What problem does it solve?
An external audit rests on evidence: invoices, contracts, bank confirmations, ledgers and the notes to the accounts. Most of it is checked by people, in samples, under a fixed timetable. The volume grows every year, and the checks are repetitive: match an expense to its invoice, look for liabilities that were not recorded, tick a disclosure checklist against the notes.
Sampling means most transactions are never looked at. Audit regulators also press on quality. The UK Financial Reporting Council published its first guidance on AI in audit in June 2025 and a review of how the six largest firms certify automated tools, because AI tools are, in its words, moving beyond experimentation to becoming a reality in certain audit scenarios. The same pressure applies to public audit offices: Google Cloud's customer story says the Auditoría Superior de la Federación in Mexico received folders and boxes of documents from audited entities and worked mainly on paper until it adopted Google Cloud tools.
How does it work?
- Load the evidence. The platform ingests the ledger, journal entries and the documents the client provides, and an AI model reads unstructured documents such as invoices, contracts and board minutes.
- Run a procedure. The AI carries out a defined step, for example comparing selected expenses with their support or reading payments after year end for liabilities that were not recorded, and records what it checked.
- Flag exceptions. Items that do not match their support go to the auditor with the evidence the AI used. A disclosure analyzer proposes answers to the checklist from the draft financial statements, and the AI drafts documentation and preliminary conclusions for the auditor.
- The auditor decides. A qualified auditor reviews the output, applies professional skepticism, follows up the exceptions and signs off the working paper. The firm certifies the tool before use and monitors it afterwards.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has published a number for the KPIs of this use case yet.
Value drivers: Employee productivity, Risk and loss reduction, Compliance quality.
Indicative value
A mid size audit firm with 40 audit engagements for listed companies
USD 72,000 to USD 864,000
Staff time cost redeployed to judgment work per year
How this is calculated
Formula: engagements * hoursPerEngagement * share * rate. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Audit engagements engagements, engagements | 40 | 40 | The reference firm. |
| Staff hours spent on routine procedures per engagement hoursPerEngagement, hours per engagement | 300 | 600 | Editorial assumption for vouching, reconciliations and checklist work. Replace with your own time records. |
| Share of those hours the AI takes over share, fraction of hours | 0.1 | 0.3 | Editorial assumption. None of the firms on this page has published an audit time saving, so this range is deliberately modest. |
| Cost per staff hour rate, USD per hour | 60 | 120 | Editorial assumption for a blended fully loaded rate, replace with your own. |
What it leaves out: Gross figure. It leaves out the cost of building and certifying the tools, review time spent on AI output and any change in audit fees.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Deloitte
Global · Professional services · 2026
Deloitte announced on 24 June 2026 a unified agentic intelligence network inside Deloitte Omnia, its global audit and assurance platform. The network brings new and existing AI agents under one framework so they can coordinate whole workflows for nearly 85,000 Audit and Assurance professionals. It was built by Deloitte's own engineering teams under its Trustworthy AI framework. Among the examples, the agents execute preliminary procedures, including data extraction, evidence analysis, drafting documentation and preliminary conclusions for professionals to review. The release gives no outcome number.
No outcome disclosed.
EY
Global · Professional services · 2026
The EY organization announced on 7 April 2026 the global roll out of a multi agent framework inside EY Canvas, its Assurance platform, covering the daily workflows of 130,000 Assurance professionals across 160,000 audit engagements. The framework runs on Microsoft Azure, Microsoft Foundry and Microsoft Fabric and is meant to help audit teams orchestrate complex tasks. EY says the roll out follows a period of testing and piloting. EY says it expects the capabilities to support all end to end audit activities by 2028 and that auditors keep professional judgment and skepticism. EY does not publish an audit quality or time outcome in the release.
No outcome disclosed.
KPMG
Global · Professional services · 2025
KPMG International and its member firms announced on 22 April 2025 the deployment of AI agents into KPMG Clara, the global smart audit platform, aimed at more than 95,000 auditors. The first agents support substantive procedures such as expense vouching and the search for unrecorded liabilities and accrued expenses, and a Financial Report Analyzer AI engine produces output that auditors use to complete disclosure checklists. KPMG describes the approach as AI enabled and people powered, with a human in the loop. The release gives no outcome number.
No outcome disclosed.
