June 7, 2026 — The American Institute of Certified Public Accountants (AICPA) today published a new auditing framework, informally called "SOC for AI," aimed squarely at enterprise AI software and services. The framework defines a baseline set of operational controls — spanning data provenance, model lifecycle governance, logging and explainability — that auditors can test and vendors can attest to when selling AI-enabled workplace software.

What SOC for AI covers

According to the AICPA guidance released this morning, the framework adapts the structure of existing SOC reporting (used widely for security and availability attestation) to the specific risks and operational realities of AI systems used in enterprise settings. Key control areas include:

  • Model lifecycle governance: processes for design, training, validation, deployment and decommissioning of models, including approved data sources and versioning.
  • Data provenance and lineage: evidence that training and input data were allowed under contracts and privacy law, and a traceable chain from raw data to model artifacts.
  • Operational logging and audit trails: per-inference metadata, RAG (retrieval-augmented generation) fetch records, and retention policies to support post-incident review.
  • Bias and performance testing: documented fairness assessments, accuracy metrics by cohort, and procedures for remedial retraining or mitigation.
  • Access and change controls: least-privilege access to model weights, training data, and operational endpoints, plus review of human-in-the-loop approvals.
  • Monitoring and incident response: SLAs and playbooks for model drift detection, erroneous outputs, and rapid rollback.

How SOC for AI differs from SOC 2

While SOC for AI mirrors SOC 2’s emphasis on controls and auditor attestations, it introduces new artifacts specific to AI operations — model cards, lineage graphs, test harness outputs, and per-request provenance logs. The AICPA framework does not replace jurisdictional regulation (for example, data protection laws or emerging AI-specific statutes), but it aims to create an industry-recognized baseline auditors can evaluate consistently across vendors.

Industry response: vendors, buyers and auditors

Enterprise software vendors and marketplaces responded quickly. Several large SaaS vendors told industry outlets this week they will move to include SOC for AI attestation in their compliance portfolios, citing procurement demand from regulated customers in finance, healthcare and government. For buyers, the framework promises to simplify vendor due diligence: procurement teams will be able to request standardized attestations rather than bespoke technical questionnaires.

Audit and compliance firms framed the development as a necessary step toward operationalizing AI governance. "Enterprises need objective, audit-ready evidence about how models are built and operated," said a partner at a Big Four firm in emailed comments. "SOC for AI gives auditors standardized checkpoints and creates a clearer procurement signal for mature vendors."

Privacy and civil-rights advocates cautiously welcomed the move — but urged stronger enforcement. "A voluntary attestation framework helps baseline vendor behaviour, but attestation quality and sampling matter," said a privacy researcher who tracks algorithmic fairness. "Without independent validation and regulator oversight, attestations risk becoming check-the-box exercises."

Practical consequences for enterprise IT

For CIOs and procurement teams, the new framework has immediate tactical implications:

  1. Update vendor checklists: Add SOC for AI attestation as a procurement requirement and require copies of the auditor report rather than vendor self-declarations.
  2. Inventory AI assets: Map which SaaS modules and internal tools use models that will be covered by SOC for AI controls (e.g., recommendation engines, automated document processing, LLM assistants).
  3. Negotiate contractual artifacts: Require access to model cards, lineage logs and sample per-inference metadata as part of contract deliverables and incident response clauses.
  4. Pilot attestations: Ask key vendors to pursue SOC for AI reports on a limited set of products to validate the report format before broad rollouts.

What vendors must build

To obtain an attestation under the new guidance, vendors will likely need to operationalize several engineering capabilities that many early-stage AI products still lack:

  • Automated lineage capture across training pipelines and production inference paths.
  • Per-inference provenance logging that records retrieval context (documents fetched), prompt templates, model version and confidence metrics.
  • Repeatable fairness and performance test harnesses with retained audit artifacts.
  • Controls to restrict who can deploy model updates, with immutable deployment logs and rollback ability.

Limitations and next steps

SOC for AI is a starting point rather than a silver bullet. It establishes auditable practices but does not mandate specific remediation levels or technical thresholds. Enforcement will depend on the rigor of audits and whether procurement teams refuse to accept lightweight reports.

The AICPA said it will publish additional guidance for auditors in the coming months, and plans to run pilot attestations with willing vendors and auditors before wider implementation. For enterprises, the arrival of SOC for AI reduces ambiguity around vendor claims, but it also raises the bar for smaller vendors that must invest in observability and governance tooling to remain competitive.

For the AI-enabled workplace, SOC for AI could become the compliance instrument that bridges legal requirements, procurement demands and engineering reality — provided auditors and buyers insist on meaningful evidence rather than checklist language.