Who can audit an agentic compliance program?

Using AI agents to operate parts of your compliance program doesn’t create a fundamentally different category of SOC 2 audit.

Every company already operates differently.

Companies use different infrastructure.

Different security tools.

Different development processes.

Different access models.

Different compliance platforms.

Different controls.

Different evidence.

An AI-operated compliance program is another variation the auditor needs to understand.

But there’s an important distinction between an audit firm being technically capable of performing the audit and being well suited to performing it efficiently and intelligently.

As more compliance activity moves to AI agents, we think that distinction will become increasingly important.

Think about the transition to cloud computing

There’s a useful historical analogy.

When cloud computing became mainstream, cloud-native companies could theoretically use the same audit firms as everyone else.

The underlying assurance standards didn’t suddenly disappear because a company moved from physical servers to AWS.

But understanding the environment mattered.

An auditor familiar with modern cloud infrastructure could more quickly understand what they were looking at.

They understood shared responsibility.

They understood infrastructure as code.

They understood managed services.

They understood cloud-native access and security models.

They knew what evidence mattered.

And they were less likely to ask questions based on assumptions inherited from an older technology environment.

Cloud-native companies therefore tended to get better outcomes from auditors who understood cloud-native companies.

We think something similar is beginning to happen with agentic compliance.

What changes when AI operates the compliance program?

The compliance objectives don’t necessarily change.

If you’re performing a SOC 2 audit, the auditor still needs to evaluate the relevant controls against the applicable criteria.

But the way those controls are implemented and operated may change substantially.

An AI agent might:

maintain compliance documentation;

monitor systems;

collect evidence;

maintain risk and vendor registers;

coordinate compliance activities;

perform first-pass reviews;

identify exceptions;

communicate with employees;

prepare audit information;

and potentially perform certain controls directly.

That creates questions an auditor needs to understand. The practical guide to building an agentic compliance program explains how role, visibility, skills, authority and escalation fit together.

What exactly is the agent doing?

What information does it rely on?

What systems can it access?

What authority has been delegated to it?

How does it identify exceptions?

What happens when it’s uncertain?

What requires human approval?

How is its company context maintained?

How do we know the agent actually performed the activity?

What evidence does it leave behind?

And where does human accountability sit?

These aren’t reasons an agentic compliance program can’t be audited.

They’re things the auditor needs to understand in order to audit it properly.

AI doesn’t remove the need for evidence

One misconception about agentic compliance is that the auditor will simply trust the AI.

That’s not how independent assurance works.

Saying:

“Our AI agent reviews access continuously.”

isn’t sufficient by itself.

The auditor needs to understand the control.

What does “reviews” mean?

Which users?

Which systems?

What information is considered?

What conditions generate an exception?

What happens to the exception?

How is the agent configured?

Who oversees it?

What records demonstrate that the process actually operated?

This isn’t fundamentally different from evaluating automation elsewhere in a technology environment.

The presence of automation changes how the control operates and how it may be tested.

It doesn’t make assurance unnecessary.

But what does it mean for an AI agent to “perform” a control?

This is where things become more interesting.

Many compliance frameworks, audit methodologies and customer expectations developed in a world where humans performed activities.

A manager reviews something.

A person approves something.

Someone signs off.

An employee performs a periodic check.

Now imagine an agent performs the underlying work.

Take an access review.

Traditionally, someone might receive a spreadsheet containing all users with access to a system.

They review the list.

They identify inappropriate access.

They sign off.

Now imagine an agent continuously reconciles:

the HRIS;

the identity provider;

system permissions;

job responsibilities;

and approved exceptions.

It identifies discrepancies, investigates the available context and escalates uncertain cases to the responsible human.

What exactly is the control now?

Is the agent performing the review?

Is the human reviewing exceptions?

Does a human still need to sign off on every population simply because that’s how the control historically worked?

Or is the combination of automated analysis, exception handling and human accountability a better control?

These are questions the profession is going to have to work through.

And the answer won’t always be:

Make the AI reproduce the human process.

Don’t automate the old control just because it’s familiar

This is an important opportunity created by agentic compliance.

If we simply make AI imitate everything humans previously did, we’ll miss much of the benefit.

Imagine a traditional control requires a quarterly user access review.

The easiest AI implementation would be:

Every quarter, generate the user list.

Have the AI review it.

Produce a report.

Have a human approve the report.

Done.

But perhaps the better design is continuous.

The agent monitors relevant identity information throughout the year.

A user changes roles.

