Third-Party ISO Certification Body in UK, Europe, UAE, MENA & Globally. (MENA HO: UAE)
SCS KNOWLEDGE CENTRE

SOC 2 for AI Enabled SaaS | Scope and Evidence | SCS

Prepare SOC 2 evidence for AI-enabled SaaS, model providers, customer data and service changes. Discuss the actual service scope and requirements with SCS.

  1. Home
  2. Knowledge Centre
  3. SOC 2 for AI Enabled SaaS | Scope and Evidence | SCS

SOC 2 for AI-Enabled SaaS and Managed Services — Assessment Scope and Control Evidence

SOC 2 for AI-Enabled SaaS and Managed Services — Assessment Scope and Control Evidence
Prepare SOC 2 evidence for AI-enabled SaaS, model providers, customer data and service changes. Discuss the actual service scope and requirements with SCS.

SOC 2 for AI-Enabled SaaS and Managed Services — Assessment Scope and Control Evidence

An existing SaaS service gains an AI-assisted feature. Customer information now passes to a model provider, a new log records prompts, and some outputs require human approval before they affect a customer workflow.

For SOC 2, explain how that change affects the described service, commitments, dependencies and controls. The assessment question is not answered by labelling the product “AI powered.” It needs a clear account of the activities and evidence within the proposed engagement.

This article concerns service-control evidence for an AI-enabled offering. A separate AI management-system requirement should be discussed through the established SCS ISO 42001 service page.

Identify the AI activity inside the service

Describe what the feature does and how its output is used. Identify the service owner, operating team, external providers and important customer commitments. A feature that drafts a response can have different operational implications from one that initiates a transaction.

Map inputs, processing, storage and outputs at a level that explains relevant information flows. Include prompt histories, support records or monitoring data where they affect the described operation. Avoid sending sensitive production prompts through an unrestricted enquiry channel.

Review boundaries and provider responsibilities

A model supplier may operate the underlying model while the service organisation controls integration, customer access, application permissions and use of outputs. Record the actual division and the information available about the supplier's service.

Do not assume that a provider's report covers the integrating application's operation. Equally, do not describe a colocation facility as governing a customer's AI model simply because it hosts equipment. The appropriate boundary depends on what the organisation actually does.

The AICPA Description Criteria resource provides context for management's system description. Explain material AI-related activities as part of the service rather than treating them as a disconnected marketing appendix.

Connect the change to control evidence

AI-enabled service activity

Information to retain

Assessment question

Feature release

Purpose, review, tests and authorisation

What was approved for production?

Model provider change

Notification, impact review and decision

What changed in the dependent service?

Customer-data handling

Flows, permissions, logging and retention

What information reaches each party?

Human approval

Responsibility, review record and exceptions

How did the described check operate?

Output-related incident

Investigation, customer effect and follow-up

How was the relevant service issue handled?

 

These examples are not a universal AI control catalogue or a claim that each item is required in every SOC 2 engagement. The relevant commitments, criteria and described controls need practitioner evaluation.

Keep model and integration changes visible

Retain the feature version, provider service, relevant configuration and effective dates where they matter to the operating explanation. If a provider changes behaviour or availability, document what management evaluated and decided.

Tests should be related to the service purpose and change being assessed. Explain the conditions, results and limits of a test rather than presenting one successful example as proof of every future output.

NIST's AI Risk Management Framework offers useful contextual guidance for AI risk considerations. It does not turn a SOC 2 report into an assurance opinion on every ethical, safety or performance characteristic of an AI model.

Trace customer information through prompts and logs

Identify what users can submit, what the service transmits, what is retained and who can access it. Compare supplier arrangements and actual application behaviour. Record exceptions and decisions where the intended handling differs from observed practice.

Relevant controls may concern access, confidential information, privacy, changes, availability or processing. Category selection still depends on the actual engagement. Use the SCS evidence checklist to organise records without creating a second generic documentation inventory.

