What Does SOC 2 Type II Actually Require for AI Decision Systems?

Auditors now want continuous AI evidence across the whole audit window.

Publication desk · · 11 min read · Updated
Cover illustration for “SOC 2 Type II Evidence Requirements for AI-Assisted Decision Workflows”
Audit-Ready AI Workflows · October 13, 2026 · 11 min read · 2,517 words

If your company uses AI to make or assist decisions, auditors now expect a much longer paper trail than they did even two years ago. Here is what they look for and how to be ready before the audit clock starts.

SOC 2 Type II is a security audit. It checks whether your controls actually worked over six to twelve months, not just on one day. It used to be a badge a vendor could point to and move on. Now it is a gate. No deal gets done, especially when a large buyer is deciding whether to trust a vendor's AI with its customer data. The five Trust Services Criteria are the five areas every auditor checks. They still date to the AICPA's 2017 framework. The AICPA is the main US accounting standards body. What has changed is how auditors read those rules against systems built on AI. On September 11, 2026, the AICPA published TQA Section 9561. That guidance explains how a company's use of AI affects SOC 1 and SOC 2 audits. The numbers show how fast things are moving. The CBIZ 2024 SOC Benchmark Study found that Confidentiality now appears in close to two-thirds of SOC 2 reports. That is up from around a third in 2023. If you run AI in production, treat auditor scrutiny as a risk you face right now, not something to plan for later.

Why the audit window matters more for AI than for ordinary software

A Type I report checks whether controls are well designed at a single moment. A Type II report checks whether those controls actually worked, continuously, across a window that usually runs six to twelve months. That difference has always mattered. But AI widens the gap in a way traditional software never did. A single snapshot cannot show a buyer that an AI prompt log was kept consistently throughout that whole period. A prompt log is a record of every instruction sent to the AI. A snapshot only proves the log existed on the day someone looked at it. An inventory of AI agents, a set of approval records, or a file of decision records built two weeks before the auditor arrives proves nothing about the other twenty-five weeks of a six-month window. Evidence either existed and was collected the whole way through, or it did not. Auditors are increasingly able to tell the difference.

How AI stretches what auditors test under each of the five areas

AI does not add a sixth area to SOC 2. It stretches what gets tested under each of the five that already exist. It pulls in evidence types that many organisations have never had to produce before. Scope follows function. AI becomes relevant to an audit the moment its use touches the system being examined, the promises made to customers, or the controls in place to keep those promises. That means you need to map every place AI shows up before sitting down with an auditor. Check the product, development, support, and how controls themselves get run. Security is required in every audit. It now covers things like who can access the model registry (the library of AI models), how API keys are secured, and the paperwork behind every outside AI supplier. API keys are the passwords that connect services to AI tools. If you decide a particular AI use has no bearing on the system being audited, write down why. Revisit that decision if the data or access involved ever changes. Skipped documentation on an exclusion shows up later as a finding during fieldwork. The sections below walk through each of the five areas in turn, with the specific evidence auditors are asking for under each one.

What auditors look for under Security when AI helps make decisions

Security is the one area every SOC 2 audit includes. AI-related evidence gaps appear there first. The old evidence set still matters. That includes access logs, multi-factor authentication records, and firewall settings. But it no longer covers everything auditors test. Model registry access controls are now part of the standard ask. Auditors want to know who can push a model to production, meaning make it live. They want to know whether an approval step gates that move. They also want to see whether the approver's identity, the timestamp, and the model version are all recorded together. Every outside AI tool that touches customer data counts as a sub-processor. A sub-processor is an outside supplier handling data on your behalf. The contracts behind that relationship need to be documented and reviewed, not assumed. This all comes down to one question: who owns the action an AI system took? Auditors need a named owner for every AI model you use. They also need a paper trail showing someone was watching it, not just that someone switched it on.

