SOC 2 and Your Digital Workforce: How AI Agents Fit the Trust Services Criteria :

Why agents now show up in SOC 2

Two things changed. Companies started putting agents to work inside the services their customers depend on: drafting responses, updating records, preparing reports. And customers started asking about it in security reviews. Auditors are paying attention too. In September 2026 the AICPA published technical guidance on how a service organization's use of AI affects SOC 1 and SOC 2 examinations. The practical point for a growing company is simple: if an agent does work inside the system your report describes, describe the agent, control it, and be ready to show the evidence.

Pick categories by what you promise

The Trust Services Criteria, currently the 2017 criteria with points of focus revised in 2022, cover five categories: security, availability, processing integrity, confidentiality, and privacy. Security, covered by the common criteria, is the foundation of every SOC 2 report. Add the others when they match commitments you make to customers. For agents, processing integrity deserves a closer look than it used to. If customers rely on what your agents produce, such as calculations, extracted data, or decisions they act on, your commitments about accuracy are exactly what that category tests.

The criteria that matter most for agents

Most of the work lands in five places. None of it is exotic, but each needs a specific answer for agents.

• Logical access. Agents should never run on a shared account or an engineer's login. Give each its own identity and a named owner, and remove it when the agent is retired.
• Monitoring. Log each agent's requests, tool calls, results, and approvals, and alarm on unusual patterns such as a jump in tool calls or repeated failures.
• Incidents. Add agent scenarios to your incident plan: an agent redirected by hostile content, an agent that sent data to the wrong place, and a model change that altered behavior.
• Change management. An edit to an agent's instructions is a change to production. Review it, test it against known cases, approve it, and record when it went live.
• Vendors. Model providers and tool services see your customers' data. Assess them like any critical vendor.

Model providers are vendors, and often subservice organizations

When an outside model provider processes data as part of your service, it may be treated as a subservice organization in your SOC 2 report. Many reports use the carve-out method, which means your report excludes the provider's controls, and readers look to the provider's own report for them. That has two consequences. You need to obtain and review the provider's SOC 2 report each year. And that report will list controls the provider expects its customers to run, such as managing API keys and configuring data retention. Those become your controls, and your auditor may test them.

Human approval and processing integrity

Auditors test whether processing is authorized and accurate. For agents, the cleanest way to show both is the same tiering used everywhere else. Routine work runs against written acceptance checks and is logged. A sample of medium-risk work is reviewed by a person each week. Consequential output, such as anything sent to a customer, anything that changes their data, or anything financial, waits for a named person. Record approvals, changes, and rejections with the action. Over a Type II period, that record shows the control operating, which is what the report is for.

A practical order of work

For a company that already has a SOC 2 report and is adding agents, or is preparing its first report with agents already in the product, this order keeps the work manageable.

• List every agent in the system description, with its job, owner, identity, tools, and data.
• Move any agent running on a shared or personal account to its own identity.
• Put agent instructions, tools, and model versions under change management.
• Turn on step-level logging with approvals linked to actions, and set alarms.
• Collect your model providers' SOC 2 reports and map the controls they expect from you.
• Tell your auditor early. Agents are new to many engagements, and surprises cost more than conversations.

If you are the buyer

When a vendor hands you a SOC 2 report and says its product uses AI agents, a few questions tell you whether the report covers them.

• Are the agents named in the system description, and is the model provider listed as a subservice organization?
• Does the report test change management for agent instructions and models, or only for code?
• Which agent actions affecting your data wait for a person, and is that tested?
• What does the model provider keep, for how long, and does it train on your data?

How LYFYE helps, and what we do not do

LYFYE designs and builds digital workforces, and we build them to be described and tested: their own identities, change-controlled instructions, step-level logs, approval records, and documented model providers. We do not perform SOC 2 examinations or issue reports. A licensed CPA firm does that. Our job is to make the agents easy for them to understand and test.