A 25-person AI agency in Denver had just closed its best quarter ever—three new enterprise clients and a 40 percent revenue increase. Then all three clients sent nearly identical requests during onboarding: please provide your most recent SOC 2 Type II report. The agency had no SOC 2 report. They had strong security practices but no formal audit attestation. One client agreed to proceed with a security questionnaire and a commitment to complete SOC 2 within 12 months. The other two paused their contracts pending SOC 2 completion. The agency scrambled to engage an auditor and spent the next nine months preparing for and completing the audit—nine months during which 800,000 dollars in contracted revenue sat frozen.
SOC 2 has become the de facto standard for demonstrating security and operational maturity to enterprise clients. For AI agencies, a SOC 2 report is not just a compliance checkbox—it is a trust document that tells clients you take the security of their data and systems seriously.
Understanding SOC 2
What SOC 2 Is
SOC 2 is an auditing framework developed by the American Institute of Certified Public Accountants (AICPA). It evaluates an organization's controls related to the Trust Services Criteria: security, availability, processing integrity, confidentiality, and privacy. A SOC 2 report is issued by an independent CPA firm that audits your controls against these criteria.
Type I vs. Type II
SOC 2 Type I evaluates the design of your controls at a specific point in time. It answers the question: are your controls suitably designed to meet the trust services criteria?
SOC 2 Type II evaluates the design and operating effectiveness of your controls over a period of time (typically 6 to 12 months). It answers: are your controls not only well-designed but also operating effectively over time?
Type II is significantly more valuable to clients because it demonstrates sustained compliance, not just a snapshot. Most enterprise clients require Type II reports. However, many agencies start with Type I to validate their control design, then progress to Type II.
The Trust Services Criteria
Security (Common Criteria). The system is protected against unauthorized access, both physical and logical. Security is required for every SOC 2 engagement—it is the foundation on which the other criteria build.
Availability. The system is available for operation and use as committed or agreed. This criterion is relevant if you provide SLA-backed AI services where uptime matters.
Processing Integrity. System processing is complete, valid, accurate, timely, and authorized. This criterion is particularly relevant for AI agencies because it addresses the accuracy and reliability of your data processing and model outputs.
Confidentiality. Information designated as confidential is protected as committed or agreed. This criterion covers client data, proprietary algorithms, business information, and other data that must be protected from unauthorized disclosure.
Privacy. Personal information is collected, used, retained, disclosed, and disposed of in accordance with the organization's privacy notice. This criterion applies if you handle personal information and have commitments about how that information is managed.
Choosing Your Criteria
Security is mandatory. Beyond that, choose criteria based on your services and client requirements.
For most AI agencies, the recommended starting combination is Security, Confidentiality, and Processing Integrity. This combination demonstrates that you protect client data (Security), keep it confidential (Confidentiality), and process it accurately (Processing Integrity).
Add Availability if you provide hosted AI services with uptime commitments. Add Privacy if you handle personal information and want to demonstrate privacy maturity.
SOC 2 Controls for AI Agencies
Common Criteria (Security) Controls
The Common Criteria are organized around the COSO Internal Control Framework and address:
CC1: Control Environment. Demonstrate that your organization is committed to integrity and ethical values, has an active board or governance structure, establishes authority and responsibility, demonstrates commitment to competence, and enforces accountability.
For AI agencies, this includes having a formal organizational structure, documented roles and responsibilities, a code of ethics, and a commitment to security at the leadership level.
CC2: Communication and Information. Demonstrate that you use relevant quality information and communicate internally and externally about matters that affect the functioning of internal controls.
For AI agencies, this includes security awareness training, incident communication procedures, and client communication about security matters.
CC3: Risk Assessment. Demonstrate that you identify and assess risks to the achievement of your objectives, including risks from fraud and changes in the business environment.
For AI agencies, this includes formal risk assessments that cover AI-specific risks such as model compromise, data poisoning, and adversarial attacks.
CC4: Monitoring Activities. Demonstrate that you monitor, evaluate, and communicate internal control deficiencies.
For AI agencies, this includes security monitoring, vulnerability scanning, penetration testing, and internal audits.
CC5: Control Activities. Demonstrate that you select and develop control activities that mitigate risks and deploy them through policies and procedures.
