The Auditor Still Decides. The Agent Does the Rest.

Every internal audit engagement spends a large share of its calendar on work that isn't audit judgment at all: reading three years of policies and prior reports before scoping a single test, building a risk and control matrix for a process the function has audited five times before, drafting evidence requests, chasing responses across email and SharePoint, and reformatting the same findings into a slide deck the audit committee will see for eleven minutes. None of that is the reason a Chief Audit Executive built the function. All of it still has to happen before the reason — the judgment — gets exercised at all.

The instinct to fix this with AI isn't new, and audit functions have largely acted on it. What's newer, and more revealing, is exactly where that AI effort has landed. A Gartner survey of 743 audit professionals published in August 2026 found that AI use in internal audit is genuinely widespread — but concentrated almost entirely in the tasks at either end of an engagement, drafting and reviewing, while the work that actually consumes the middle of the cycle — testing, sampling, evidence validation, quality assurance — remains largely untouched.

That gap is the starting point for this piece. Agentic internal audit isn't a rebrand of "AI for audit." It's a specific claim: that the same discipline already applied to isolated drafting tasks can be extended, safely and with the auditor still in charge of every conclusion, across the full engagement lifecycle — from reading the policy library to briefing the audit committee.

AI Adoption Isn't Internal Audit's Problem

It would be easy to read "93% of audit teams already use AI" as evidence the profession has moved on from the adoption question. Gartner's own analysts read the same number differently. James Bourke, a director analyst in Gartner's Risk & Audit Practice, summarized the survey's finding directly: audit's current use of generative AI concentrates less on strategic use cases and more on moderate productivity improvements, so despite high adoption rates, it is not generally transforming audit processes or delivering better strategic insight.

The pattern makes sense once you separate the two kinds of work an audit engagement actually contains. Drafting and reviewing text is a bounded, single-session task — the kind a general-purpose AI assistant handles well because it doesn't require persistent context or a connection to live systems. Scoping a risk universe, building a control matrix, testing a full transaction population, and validating evidence against a specific control requirement are different in kind: they require the tool to hold context across the whole engagement, reach into the systems that hold the evidence, and carry findings through from identification to closure. Most audit functions have adopted AI for the first category and stalled at the boundary of the second — not because the second category matters less, but because it requires something closer to a connected workflow than a chat window.

That boundary is exactly where agentic internal audit is built to operate. Not instead of the drafting and review use cases already in place — alongside them, extended into scoping, evidence, sampling, and remediation tracking, where the actual engagement hours go.

What Agentic Internal Audit Actually Means

Agentic internal audit is the deployment of AI agents across the full audit engagement lifecycle, rather than at a single point inside it. In practice, that means agents that read the existing policy library, prior audit reports, and relevant SOPs before a risk and control matrix gets built, so the matrix reflects institutional history instead of starting from a blank template each cycle. It means agents that translate a risk universe into a risk-based audit programme, weighting scope toward the areas where prior findings, control changes, or business change concentrate real exposure. It means an evidence-request workflow that goes out to the right system or process owner automatically, checks what comes back against the specific control requirement it's meant to satisfy — not just whether something was submitted — and flags gaps for a human reviewer rather than assuming completeness.

It also means risk-based sampling that can test an entire transaction population instead of a small manual sample, surfacing exceptions a traditional 25-item sample would likely miss entirely. And it means agents that draft a first-pass observation and root-cause hypothesis from the evidence gathered — not a final finding, but a starting narrative the auditor edits, challenges, and owns before it goes anywhere near a report.

This connects directly to work Crest.Digital has already covered on the controls side. Continuous controls monitoring tests the control environment itself, continuously, across ERP, procurement, payroll, and other operational systems. Agentic internal audit operates one layer above that — it runs the engagement, and can treat CCM's continuous exception data as a direct input into scoping and sampling, so an audit plan reflects what's actually happening in the control environment rather than a snapshot taken months earlier.