The agent identifies access that no longer appears appropriate.

An employee leaves.

The agent confirms access has been removed.

A privileged permission appears unexpectedly.

The agent escalates it.

Uncertain situations go to a human.

Now the company isn’t merely using AI to automate a quarterly control.

It’s reconsidering the control itself.

A good auditor needs to be able to understand that distinction.

Human accountability still matters

Agentic compliance doesn’t mean removing humans from responsibility.

We think the useful mental model is:

AI can be the compliance manager. A human still needs to be the compliance leader.

That applies at the control level too.

How much an agent can perform autonomously will depend on the control, the environment, the risks involved and the reliability of the system.

For one activity, the agent might perform almost everything.

For another, it may prepare the analysis and require a human decision.

For higher-risk decisions, human approval may remain important.

The important question isn’t:

Did a human click approve?

It’s:

Is the control appropriately designed, is it operating effectively, and is there appropriate accountability for the outcome?

Agentic compliance forces auditors and companies to think carefully about that distinction.

The audit firm needs to understand the agent itself

If an AI agent performs meaningful compliance activities, it becomes part of the environment the auditor needs to understand.

That doesn’t necessarily mean auditing the underlying AI model from first principles.

It means understanding the parts relevant to the controls being relied upon.

For example:

Instructions. What has the agent been told to do?

Context. What company information does it use to make decisions?

Access. Which systems and information can it access?

Authority. What actions can it perform autonomously?

Escalation. When does it involve a human?

Change. How are important changes to the agent, its instructions or its context managed?

Evidence. What records exist showing what it did?

Oversight. Who inside the company is accountable for its operation?

The significance of each question depends on how heavily the company relies on the agent.

An AI assistant drafting policies creates a very different assurance question from an autonomous agent performing important controls.

The more autonomous the agent, the more its reliability matters

There’s a spectrum.

At one end:

AI assists a human.

The agent prepares an access review and the human performs the decision-making.

The auditor may largely be concerned with the resulting human control and supporting evidence.

Further along:

AI performs the activity and a human reviews exceptions.

Now the auditor needs greater understanding of how the agent reaches conclusions and determines what constitutes an exception.

Further still:

AI performs the control autonomously.

Now the reliability of that automation becomes much more significant to the effectiveness of the control.

That’s not an argument against autonomy.

Automation can be more consistent than manual human activity.

It doesn’t get tired.

It doesn’t forget that the quarterly review is due.

It can analyse far larger populations.

It can potentially operate continuously rather than periodically.

But those benefits only matter if the automation itself can be appropriately understood and relied upon.

That’s where an audit firm that understands agentic systems becomes valuable.

Why audit methodology matters

There has been another important change in SOC 2 over the last several years.

Compliance automation platforms helped make SOC 2 significantly more accessible.

Audit firms developed around that ecosystem.

In many cases, this created highly efficient audit methodologies built around standardized compliance-platform outputs.

There are real benefits to that.

Standardization makes audits easier to scale.

Auditors know where evidence will appear.

Companies know what information to provide.

Controls can be mapped to familiar templates.

But the trade-off is flexibility.

If the audit methodology assumes the client will arrive with a particular style of control environment and standardized set of evidence, a highly bespoke or agentic compliance program can create friction.

The client might find themselves changing the way they represent their controls simply so they fit the auditor’s methodology.

That’s backwards.

The audit methodology should be capable of understanding how the client actually operates.

Your auditor should be able to work where you work

Imagine your company doesn’t use a compliance platform.

Your policies live in Notion.

Your people processes live in your HRIS.

Your engineering controls live in GitHub.

Your infrastructure controls live in AWS.

Your compliance agent understands the connections between them.

Your evidence is available directly from those systems.

An adaptable auditor should be able to work with that environment.

That might mean working with information in Notion.

It might mean inspecting source systems.

It might mean receiving evidence collected by your agent.

It might increasingly mean interacting directly with your compliance agent to understand the program and retrieve relevant information.

You shouldn’t necessarily have to reconstruct your entire compliance program inside another interface simply because that’s where the auditor prefers to work.

This principle matters well beyond AI.

A good audit should adapt to understand the client’s environment rather than requiring the client’s environment to resemble the audit firm’s template.

AI-native companies may also expect a different audit experience

There’s another aspect of compatibility that isn’t purely technical.

Companies building around AI tend to work differently.

They expect AI to participate in workflows.

They may use agents to research, analyse, coordinate and execute tasks.

They may move quickly.