This is the broadest category and covers access controls, change management, logical and physical security, system operations, and risk mitigation.
CC6: Logical and Physical Access Controls. Implement controls that restrict logical and physical access to information and systems. For AI agencies, specific controls include:
- Role-based access control for all systems including AI development platforms
- Multi-factor authentication for system access
- Access provisioning and deprovisioning procedures
- Periodic access reviews
- Encryption of data at rest and in transit
- Network segmentation
- Endpoint security
CC7: System Operations. Implement controls for detecting and managing security events, monitoring system components, and recovering from identified security events.
For AI agencies, this includes security information and event management (SIEM), intrusion detection, vulnerability management, incident response, and business continuity planning.
CC8: Change Management. Implement controls for authorizing, designing, developing, configuring, documenting, testing, approving, and implementing changes to infrastructure, data, software, and procedures.
For AI agencies, this is critical because AI development is inherently iterative. Your change management process must cover model changes, data changes, infrastructure changes, and configuration changes without stifling the experimental nature of AI development.
CC9: Risk Mitigation. Implement controls to identify, assess, and manage risks from business partners and vendors.
For AI agencies, this covers your relationships with cloud providers, third-party AI tools, data providers, and subcontractors.
Processing Integrity Controls
Processing Integrity is especially relevant for AI agencies because it addresses the accuracy and completeness of your data processing. Controls include:
- Input validation to ensure data entering your systems is complete and accurate
- Processing validation to ensure models produce accurate and consistent results
- Output validation to ensure results are complete and delivered to the right recipients
- Error handling and correction procedures
- Data reconciliation procedures
For AI systems, implement model validation procedures that demonstrate processing integrity. Document model accuracy metrics, testing methodologies, and monitoring approaches that verify your models process data correctly.
Confidentiality Controls
Confidentiality controls protect information designated as confidential, including client data, model code, and business information. Controls include:
- Data classification policies that define confidentiality levels
- Access restrictions based on classification level
- Encryption for confidential data at rest and in transit
- Secure disposal procedures for confidential data
- Non-disclosure agreements with employees and contractors
- Confidential data handling procedures for AI development (training data, model artifacts, evaluation results)
The SOC 2 Audit Process
Selecting an Auditor
Choose a CPA firm with experience auditing technology companies and ideally AI or ML companies specifically. Consider:
- Experience with companies similar to yours in size and industry
- Understanding of AI and ML technologies
- Availability and timeline
- Cost (Type I audits typically cost 25,000 to 50,000 dollars; Type II audits typically cost 40,000 to 100,000 dollars)
- Reputation and client references
Readiness Assessment
Before the formal audit, conduct a readiness assessment. This can be done internally or with the help of a consultant. The readiness assessment evaluates your current controls against SOC 2 requirements and identifies gaps that need to be addressed before the audit.
A readiness assessment typically takes 4 to 8 weeks and costs 10,000 to 25,000 dollars if done by an external consultant. The investment is worthwhile—it is far better to identify gaps before the audit than during it.
Pre-Audit Preparation
Based on the readiness assessment, address identified gaps:
- Implement missing controls
- Document existing but undocumented controls
- Gather evidence of control operation
- Train team members on their responsibilities during the audit
- Prepare a controls matrix mapping your controls to SOC 2 criteria
The Audit
Type I audit process. The auditor reviews your control descriptions and documentation, tests the design of your controls, and issues a report on whether your controls are suitably designed. The audit typically takes 2 to 4 weeks of active auditor engagement.
Type II audit process. The auditor observes your controls operating over the audit period (typically 6 to 12 months), tests the operating effectiveness of your controls by examining evidence, and issues a report on whether your controls are both well-designed and operating effectively. The auditor will request evidence throughout the audit period and conduct focused testing during the final weeks.
Evidence You Will Need
Prepare to provide evidence for every control including:
- Policies and procedures documentation
- Access control configurations and access review records
- Change management records (tickets, approvals, test results)
- Incident response records and post-mortem reports
- Vulnerability scan results and remediation records
- Penetration testing reports
- Employee onboarding and training records
- Vendor assessment records
- Risk assessment documentation
- Monitoring dashboards and alerting configurations
- Backup and recovery test results
- Meeting minutes from security reviews
The Report
The SOC 2 report includes:
- Management's description of the system. Your description of the services, infrastructure, controls, and processes in scope.
