controls around them are in scope.
Within the organisation, the people involved are workspace operators, agent owners and builders, approvers of held actions, engineers who manage agent connections and credentials, incident responders, and vendor managers overseeing model providers.
What competence does SOC 2 imply when agents act?
Role-based competence. Each role touching agents needs defined responsibilities and training to match. An operator who releases held actions must understand the agent's certified depth, the depth each action needs, its sources and the four verdicts. A builder must understand how depth is earned through certification on real cases. An engineer must understand the privileges an agent's connections carry.
Access tied to training. The strongest control links access to completed, current training. Release rights at a given depth, edit rights to an agent's brief, and access to agent credentials should be granted only after the relevant assessment, and removed when it lapses.
Change management. Changing an agent's brief, sources or bounds is a change to the system. It should go through the same approval and testing as any other change, by people trained to assess agent risk.
Processing integrity. Where processing integrity is in scope, the auditor will want evidence that agent outputs are complete, valid and accurate. Certification evidence helps: Attest replays an agent against real cases and scores every answer against a floor, and certificates lapse after 90 days. Operators must know what certification status means for the work.
Incident response. Teams must recognise agent incidents and follow the response process. Every verdict is filed in an append-only record, which is what responders and auditors will both use.
Vendor oversight. Model and platform providers are vendors. Vendor managers need to know what data reaches them and what assurance they provide.
How do you meet the criteria and leave evidence that survives a sample?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Commitment to competence | Role definitions and training for each agent role | Role descriptions; training records by person |
| Access tied to competence | Release, edit and credential access granted after assessment | Access grants referencing assessment IDs and dates |
| Security awareness | Agent-specific security module | Completion records; acknowledgements |
| Change management | Agent brief and source changes approved and tested | Change tickets with approver and training status |
| Processing integrity | Agents certified on real cases; recertification tracked | Certification history |
| Incident response | Responders trained on agent incidents | Incident records; exercise results |
| Vendor management | Model provider assessments | Vendor files |
| Traceability | Every verdict filed with the releasing person | Append-only decision record |
The test an auditor applies is simple: for a sampled release, was the person trained, assessed and authorised on that date? If the answer needs a spreadsheet hunt, the control is weak. The case for treating records with that rigour is made in training records are evidence; treat them that way, and a ready structure is in the training evidence pack template for AI audits.
What will a SOC 2 auditor ask?
- Give us the population of people who released agent actions during the period.
- For this sample, show training, assessment and access grant dates.
- Who changed this agent's brief, and who approved it?
- Who holds credentials for agent connections?
- How are agents certified, and what happens when a certificate lapses?
- How do you oversee your model provider?
What does SOC 2 not cover?
It is an attestation, not a certification or a regulation. It does not set training content. It does not cover controls outside the system description. It does not address legal duties such as data protection law. And for EU financial entities, DORA sets its own ICT risk and third-party requirements; see DORA and ICT training for teams running AI agents. The guide to AI literacy and certification covers the wider model.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where the Academy fits
Pryme Intelligence Academy produces the kind of evidence a SOC 2 sample needs. Certificates carry an issue date and expiry, are issued through VerifyDoc.ai and can be checked by number or QR code. The admin view shows each person's progress, results, certificates and expiry dates, so an organisation can match a sampled release to a current certificate in one step.
Questions people ask
Does SOC 2 require AI-specific training?
It requires competence for each role. Where roles involve agents, training must cover them for the control to hold up.
What sample size will the auditor use?
That depends on the auditor's approach and the population size. Plan for every person in scope to be sampled.
Is a certificate enough evidence on its own?
It helps, but the auditor will also want to see that access was granted after the certificate and removed when it lapsed.