What the Standards Already Expect

The profession's own standard-setter has already moved. The IIA's Global Internal Audit Standards, released in January 2024 and effective from January 9, 2025, consolidated the prior International Standards for the Professional Practice of Internal Auditing into a single 120-page framework spanning five domains, 15 principles, and 52 standards — with an explicit emphasis on technology-enabled audit practices, data analytics, and the internal audit function's own governance of the tools it uses. A function still running risk assessments and evidence requests manually isn't just leaving efficiency on the table; it's increasingly out of step with what its own standard-setter expects a modern audit practice to look like.

ISACA's COBIT framework has separately built out continuous assurance and continuous auditing guidance for exactly this shift — testing controls and evidence on an ongoing basis rather than through a periodic, calendar-driven cycle — and the COSO Internal Control–Integrated Framework remains the reference point most audit functions already map their control universe against, which is precisely the structure an agentic risk-and-control-matrix workflow needs to plug into rather than replace.

Read together, the direction is consistent: standards bodies are pointing audit functions toward technology-enabled, continuous, evidence-heavy practice, while Gartner's own survey shows most functions have adopted AI only at the margins of that ambition. The gap between where the standards point and where adoption has actually landed is the opportunity — and the risk, for functions that mistake high AI usage for having closed it.

An 8-Point Framework for Agentic Internal Audit

Extending AI across the full engagement doesn't require reinventing the audit methodology. It follows the same lifecycle every internal audit function already runs — the framework below simply maps where an agent supports each stage, and where the auditor's judgment stays firmly in place.

Point six is where most functions instinctively want reassurance, so it's worth saying plainly: a drafted observation is not a finding. It becomes one only after an auditor evaluates the underlying evidence, applies professional skepticism, and decides the conclusion holds. The agent's draft exists to save the time spent staring at a blank page, not to substitute for the evaluation that gives a finding its authority.

Building the Programme: A Six-Step Delivery Playbook

Crest.Digital positions this work as a configurable "Risk Automation Pod" rather than a bespoke software build — a shared underlying stack of integration connectors, workflow engine, evidence repository, and dashboards, customized around a specific audit function's systems, control universe, and existing methodology.

Starting with one or two engagements rather than a function-wide rollout matters more than it sounds. It lets the validate step do its job — comparing agent output against what an experienced auditor would have produced manually — before the function's confidence in the tool outruns the evidence for trusting it.

The Auditor Still Decides

The question every CAE eventually asks, reasonably, is where the line sits between what the agent does and what stays human. The honest answer is that agentic internal audit is built around that line rather than around erasing it. Agents handle three things well at scale: reading and cross-referencing volumes of policy, prior-report, and evidence text no team could process manually in the same time; testing entire populations instead of small samples; and assembling a first-pass narrative from evidence that's already been gathered and checked.

What agents don't do is decide whether a control gap is significant, how a finding should be rated, whether evidence that technically satisfies a checklist is actually sufficient in context, or what a root cause genuinely is rather than merely correlates with. Those calls require the kind of contextual, organizational judgment that is precisely why the auditor role exists — and precisely why the framework above routes every agent-assisted step back to a human checkpoint before it becomes a conclusion.

This human-in-the-loop discipline is the same one Crest.Digital has built into its wider agentic risk practice — including the audit-trail question that follows naturally once agents are doing more of this work. If an agent drafted the observation, ran the sample, or flagged the evidence gap, an audit function needs to be able to reconstruct exactly what it did and why, which is the specific problem Crest.Digital has addressed directly elsewhere. The ownership question — who is accountable when an agent acts inside a GRC or audit workflow — is covered separately in that companion piece, and the shift from small manual samples to full-population testing specifically is explored further in Crest.Digital's earlier work on AI-driven audit sampling. All three connect back to the same underlying thesis: extending AI further into the audit lifecycle is only defensible if every extension comes with a reconstructable record of what the agent did and a named human who evaluated it.

That same thesis is what separates agentic internal audit from the vendor-questionnaire version of the same problem Crest.Digital has covered in TPRM — a periodic checklist tells you what was submitted, not what's actually true. Internal audit's version of that gap is a drafted finding that was never actually evaluated against the evidence behind it. The fix in both cases is the same: keep the human judgment call intact, and use the agent to make sure that judgment has better evidence in front of it, faster, than a manual process could ever assemble.

Frequently Asked Questions