Throughline.
    Insights
    Controls in Practice

    AI Controls in Practice: Three Worked Examples, From Risk to Audit Evidence

    Three worked SOC 2 examples: AI-written code, agent-led access reviews and agents changing customer data. Follow each risk through its control, audit test and evidence.

    Paul Wenham
    Updated 10 November 2026

    AI now shows up in SOC 2 environments in three ways: supporting the people who operate the system, performing compliance controls, and acting directly in production. Each changes the risk differently, and so changes what a credible control looks like.

    This guide works through one example of each, from the risk to the control that SOC 2 and enterprise customers would expect, to how an auditor tests it and the evidence that shows it worked.

    The proposed 2027 SOC 2 criteria update would make AI risks more explicit through points of focus. These examples show how to address those risks in practice today.

    These are illustrative control designs and audit procedures, not a mandatory checklist. The final approach depends on your system scope, risks and service commitments. For the broader criteria mapping, see SOC 2 controls and criteria for AI agents in production.

    How should you think about AI in your control environment?

    By how much the AI acts on its own. The more it acts without a person in between, the more the controls have to govern the AI itself, not just the people using it.

    PatternWho actsWho owns the outcomeWhere the risk concentrates
    AI supports operationsA person, using AI outputThe personOver-trust and volume: people approve more, faster
    AI performs a controlThe AI, with a person overseeingThe control ownerMissing context and silent errors in the AI's judgement
    AI acts in productionThe AI, on live systems and dataThe company, and its customersAuthorisation, accuracy and the AI straying from its purpose

    What controls do you need when AI writes code and a human reviews it?

    Worked example 1: AI writes code, a human reviews it. This is the simplest case, and it looks like nothing has changed. A person still owns each change, and a separate person still reviews and approves it before it reaches production.

    But the risk profile has shifted. AI raises the volume of code and the size of each pull request, so reviewers are stretched. People can also over-trust AI output, while their experience tells them where humans make mistakes, not necessarily where AI does. AI can make confident assumptions with nothing behind them, and those are easy to wave through.

    The control framework still fits, but it needs more weight on independent review and the automation around it.

    RiskControlAudit testEvidence
    AI-assisted changes reach production without real reviewBranch protection requires approval from someone other than the account that pushed the changeSample merged changes; confirm the approver differs from the author and approved before mergeRepository settings; pull request history with authors, approvers and timestamps
    Reviewers over-trust AI output, and defects slip throughReview guidance covering AI failure modes; review time scaled to change size; limits on pull request sizeInspect guidance; for a sample, look for substantive review and time-to-approve that fits the change sizeReview guideline; review comments; approval timing
    Volume outpaces human reviewAutomated tests, code quality checks and security scanning as required checks before mergeConfirm checks are mandatory, not advisory; sample changes for passing checksCI configuration; check results on sampled changes
    Insecure code or invented dependenciesDependency and vulnerability scanning, with findings tracked to resolutionSample findings and trace remediationScan reports; remediation tickets

    An auditor will notice patterns that suggest rubber-stamping, such as a 2,000-line change approved within a minute. Relevant criteria include CC8.1 (change management) and CC7.1 (vulnerability detection). Approval timing is a signal to investigate, not proof on its own that a review was inadequate.

    What changes when AI performs a user access review?

    Worked example 2: AI performs a control. Take a user access review. An agent compares access lists against current staff to find people who have left but still have access. In some setups it reports what it finds to a responsible person, who acts on it. In others, it has permission to remove the access itself.

    This can make the control stronger than the human version. AI doesn't get tired or skim a long list, and the owner can focus on the actions that matter. But AI makes mistakes people wouldn't. If someone changed roles or was deactivated and that never reached a system the agent can see, the agent may judge their existing access appropriate. People pick up that context from the office; AI only has what's recorded.

    RiskControlAudit testEvidence
    The agent lacks context, such as an unrecorded role change, and passes inappropriate accessHR records are the source of truth for joiners, movers and leavers, and feed the agentSample role changes and leavers in the period; confirm each appears correctly in the agent's reviewHR change records; agent review outputs
    The agent makes a wrong callThe control owner reviews the agent's findings, and a sample of its 'no issue' results, before signing offInspect sign-off; reperform the review for a sample of accountsReview report; sign-off record
    An agent with write access removes the wrong access or goes beyond its remitLeast-privilege permissions for the agent; every action logged and reviewedInspect the agent's permissions; trace logged actions to review findingsAgent role configuration; action logs
    A change to the agent's prompt, tools or model silently weakens the controlThe agent's configuration is under change managementSample configuration changes for approval and testingChange records for the agent

    The last row is the shift auditors care about most. When AI performs a control, the auditor needs comfort over the AI itself, as with any system: how it's configured, who can change it, and whether its outputs are complete and accurate.