Processing Integrity: proving AI decisions came from a valid, traceable process

Processing Integrity checks whether the system does what it is supposed to do, completely and accurately. This is where most AI-assisted decision workflows fail their first audit. The failure is usually not a lack of evidence that AI made decisions. It is a lack of evidence that those decisions came from a valid, traceable, authorised process. Raw AI output is just data. The most common breakdown looks like this: the AI flags something correctly, but the incident record does not explain why a person investigated it or chose to dismiss it. Every AI alert needs to produce a ticket with real investigation notes attached. That creates the documented human review step auditors are asking for, in the form they want it. Organisations that build this into the workflow from the start walk into the audit with a much cleaner file than those who try to reconstruct it after the fact.

Confidentiality and Privacy: contracts, data location, and AI training commitments

Confidentiality and Privacy obligations under SOC 2 now reach every AI tool provider that touches customer data. The contractual proof auditors want goes further than what most vendor agreements currently offer. A reference to a provider's Terms of Service is no longer enough. Auditors want a signed Data Processing Agreement for every AI tool provider in scope. A Data Processing Agreement is a formal contract that sets out exactly how a supplier may use your data. They also want written confirmation that customer data is not used to train the provider's models unless the customer explicitly opted in. The reasoning is simple. Promises about Confidentiality and Privacy made to a customer extend to whatever the AI provider does with that customer's data. If the contract does not say so in writing, the auditor has nothing to accept. One technical setup satisfies more of these requirements at once than any other. That is VPC-isolated deployment, meaning the AI runs inside a private, walled-off section of the cloud rather than on shared public infrastructure. A related approach called Bring Your Own Cloud is gaining ground in regulated industries. With Bring Your Own Cloud, a vendor's software and AI layer run inside the customer's own private cloud environment. This approach is especially common in healthcare, fintech, and organisations under strict GDPR enforcement. Choosing a setup where you can swap out the underlying AI model without rebuilding your compliance case keeps the compliance decision separate from the capability decision.

Availability and agent governance: when AI acts on its own, not just recommends

Once an AI agent moves from making a recommendation to carrying out approved work by itself, the evidence burden grows. An AI agent is a piece of software that carries out tasks on its own, not just produces suggestions. Availability is the area that checks whether the system is reliably up and running. A new category, agent governance, also becomes something auditors test directly. Most organisations have not built this into their control environment yet. The basic Availability evidence has not changed. That includes uptime monitoring, capacity planning, backup and restore testing, and disaster recovery runbooks. What is new is the layer specific to AI. If the only AI provider you rely on goes down, auditors want to see what happens to the service-level promise made to customers. For systems where agents act rather than just advise, the same human-review logic from Processing Integrity applies. Auditors want a clear, written definition of what an agent may do and what it is barred from doing. That definition needs to be agreed before the audit period starts. Good practice means logging every agent action, including its inputs, outputs, and the event that triggered it, into a record that cannot be changed after the fact. Set alerts for anything unusual, such as an unexpected call to an outside service. Keep a step-by-step response plan ready for containment and fixing problems. Auditors now expect runbooks for AI incidents. A runbook is a documented step-by-step response plan. It lets an investigator trace a live incident all the way back to the AI decision that caused it. Auditors want proof that someone followed a documented path from the alert to the root cause, not just proof that an alert fired. Nodes connects company systems, finds problems, and does the work when a person says yes. It includes role-based access controls, approval gates, an agent inventory, telemetry, audit logs, monitoring, and pause and revoke controls. It also provides case-by-case oversight of third-party agents, giving auditors the governance record they now expect as a standard part of fieldwork. All of this needs to be running from the first day of the audit period. A control introduced partway through a six-month window cannot show the consistent operation a Type II report is built to confirm.

