Conversations about AI agents in continuous audit tend to collapse into one of two unhelpful positions. The vendor position implies an agent that audits your company. The professional-sceptic position is that none of it constitutes audit evidence.
Both are wrong, and the territory in between is more specific than either, and more useful.
Where AI agents in continuous audit genuinely work today
Full-population testing instead of sampling. Traditional audit sampled because examining everything was impractical. For structured data that constraint has largely gone. A sample of forty journal entries supports a statistical inference; testing every entry posted in the year, filtered for entries by unexpected users, unusual account combinations, round-dollar amounts, entries outside business hours, period-end entries reversing shortly after, identifies specific items rather than inferring a rate.
This is not new in principle. Audit data analytics predate the current wave by decades. What language models changed is the cost of building the analysis. Work that needed a specialist and several weeks can now be specified and refined in days, which moves it from something reserved for large engagements to something available on ordinary ones.
Unstructured data, which resisted automation entirely. This is the genuinely new capability. Extracting revenue recognition terms, change-of-control provisions and liability caps across hundreds of contracts. Analysing free-text journal entry descriptions and expense narratives, where the signal depends on meaning rather than wording and keyword search misses it. Comparing a written procedure against evidence of what was actually done, systematically rather than by sampling.
First-draft documentation. Process narratives and control descriptions drafted from interview transcripts and system evidence, then corrected by a person. Unglamorous, immediately valuable, and where most teams should start.
Continuous controls monitoring
Annual testing tells you a control worked on the days you sampled. Continuous monitoring tells you whether it is working now, which matters disproportionately for SOX programmes and SOC 2 observation windows, where a control failing in month two and discovered in month eleven is an exception in the report rather than a fixed problem.
Controls that automate well: segregation of duties conflicts detected as they arise, new vendor creation matched against the payment file and against employee address data, access changes reconciled to approval records, journal entries flagged against defined risk criteria, expense claims tested against policy.
The design constraint that determines success or failure is alert volume. A system generating two hundred alerts a week gets ignored within a month, and an ignored control is worse than no control, because it produces documented false assurance. Tuning to a volume the responsible person can genuinely review, and defining what happens to each alert, is most of the implementation work. It is also the part that gets skipped.
What makes an AI-assisted procedure defensible
Every auditor and regulator will ask the same question: how do you know the output is right? "The model is usually accurate" is not an answer.
What makes these procedures hold up is the same thing that makes any procedure hold up, a controlled, documented, reproducible process:
- Deterministic where it matters. Extraction, filtering and matching logic is code, not a prompt. Models classify and extract; conclusions come from rules producing identical output every run.
- Validated against a known-answer set. Before relying on it, run the procedure against a population already tested manually and measure the error rate in both directions. Without this there is no basis for reliance.
- Human review of every exception. The system identifies; a person concludes. That boundary should be visible in the workpapers.
- Change control over the automation. Which version produced which result, who changed it, when, who approved it. The same discipline applied to any system relied upon.
- A retained trail. Inputs, logic version, outputs and dispositions preserved so the procedure can be re-performed.
- Data governance. What was sent to which model, under what contractual terms. Sending client financial data to a consumer AI service is a confidentiality breach regardless of the quality of the analysis.
What will not survive review
An agent that forms a conclusion. Professional standards require the practitioner to form the opinion, and any product implying otherwise is misrepresenting both the technology and the standards.
A procedure nobody can explain. If you cannot describe to an auditor what the logic does and why it is reliable, it will not be relied upon however well it performs. Explainability here is a practical requirement, not a philosophical one.
Output with no validation baseline. Models produce confident output on questions they have handled poorly; which is precisely the failure mode that matters in assurance. Confidence is not accuracy, and the two are indistinguishable without a known-answer test.
A prompt-and-spreadsheet workflow. Non-reproducible, unversioned, and unable to demonstrate what was actually examined.
Where a mid-market team should start
Duplicate payment detection or journal entry analysis. Both use structured data you already hold, both have clear rules, both replace work someone does manually or does not do at all, and both produce results that are immediately verifiable, which builds the credibility needed for anything more ambitious.
Where not to start: a general-purpose assistant with access to everything. Broad scope produces impressive demonstrations and no measurable outcome, and it makes the data governance question much harder to answer.
The honest summary
AI reduces the cost of finding things by a large factor and changes the process of concluding about them not at all.
That is a genuinely significant shift (full-population testing and continuous monitoring were previously affordable only to large engagements) and it is a smaller claim than the market is currently making. Teams that build on the smaller claim end up with procedures that survive review. Teams that build on the larger one end up rebuilding.
The AI audit automation service page sets out how these engagements are structured, and the finance automation page covers the adjacent question of automating the close itself.
Javed Peeran
Licensed by the California Board of Accountancy and the author of every article published here. Thirty years of practice covering external audit of banks, insurers and mortgage companies, fifteen years as CFO and Corporate Controller inside technology companies, and IT governance and security compliance work spanning SOX 404, SOC 1 and SOC 2, ISO 27001, FISMA, FedRAMP, PCI DSS, HIPAA/HITECH, CCPA and GDPR, plus Oracle ERP migrations and, more recently, generative-AI audit automation.