- Management's assertion. Your statement that the description is fairly presented and the controls are suitably designed (Type I) or suitably designed and operating effectively (Type II).
- Auditor's opinion. The auditor's opinion on whether your controls meet the applicable trust services criteria.
- Description of tests and results. For Type II, the specific tests the auditor performed and the results.
The report may contain exceptions—instances where controls were not operating effectively. Exceptions do not necessarily prevent a clean opinion, but they will be visible to anyone who reads the report.
Common SOC 2 Challenges for AI Agencies
The Experimentation Problem
AI development is inherently experimental—data scientists try different approaches, iterate rapidly, and discard unsuccessful experiments. Traditional change management (formal request, review, approval, test, deploy) does not map well to experimental model development.
Solution: Implement a two-tier change management process. Experimental changes in development environments follow lightweight controls (version control, experiment tracking, peer review). Changes to production systems follow formal change management (request, review, approval, test, deploy). The line between experiment and production change must be clear and enforced.
The Data Problem
SOC 2 requires controls over data access, data processing, and data integrity. AI agencies often work with large, complex datasets that are accessed by multiple team members and processed through multi-step pipelines.
Solution: Implement data governance controls including classification, access control, lineage tracking, and quality monitoring. Automate data access logging. Implement data pipeline monitoring that detects quality issues.
The Evidence Problem
SOC 2 auditors need evidence that controls are operating effectively. AI agencies often lack formal evidence because processes are informal or ad hoc.
Solution: Automate evidence collection from the start. Configure systems to generate and retain audit evidence automatically. Use GRC platforms or structured evidence repositories to organize evidence by control.
Implementing SOC 2 for AI-Specific Operations
Model Development Controls
Document and implement controls specific to AI model development:
- Model development lifecycle with defined phases and gates
- Code review requirements for model code
- Version control for model code, configurations, and data references
- Testing requirements including accuracy, bias, and robustness testing
- Documentation requirements for model design, assumptions, and limitations
- Approval requirements for model deployment
Data Handling Controls
Implement controls for handling client data in AI workflows:
- Data ingestion procedures with validation checks
- Data classification and handling based on sensitivity
- Access controls for training data, evaluation data, and model outputs
- Data retention and deletion procedures
- Data lineage tracking from source through model training to outputs
Production Monitoring Controls
Implement monitoring controls for deployed AI systems:
- Model performance monitoring with defined metrics and thresholds
- Data drift detection
- Anomaly detection in model inputs and outputs
- Automated alerting for performance degradation
- Incident response procedures for model failures
Maintaining SOC 2 Compliance
Continuous Monitoring
SOC 2 is not a once-a-year effort. Implement continuous monitoring of your controls:
- Automated compliance monitoring using GRC tools
- Regular evidence collection rather than scrambling before the audit
- Ongoing risk assessment and control updates
- Regular security training and awareness activities
Annual Renewal
SOC 2 Type II reports cover a specific audit period and must be renewed annually. Plan for the annual audit cycle by maintaining your controls consistently throughout the year, collecting evidence continuously, conducting an internal readiness review 2 to 3 months before the audit period ends, and addressing any identified gaps promptly.
Expanding Scope
As your agency grows, you may need to expand the scope of your SOC 2 report to cover new services, new trust services criteria, or new systems. Plan scope changes during the annual planning cycle and give yourself adequate time to implement controls for the expanded scope.
Your Next Step
This week: Determine whether SOC 2 is a priority based on your client requirements and competitive landscape. Survey your top 10 clients and prospects to understand their SOC 2 expectations. Begin a preliminary self-assessment against the Common Criteria to understand the size of the gap.
This month: Select your trust services criteria. Engage a CPA firm for a readiness assessment or conduct one internally. Develop a remediation plan for identified gaps. Begin implementing the highest-priority controls, focusing on access management, change management, and monitoring.
This quarter: Complete your readiness remediation. If pursuing Type I first, schedule the audit. If going directly to Type II, begin your audit period with all controls in place. Implement a continuous evidence collection process so you are always audit-ready.