Continuous attestation: what audit looks like when agents are doing the work
Published 25 April 2026
The annual audit, in its current form, is built on assumptions that have not been re-examined in decades. The books close monthly. The team produces a set of financial statements. The auditor tests a sample of transactions, evaluates the controls, and signs the opinion. The cycle repeats.
That model was already under pressure before AI agents arrived. Continuous accounting, real-time reporting, and the demand for faster assurance were all eroding the annual model from the outside. The IAASB’s October 2024 Technology Position Statement committed the standard-setter to eight specific technology-related actions for the audit and assurance standards. The FRC followed up with landmark AI-in-audit guidance in June 2025, then in March 2026 published the first regulator guidance globally specifically on generative and agentic AI in audit engagements. (Sources: IAASB Technology Position; FRC June 2025 guidance; FRC March 2026 agentic AI guidance.) The arrival of finance agents that do real work, continuously, against governed data is the development that makes the question urgent rather than theoretical.
This post is the practitioner read on what continuous attestation actually means, what changes about the audit when agents are in the loop, and what a finance function should be doing about it now.
What continuous attestation means in practice
Continuous attestation is the idea that the audit is not a separate event after the period closes. It is a layer that runs alongside the work, with the auditor (or an auditing system) checking outputs against controls as they happen.
The annual audit becomes the formalised sign-off on a continuous assurance process, rather than the substantive testing event in itself.
The model has been talked about for years. The Rutgers continuous auditing research programme led by Miklos Vasarhelyi has been making the academic case since the early 1990s, and Vasarhelyi’s 2024 work on AI-assisted real-time cross-verification (SSRN paper) extends that line into the LLM era. It has not happened in practice at scale for two reasons. The first is that the underlying data was not continuous, governed, and observable at the level the model requires. The second is that the work the audit was reviewing was done by humans on a periodic basis, and the audit reasonably looked like the work it was reviewing.
Both of those reasons are softening. The data layer is improving. The work is starting to be done continuously, by agents that produce structured outputs with full audit trails. The audit can now reasonably begin to look like the work it is reviewing, which is to say, continuous.
The agentic AI piece covers the underlying capability shift. This post is about what changes downstream of it.
What changes for the audit
Three things, in approximate order of impact.
The substantive testing shifts upstream. When an agent reconciles accounts, posts journal entries, and produces variance commentary, the substantive testing that used to happen on the entries themselves can shift to testing the agent’s design. The audit is no longer “sample 25 entries and recompute.” It is “evaluate the skill the agent is following, the connectors it is using, the subagents it is calling, and the audit trail it produces. Then sample the agent’s outputs to confirm the design is operating as documented.”
That is a different kind of audit. It is partly engineering review. It is partly process audit. It is partly traditional substantive testing on a smaller sample. The audit firms that adapt fastest will be the ones whose teams have the technical depth to do the engineering review credibly.
The controls review gets sharper, not softer. This is the bit that surprises finance leaders. The instinct is that AI in the workflow weakens controls because there is a model in the loop. The reality, for a properly designed deployment, is that the controls become more legible and the testing becomes easier, because every step of the agent’s work is logged in a way that a human’s work usually is not.
The audit firm that knows how to read an agent’s audit trail can test more of the population in less time than the audit firm that is still sampling a manual ledger. The function whose audit trail meets that standard is the function whose audit gets less, not more, painful.
The interim audit comes back, in my view. This is the part of the argument that is forward-looking rather than regulator-stated. The interim audit, the pre-year-end audit visit, became less rigorous in many engagements as the focus shifted to substantive testing at year-end. If the substantive testing shifts upstream to continuous review of the agent layer, the interim becomes the place where most of the substantive work happens, and the year-end becomes the confirmation. The current standards do not yet require that. The direction of travel suggests it is where the model ends up.
What the finance function should be doing now
The function does not have to be at continuous attestation in 2026 to act on this. The work that prepares for it pays back regardless. Four things.
Make the agent’s audit trail meet the standard. This is the prerequisite for everything else. The AI governance framework covers what the audit trail needs to contain. The shorter rule: if a competent auditor cannot reconstruct, from the trail, what the agent did and why, the audit trail is not yet adequate.
Document the agent’s deployment design. The skill, the connectors, the subagents, the review layer, the escalation logic, the rollback plan. All of it documented, versioned, and reviewable. The audit preparation post is the long version. The shorter rule: the documentation that lets your internal team operate the agent is the documentation the auditor will want to see.
Talk to the audit partner early. The audit firm’s view on AI in the workflow varies considerably by firm and by partner. The function that finds out where its auditor stands in December is the function that has a difficult January. The conversation should happen in advance of any production deployment, not in response to it.
Invest in your internal audit function’s capability. Internal audit needs to be able to review the agent’s work the same way external audit will. That requires people who understand AI agents at the level of how they work, not at the level of marketing. Most internal audit functions today do not have that capability. The ones that do are buying themselves an option that becomes valuable as the deployment matures.
What does not change
The accountability does not move. The finance team’s signature is on the financial statements. The fact that an agent produced the first draft of the close, or the reconciliation, or the variance commentary does not transfer the accountability to the agent. The audit confirms the function’s representations. The function’s representations are the function’s.
The materiality framework does not move. The agent does not lower the threshold at which an error matters. If anything, the agent’s volume capacity raises the question of what happens when an error replicates at scale. A single agent producing five thousand outputs an hour, with a 3% error rate that the team has not designed the review around, is a different exposure than the same error rate on five human outputs an hour.
The professional judgment requirement does not move. The audit opinion is still a professional judgment by the audit partner. The agent does not produce that judgment, and neither does an agent on the other side of the engagement. The professional judgment of the people in the room remains the foundation of the audit.
What I would do with this
If I were running a finance function that has any kind of finance AI in production this year, I would put continuous attestation on the agenda for the next audit committee meeting. Not as a project to launch this year. As a framing for the conversations the committee will have over the next three years. The Big 4 are already building the tooling they will need. Deloitte’s Zora AI launched in March 2025; PwC’s Agent OS shipped in 2025 with around 25,000 agents reportedly deployed across client operations; KPMG launched Workbench in June 2025 and underwrites it with a $2 billion five-year AI commitment; EY’s agentic platform now reportedly supports 150 tax agents across 80,000 tax professionals. None of these is continuous attestation in production yet. All of them are the infrastructure for it.
The conversation has three parts. What does the function’s audit trail look like today, and what would it need to look like for continuous attestation to be a credible direction? What is the audit partner’s view on AI in the workflow, and is that view stable? What is the internal audit function’s capability to review AI-assisted work, and what is the development plan?
The function that has those conversations early will not be the function that scrambles when the audit firms start to lean into AI-assisted assurance. The function that does not will be.
Where this lands
Continuous attestation is not a 2026 reality for most finance functions. It is also not a hypothetical. The combination of agents doing more of the work, audit firms developing the capability to review AI-assisted outputs, and regulators expecting more frequent and more granular assurance will move the audit model in this direction over the next three to five years.
The finance functions that prepare for it will spend less time on the audit and more time on the business. The ones that do not will find out, around 2029, that the audit they used to have is not the audit they are getting any more, and they are not ready.
The audit model is moving. The work to be ready for it starts now.
Maebh Collins is a Fellow Chartered Accountant (FCA, ICAEW) with Big 4 training and twenty years of operational experience as a founder and senior finance leader.