    The same applies when AI drafts board minutes, updates the risk register or performs a risk assessment. Fed the right context by the right people, it can be more thorough than a person. The risks are excessive detail that others struggle to work with, and missing a critical risk that people know about but nobody has written down. The control is a human owner who supplies the inputs and approves the result.

    Relevant criteria include CC6.2 and CC6.3 (granting and removing access), CC4.1 (monitoring), CC3.2 (risk assessment) and CC8.1 (changes to the agent).

    For the operating model, see how to build an agentic compliance program. For the context an agent needs, read Compliance Brain.

    What controls do you need when an agent changes live customer data?

    Worked example 3: an agent changes live customer data. Now an agent runs part of a call centre. It works as a data administrator: when a customer needs a record fixed or updated, it makes the change in live production data while talking to them. It can be fast and effective, and it's now part of your system, acting on your customers' data.

    RiskControlAudit testEvidence
    The wrong people direct changes through the agentA clear authority matrix: who can request and approve each type of change; requesters verified before the agent actsInspect the matrix; sample changes and confirm each requester was authorisedAuthority matrix; change logs showing requester and approval
    The wrong people can use, configure or change the agentAccess to the agent restricted to approved users and isolated to each customer's environment; configuration access reviewedInspect access lists and isolation settings; review access review recordsAccess lists; environment configuration; access review sign-off
    The agent misunderstands a request and makes the wrong changeAudit logging of every change with before and after values; human approval for high-impact changes; a way to reverse changesSample changes and trace each to the request, approval and logTamper-resistant logs; approval records; reversal records
    Prompt injection, or the agent drifting from its purposeScope limited to approved tools, data and change types; guardrails on inputs; alerts on out-of-scope attemptsInspect the allowed scope and guardrails; review alerts and how they were handledTool and permission configuration; guardrail rules; alert and incident records
    Changes to highly sensitive data, or accounts customers haven't approved, causing a breachAgent permissions follow data classification; customers opt in before the agent can act on their accountsInspect permission mapping against classification; sample customer approvalsClassification policy; permission matrix; customer approval records

    The agent also belongs in your system description, with its purpose, boundaries and the controls around it. Relevant criteria include CC6.1 and CC6.3 (access), CC7.2 to CC7.4 (monitoring and incident response), CC8.1 (changes to the agent), and Processing Integrity, Confidentiality and Privacy where they're in scope.

    For choosing those additional categories, see the SOC 2 Trust Services Criteria explained.

    What changes when AI becomes part of your system?

    When an agent works in production, the model provider behind it becomes a technology dependency to assess within your system scope. It's a third party to you and a fourth party to your customers. Customer data, privacy and confidentiality commitments need to be considered through that chain. How the provider is treated in the SOC 2 report depends on the service and reporting approach; it does not mean the provider's own controls are automatically tested by your auditor.

    That brings the model provider into the controls you already run for critical vendors:

    • Due diligence. Review the provider's own SOC 2 report, and map the controls it expects you to operate on your side.
    • Data terms. Confirm in contract how the provider uses, retains and trains on your data, and where that data is processed and stored.
    • Privacy obligations. Where the provider processes personal data on your behalf, cover it in the relevant data processing agreements and subprocessor disclosures so customers can see it.
    • Change monitoring. Track model and version changes from the provider, since they can change how your agent behaves.
    • Disclosure. Describe the AI components in your system description, including how the provider is treated in the report.

    Enterprise customers increasingly ask about this directly. Clear answers about the model provider, and the controls around it, help their risk teams get comfortable. Our guide to getting through enterprise security reviews explains how this supports customer due diligence.

    What questions will your auditor and enterprise customers ask about AI?

    The following eight questions turn the worked examples into a practical review of your own environment. The answers should point to real owners, settings, decisions and records, not just a policy saying AI is governed.

    Where is AI used across your development, operations, controls and product?

    Identify where AI supports a person, performs a control or acts directly in production. Record each material use case, its purpose and its boundaries. That makes the scope visible before you decide what controls and evidence are needed. The ISO 42001 startup guide gives a worked structure for an AI use case and agent register.