What regulated industries like insurance face on top of the standard requirements

Regulated industries carry an evidence burden on top of the SOC 2 baseline. Insurance sits near the top of that list. So much of what AI does there touches decisions about actual people. You need to set the evidence requirements to match the decision boundary, and do that before any technical work starts. An agent that looks up a claim's status and drafts a handoff note carries a far lighter evidence load than a system that recommends coverage terms or routes a claim based on a risk score. A system that feeds into a hiring decision carries a heavier load still. That boundary has to be written down before testing begins, not inferred from what the system happens to do once it is live. Hiring decisions carry a separate regulatory weight beyond SOC 2. None of this involves insurance agents in the human sense of the word. That is a distinct category from the AI agents doing the work described throughout this piece.

Building an evidence system that holds up for the entire audit period

Organisations that walk away from AI-expanded SOC 2 Type II audits with the fewest findings are not the ones that gathered the most evidence during fieldwork. They are the ones that built evidence collection into the workflow itself at the point of deployment. Chain-of-custody metadata is not optional. This is the basic information that proves a piece of evidence is genuine. Any AI-generated evidence needs a timestamp, the version of the source system, the query settings used, and the date a person signed off on it. Without all four, auditors reject the evidence outright. The organisation then has to rebuild it by hand, usually under time pressure. Trying to build that mapping after the fact does not hold up under scrutiny. Start with one named workflow, one accountable owner, and one measurable outcome. Agree on the baseline, the success measures, who owns the measurement, and the time period before deployment begins. Before deployment, work through the plausible ways a workflow could fail. Consider how likely each failure is, what it would cost, how easily it would be caught, and whether it could be reversed. Write that assessment down and revisit it on a regular schedule. A staged approach to rollout helps manage this. Start with investigation and recommendations. Only authorise full independent action once quality and value are shown. This stages the evidence requirements the same way, keeping the early compliance footprint manageable while the system earns its way toward full independent action. Nodes delivers first findings within 72 hours of complete agreed data access. It identifies at least $1M in potential savings for US companies with more than 2,000 employees, or the fee back. It lays out potential business value alongside findings, assumptions, evidence gaps, options, and the cost of acting versus not acting. That structure, with estimates kept distinct from observed results and assumptions named up front, produces the documented baseline SOC 2 auditors need before they can judge whether a control actually worked.

What senior leaders should do before the audit period opens

Most of the decisions that get you through an AI-expanded SOC 2 Type II audit are made months before an auditor shows up. Those decisions belong to business leaders, not engineers. Start by taking an honest inventory of scope. List every AI use across product, development, support, and control operations. Assess whether each one touches the examined system or the promises made to customers. Attach a written reason to any exclusion, because an undocumented exclusion becomes a finding later. Each AI-assisted workflow in scope needs a decision boundary set before the audit period opens. Write down what the system may recommend, carry out, or route and escalate. Name who owns that workflow. Name who holds approval authority. Set out what the override path looks like if something needs to be stopped. Finalise vendor contracts before the observation window opens, not during it. Choose your deployment setup with Confidentiality and Privacy evidence in mind from the start. A private, isolated cloud environment produces data-location and network-isolation evidence that shared infrastructure simply cannot. A setup that lets you swap AI models keeps you free to change providers without rebuilding your compliance case from scratch. Build human review steps into the workflow design itself, not added on top afterward. Every AI alert or recommendation feeding into a material decision needs a ticket, a documented human judgment, and a sign-off. A resolution code standing in for all three is not enough. Start capturing decision records on day one of deployment. A decision record captures the evidence, recommendation, human reasoning, approval, action, and outcome in one place. It serves three purposes at once. It is the Processing Integrity evidence auditors ask for. It is the governance record you need to learn from your own outcomes. And it is the institutional memory that lets future AI systems build on what you already know instead of starting over. Treat the audit period itself as the measurement period. Agree on a baseline, success measures, and a named measurement owner before deployment. Weigh the cost of acting against the cost of delay and failure, including AI usage, human review time, corrections, change management, and ongoing monitoring. Done this way, the same work that proves control effectiveness to an auditor also proves business value to the leaders who have to answer for it.

Sources

  1. SOC 2 Compliance Guide: 2026 Requirements & AI Controls

More in Audit-Ready AI Workflows