Auditoría Superior de la Federación
Mexico · Government and public sector · 2025
The Auditoría Superior de la Federación, Mexico's federal audit office, deployed ScriptoMind in early 2025 in the unit that audits federalized spending, which reviews more than 40% of the federal resources transferred to the 32 states and more than 600 municipalities. Google Cloud's customer story says it uses Vertex AI, the Gemini API, Document AI and Vision AI to read, analyze and reconcile the documents that audited entities submit. The story reports that audit processes that took about 10 months now take hours, or even minutes. The figure is Google Cloud's account and is not recorded as a metric because it gives no comparable before and after unit.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Digital ledgers, journal entries and supporting documents from the audited entity
- A library of audit procedures written as steps an agent can follow
- Reviewed examples of past working papers and exceptions for testing
Systems to integrate
- Audit platform and working paper system
- Client data transfer and secure document intake
- Document extraction and OCR
- Firm quality management and AI inventory
Complexity: High
A firm needs the audit platform and client data feeds, a document extraction layer, a tool certification and documentation process that satisfies the audit regulator, and training for auditors. Public audit offices also need secure handling of the documents that audited bodies submit.
- 1
Start with one procedure
Pick a procedure with clear inputs and a clear answer, such as expense vouching or a disclosure checklist, and write it as numbered steps with the evidence required at each step.
- 2
Certify the tool before use
Document what the tool does, its inputs, its limits and how it was tested, to the standard your audit regulator expects. The FRC guidance on AI in audit describes the documentation it looks for.
- 3
Test against past engagements
Run the agent on closed engagements where the exceptions are known, and measure how many it finds and misses before it touches a live audit.
- 4
Keep the auditor in the loop
Every output is a proposal. The auditor reviews exceptions and a sample of items the agent passed, and the working paper shows who concluded what.
- 5
Monitor quality, not only usage
Track errors found in review, rework and regulator inspection findings, as well as licenses and active users.
Guardrails
- The agent proposes and the auditor concludes, and the sign off is always human
- Client data stays in the firm's controlled environment and is not used to train shared models
- Tools are certified and documented before use on a live audit
- Every agent step is logged so a reviewer can reproduce it
KPIs to instrument
- Exceptions found by the agent versus found later by a reviewer
- Staff hours per procedure before and after
- Share of engagements and procedures using the tool
- Review findings and regulator inspection findings that involve the tool
Human in the loop
Auditors review all exceptions and a sample of the items the agent passed, and a partner signs the audit opinion. The firm's quality function reviews tool performance and changes.
Common failure modes
- Usage measured, quality not
- A firm counts teams and engagements that use the tool but never measures what it misses. Add quality measures, such as exceptions found later in review, from the start.
- Over reliance on the output
- Auditors accept a clean result without testing it. Sample passed items and keep professional skepticism in the methodology.
- Documents misread
- Poor scans and unusual contract wording produce wrong extractions. Route low confidence documents to a person.
- Client confidentiality
- Evidence leaves the controlled environment. Keep processing inside approved environments and log access.
What are the risks and rules?
EU AI Act
Minimal risk
A tool that helps auditors check an entity's financial records and documents is not an Annex III use, and a qualified auditor reviews every output. It would become high risk only if the same tool were used to evaluate or monitor the audit staff themselves (Annex III point 4). Audit regulators such as the FRC supervise audit quality and the documentation of the tool rather than classify it.
Guidance
- FRC publishes landmark guidance providing clarity to audit profession on the uses of AI (Financial Reporting Council, Europe). First UK guidance on AI in audit, with a thematic review of how the six largest firms certify automated tools.
Controls to put in place
- Tool inventory with owner, purpose and certification date
- Documented testing and limits per tool
- Logs of every agent step and every auditor decision
- Client confidentiality controls for all evidence
Frequently asked questions
- Which audit firms use AI agents in their audit platforms?
EY announced in April 2026 a multi agent framework in EY Canvas for 130,000 Assurance professionals, KPMG announced AI agents in KPMG Clara in April 2025 for expense vouching and unrecorded liabilities searches, and Deloitte announced in June 2026 a network of AI agents in Deloitte Omnia that executes preliminary procedures for professionals to review. None of the three has published a measured audit time or quality result.
- What do audit regulators expect when AI is used in an audit?
The UK Financial Reporting Council published guidance in June 2025 on documenting AI tools used in audit, and a thematic review of how the six largest firms certify automated tools. The auditor remains accountable for the conclusion, so firms certify tools, document them and review their output.
- Do public audit offices use AI too?
Yes. Google Cloud reports that Mexico's Auditoría Superior de la Federación uses ScriptoMind, built on Vertex AI, Gemini and Document AI, to read and reconcile documents from audited entities, and says audit processes that took about 10 months now take hours, or even minutes. That figure is the vendor's account.
How to cite this page
Blits.ai AI Use Case Library, "AI assisted financial statement audit", last verified 12 November 2026, https://www.blits.ai/ai-use-cases/ai-assisted-financial-statement-audit. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 12 November 2026: First published