    Who owns the outcome of each AI-performed task, and how do they oversee it?

    Name an accountable human owner. Show how that person supplies context, reviews findings and checks that the task worked, including a sample of results the agent marked as having no issue. Delegating execution to AI does not remove ownership of the outcome.

    What can each agent access and change, and who approved that scope?

    Show the agent's permissions, approved tools, data boundaries and authority to act. Trace those permissions to an approval by the appropriate owner. For customer-facing agents, include customer authorisation and isolation between customer environments.

    How are changes to prompts, tools and models controlled and tested?

    Treat consequential changes to the agent as system changes. Keep approval and testing records that show the configuration was checked before use, including whether the change weakened a control or expanded the agent's authority.

    Where do agents get their context, and how do you know it's complete and current?

    Identify the sources the agent relies on and who keeps them current. In an access review, that includes joiner, mover and leaver records. Check actual changes against the agent's inputs and results, rather than assuming a connection to a system makes the context complete.

    What is logged, and could you reconstruct and reverse an agent's action?

    For consequential actions, retain a trace of the request, requester, approval and result. For live data changes, include before and after values and a practical reversal path. Auditors need evidence that can be traced back to what happened, not just an agent's summary. See collecting evidence with connectors, browser agents and a human in the loop.

    How do you detect and respond to an agent acting outside its purpose?

    Define the allowed scope, guardrails and alerts for out-of-scope attempts. Keep records of how alerts were reviewed and incidents handled. Monitoring should show that someone acted on the signal, not only that the system generated it.

    Which model providers are involved, and what do your contracts say about data use, retention and location?

    Identify the providers behind each material AI use case. Check contractual data use, training, retention, processing and storage terms, and align permissions and disclosures with those commitments. Keep the due diligence and data terms available to support customer questions.

    AI can do more and more of the work, but people must own the outcomes. A control is only as good as the context, scope and oversight around the AI performing it.

    Related reading

    Paul Wenham
    Written by
    Paul Wenham

    Former PwC manager, founder of AssuranceLab and former Sensiba partner and Chief Innovation Officer. Now founder of Throughline, focused on compliance and assurance for AI-native companies.

    More from Paul
    Compliance Planning · Audit & Assurance
    SOC 2 Cost and Timeline Benchmarks by Company Size

    Typical SOC 2 costs and timelines for startups, mid-sized companies and enterprises, by report type. Compare planning, readiness, observation, fieldwork and reporting, plus related costs in USD.

    Audit & Assurance
    How Australian Companies Should Choose a SOC 2 Audit Firm

    Australian companies need a SOC 2 report issued under the AICPA standards by a US-licensed CPA firm, delivered by a strong local team in Australian time zones. Here's how the provider options compare, and what to check.

    Compliance Planning · Audit & Assurance
    How Much Does SOC 2 Cost? The True All-In Cost, Category by Category

    A category-by-category SOC 2 budget in USD, covering audit fees, readiness, platforms, penetration testing, consultants, internal assessments and operational costs. Learn what is required, what is optional, and what a quote can leave out.

    Audit & Assurance
    SOC 2 Isn't Pass/Fail: What It's Really For

    SOC 2 is a risk report, not a certificate. Understand how enterprise customers use it, why having a report is only base camp, and how to turn your audit into a trust asset for growing revenue.

    Compliance Planning
    SOC 2 Readiness Assessments: Your Questions Answered

    What readiness should cover, when to involve your auditor, which gaps matter most, and how readiness, Type 1 and Type 2 can work as one continuous engagement. A practical guide to preparing for a report your customers can rely on.

    SOC 2
    What makes a SOC 2 report quality?

    Enterprise buyers use SOC 2 as an initial qualifier, then look deeper. A quality report is what carries you through that assessment and turns your compliance investment into real trust.

    SOC 2
    What makes a SOC 2 audit high quality

    Audit quality isn't service quality, and it isn't only professional standards. The real test is whether a reader of your report thinks: this tells me what I need to know.

    Audit & Assurance
    Getting Through Enterprise Security Reviews

    Enterprise security teams review you in the background and rarely give feedback. What they look at beyond SOC 2, what counts as a genuine pen test, and how a quality report does the heavy lifting.

    Agentic Compliance
    Agentic Compliance: AI-Native, Human-Led

    Compliance from first principles, the 95% problem, the compliance metadata AI needs, rethinking screenshots, spreadsheets and policies, and why audits should meet clients where they are.