Explain the role of human review

If human approval is described as a control, identify who performs it, what they review, the relevant authority and what record is retained. An optional user warning and a recorded approval step are different operating arrangements.

Where the output supports a customer decision, explain who retains responsibility for that decision and any limits communicated through the service. Do not claim that a reviewer guarantees every output is correct simply because a person is involved.

A fictional AI-assisted support feature

A Malaysian SaaS company adds AI-generated draft replies to its support platform, with development support in India. Staff review drafts before sending them to customers. A provider update changes the draft's handling of account context.

The company records the update, relevant evaluation, approval and deployment decision. It checks prompt/log handling and whether staff-review records support the control it describes. An identified incorrect draft is investigated with its actual effect and response retained.

The example shows a service-evidence question, not a model certification result. The Malaysia and India SOC 2 guides provide wider country context.

Keep supplier due diligence and assurance purposes clear

For the separate task of evaluating a purchased AI supplier, use the existing SCS AI-vendor security and privacy guide. That buyer task differs from describing your own AI-enabled service for examination.

If an enquiry also requests ISO/IEC 42001, clarify both outputs and roles. ISO's official overview describes an AI management system; a SOC 2 report has a different reporting purpose. Neither should be presented as a guarantee that every model output is safe or error-free.

Discuss the proposed AI-enabled service with SCS

Provide the service, AI activity, data-flow summary, provider dependencies, changes and current evidence status. Identify the customer requirement and any existing assurance.

Contact SCS to discuss the proposed scope, requested support and independent examination arrangements. Keep the enquiry tied to actual service responsibilities rather than the presence of AI terminology alone.

Share this article

Need ISO Certification for Your Business?

Speak with our certification specialists to understand certification requirements, audit process, implementation timelines and accredited certification services.

Frequently Asked Questions

Describe the change and its effect on the service for practitioner review rather than assuming the old description remains sufficient.
State what the feature does, how outputs are used, who operates it and which important dependencies support it.
Not automatically. Explain the activities and responsibilities your organisation actually performs.
It may support a dependency review, while your integration and retained controls need their own explanation.
Consider them where they affect relevant data handling within the described service.
They may. Review actual content, permissions, supplier handling and retention rather than assuming prompts are harmless metadata.
Record the change, affected functionality or responsibilities, management evaluation and resulting decision.
Retain relevant versions and dates where needed to explain the service and control operation.
No. Explain the test's purpose, conditions, results and limitations.
Connect purpose, review, relevant testing, authorisation and the actual production change.
Inform the practitioner and explain dates, changes, controls and available operating evidence.
It is an input, alongside the actual service, risks and engagement basis; avoid deriving all controls from a marketing label.
Identify the responsible role, review activity, authority and record showing how the described check operated.
No. Explain the actual operational step rather than treating different arrangements as interchangeable.
No. Review what the control is designed and evidenced to achieve within its stated limits.
Retain relevant investigation, decisions, customer effects and actions under the service's arrangements.
Yes. Their authority and effect on the service can differ materially.
Explain commitments and actual checks, approvals or reconciliations relevant to the service and proposed coverage.
Do not assume that. Proposed categories require consideration of the service, commitments and engagement.
Yes, as contextual guidance; it does not automatically become the complete SOC 2 assessment basis.
No. Describe the report's actual scope rather than extending it to every AI characteristic.
No. Clarify the separate management-system and service-control assurance requirements.
Review consistency where those arrangements affect the described service and customer commitments.
Evaluate the actual effect, dates and responsibilities, and retain relevant decisions and actions.
It may describe intentions; actual control operation requires appropriate evidence of its own.
Record its status and any remaining data, access, dependencies or customer responsibilities relevant to the service.
Explain their activities and access where they affect the scoped offering.
No. Agree a suitable secure method and share only information relevant to the discussion or assessment.
No. Confirm the applicable engagement and responsible independent practitioner.
Provide the offering, AI use, dependencies, information flows, changes, evidence status and actual customer requirement.