They may have fewer traditional administrative processes.

They may deliberately avoid manual workflows that exist primarily for compliance.

And increasingly, they may expect their service providers to be capable of interacting with that operating model.

That includes auditors.

If the client’s compliance agent can answer an audit question, locate the evidence and explain the surrounding control in seconds, there is little value in forcing a human compliance manager to manually perform the same administrative task simply because that’s how the audit has historically operated.

The auditor still needs to independently evaluate the evidence.

But the interface can change.

Does an agentic compliance program need a specialist auditor?

Not necessarily.

There isn’t a separate category of SOC 2 called an “agentic SOC 2.”

Using AI doesn’t automatically require a different audit standard.

And companies shouldn’t assume that an auditor needs some special agentic-compliance badge before they can perform the engagement.

The better question is:

Does the audit firm understand how we operate well enough to audit us effectively?

That includes whether the firm:

understands modern AI agents;

is comfortable with controls operated partly or substantially by AI;

can evaluate the associated risks;

understands where human accountability still matters;

is prepared to evaluate evidence produced through agentic workflows;

can distinguish meaningful automation from compliance theatre;

can adapt its audit methodology to the company’s environment;

and is willing to work with the client’s systems rather than requiring everything to be translated into a predetermined template.

That’s a compatibility question rather than a qualification category.

Agentic compliance doesn’t mean lower audit quality

There’s also a risk that “AI-native audit” gets interpreted as “faster and easier audit.”

That’s not what we mean.

AI can make the process dramatically more efficient.

An agent can organize evidence.

It can answer questions.

It can explain the control environment.

It can surface relevant populations.

It can reconcile information.

It can eliminate huge amounts of administrative work.

But efficiency and rigor are different dimensions.

The objective shouldn’t be to use AI to ask fewer meaningful questions.

It should be to spend less time on low-value administration so that more attention can go toward the questions that actually matter.

What does this control really achieve?

What could go wrong?

Does the evidence actually demonstrate what the company says it does?

Is the automation reliable?

Are important exceptions being surfaced?

Do the company’s commitments match its practices?

Would a customer reading this report receive a meaningful basis for trust?

That’s where we think the value of audit should sit.

The auditor shouldn’t be afraid of a different-looking compliance program

Agentic compliance may produce programs that look different from the compliance programs auditors are used to seeing.

That is not inherently a problem.

The company may not have a conventional compliance platform.

Its controls may be embedded more deeply into source systems.

Its agent may maintain much of its compliance context in a Compliance Brain.

Some traditionally periodic controls may operate continuously.

Humans may spend considerably less time performing administrative compliance activities.

Evidence may be assembled dynamically rather than stored throughout the year in a dedicated compliance repository.

The auditor’s job is not to make those things look traditional.

The auditor’s job is to understand what is actually happening and determine whether it satisfies the relevant assurance requirements.

That’s an important distinction.

What should I ask an auditor before choosing them?

If you’re operating an agentic compliance program, there are a few useful questions to ask a prospective audit firm:

Are you comfortable auditing controls performed or supported by AI agents?

How do you approach determining when reliance can be placed on automated or agent-operated controls?

Are you comfortable working with our source systems rather than requiring a particular compliance platform?

Can our AI agent participate in evidence collection and audit coordination?

How do you think about human accountability where an agent performs most of a control?

Will you evaluate our actual control design, or expect us to adopt your standard control template?

Are you comfortable with controls that achieve familiar objectives through non-traditional methods?

The exact answers matter less than the conversation.

You’ll quickly discover whether the auditor is trying to understand your operating model or trying to fit it into theirs.

So, who can audit an agentic compliance program?

In principle, any appropriately qualified audit firm capable of understanding and evaluating your control environment.

You don’t need a separate kind of SOC 2 auditor simply because you use AI.

But just as cloud-native companies benefited from auditors who understood cloud-native environments, AI-native companies are likely to benefit from auditors who understand agentic ones.

The difference isn’t the badge on the audit.

It’s the quality of the interaction.

A compatible audit partner should understand what your agents do, where they’re reliable, where they aren’t, how human accountability works, what evidence matters and how your controls fit together.

They should be prepared to work with your technology rather than forcing you into theirs.

And they should be able to embrace the efficiencies AI creates without confusing efficiency with reduced rigor.

At Throughline, that’s how we’re approaching it.

Agentic compliance doesn’t need a different standard of assurance.

It needs auditors capable of applying the standard intelligently to a different way of operating.