Can AI Improve Advisory Quality?
AI makes a competent first draft cheap for every firm at once. That lifts the floor of advisory work, but it also moves the point of difference to the judgement applied on top. The evidence on that floor is thin, the quality effect is mostly unmeasured, and the failures so far are failures of checking.
On this page
Better: Note and confidence
Low confidenceLifts the baseline; the effect on deliverable quality is barely measuredA peer-reviewed audit study links AI investment to fewer restatements, but it covers 2010 to 2019 and machine learning, not generative tools, and it is observational. The UK audit regulator found the largest firms did not formally measure the quality impact of their automated tools. We found no independent study of generative AI and the quality of consulting or audit deliverables.Faster: Note and confidence
Moderate confidenceDrafting is faster; checking is notThe controlled experiments on drafting speed are covered in our delivery-costs review and not re-examined here. In our reading, the checking a client relies on does not compress at the same rate.Cheaper: Note and confidence
Low confidenceA small fee effect in one audit studyIn the same audit study, AI investment was associated with 0.9% lower audit fees. The Deloitte refund to the employment department shows the cost when unchecked output reaches a client.Safer: Note and confidence
Governance-gatedOnly where a named person checks every sourceFabricated citations reached clients in published reports in Australia and Canada. The UK regulator states that the human auditor is always accountable, and the APS AI Plan will keep consultants responsible whether or not they use generative AI.The problem, in numbers
Advisory quality is hard to see from outside, which is why inspections matter. In 2024-25 ASIC inspected 10 company audit files across eight firms. It chose them because its reviews of financial reports had raised concerns, so in our reading they are not a representative sample. Seven of the 10 had at least one audit finding, meaning that in ASIC's view the auditor did not obtain reasonable assurance that the financial report was free of material misstatement.Source 1 ASIC notes that a finding does not necessarily mean the report was misstated. Most findings sat in revenue and receivables, impairment and asset values, and investments and financial instruments.Source 1
In our reading, consulting has no comparable inspection regime, so its failures surface when a reader checks the footnotes. As The Register reported, a review Deloitte produced for the Department of Employment and Workplace Relations contained fake citations and a made-up quote from a Federal Court judgment, spotted by a University of Sydney academic after publication, and Deloitte agreed to repay the final instalment of its contract.Source 2 The corrected report, published by the department, discloses a tool chain built on Azure OpenAI GPT-4o and states that the corrections do not affect its findings or recommendations.Source 2 As PS News reported from Senate estimates, the repayment came to $97,587 of a fee of about $440,000.Source 2 Our delivery-costs review covers its cost; here the question is quality.
What improves, and why
In our reading, synthesis tools change advisory work in two ways. They make a competent first draft cheap to produce. In audit, the UK regulator notes, machine learning can analyse full populations of journals or revenue transactions.Source 6 Both lift the baseline. Neither supplies the judgement a client pays for.
In our reading, the best evidence on audit quality is an association study that predates generative AI. A peer-reviewed study in the Review of Accounting Studies used more than 310,000 employee resumes at the 36 largest US public accounting firms, from 2010 to 2019, to measure investment in staff with AI skills such as machine learning. A one-standard-deviation change in recent AI investment was associated with a 5.0% lower likelihood of a restatement.Source 4 That is an observed association, not a trial. The 17 audit partners the authors interviewed said the main aim of AI in audit was quality, through anomaly and fraud detection, risk assessment and moving people to higher-risk areas.Source 4
For written work, the experiment with Boston Consulting Group consultants covered in our delivery-costs review found the largest gains among lower performers. A peer-reviewed experiment in Science Advances suggests what happens to the ceiling. Of 293 writers recruited online to write an eight-sentence story, those offered GPT-4 ideas produced stories rated more creative and better written. Among the less creative writers, five AI ideas lifted novelty ratings by 10.7% and usefulness by 11.5%. But the AI-assisted stories were more similar to each other than stories written alone.Source 5 That is fiction, not advice, but in our reading the pattern carries: when every firm drafts with the same tools, the draft stops being what sets one firm apart.
When every firm can produce a competent first draft, the draft stops being the product. The client is paying for the judgement about which parts are true and which parts matter.
What the evidence actually says
Firms are not yet measuring the quality effect. In June 2025 the UK's Financial Reporting Council reviewed how the six largest audit firms certify automated tools. Monitoring mostly counted how many audit teams used a tool, typically for licensing, and the regulator found "no formal monitoring performed by the firms to quantify the audit quality impact" of using them.Source 6 Generative AI was then limited to tools that assist audit teams or aid productivity, such as chat bots, and firms acknowledged that existing certification processes may not address AI's additional risks.Source 6
Fabricated sources are a recurring failure. In Canada, The Independent reported that a 526-page health workforce plan Deloitte wrote for Newfoundland and Labrador, at a cost it put at $1.6 million, cited research articles that do not exist. Deloitte said it used AI "to support a small number of research citations", that it stands behind the recommendations and that the corrections do not affect the findings.Source 3 In March 2026 the outlet reported that the provincial accounting body's complaints committee had decided to investigate; no finding on sanctions had been made.Source 3 In both the Australian and Canadian cases, the conclusions were said to stand. In our reading, a client cannot easily test that claim once the evidence beneath the conclusions has failed.
Regulators keep accountability with people. The FRC's March 2026 guidance on generative and agentic AI states that "the human auditor is always accountable", and that regulatory accountability for the quality of audit outputs is unchanged.Source 6
Where it falls short
The quality evidence is old or indirect. The strongest audit study ends in 2019, measures investment in machine learning staff and shows association, not cause.Source 4
Nobody is keeping score. The UK's largest audit firms did not formally quantify the quality impact of their automated tools.Source 6
Samples are small. ASIC's 10 audit files were chosen because of concerns, so they show where quality failed, not how often.Source 1
Assistance converges. AI-assisted stories were more alike than unassisted ones.Source 5 In our reading, advice drafted from the same tools risks the same drift.
Fluent errors get through. In both Deloitte cases, false citations reached a published report.Source 2Source 3
Where governance decides the outcome
Australian expectations are moving to the contract. Under the APS AI Plan released in November 2025, consultants and contractors will remain responsible for their work whether or not generative AI is used, and a new clause will cover transparency in providers' use of it, as The Canberra Times reported.Source 7 For audit, the FRC's position is that accountability stays with the human auditor.Source 6 In our reading, an advisory firm needs four things: a named person who checks every source and quotation against the original before release, disclosure to the client of where AI was used, a quality measure tracked before and after adoption, and senior time protected for the judgement calls the tools cannot make. Client data handling is covered in our confidentiality review.
How to think about adoption
Treat a better first draft as the new minimum, not the product. Decide in advance which parts of a deliverable a person must verify line by line: every citation, quotation and figure, at least. Measure quality, not only usage: track review findings, client corrections and restatements before and after the tools arrive. Expect clients and governments to ask how AI was used, and answer in writing. Then put the time the tools save into the judgement that clients cannot get from the same tools themselves: what the evidence means for this client, what is missing from it, and what to do next.
BFCS.ai does not fabricate figures. The ASIC and FRC findings are regulator observations from small, targeted reviews; the audit study is peer-reviewed but observational and predates generative AI; the writing experiment is peer-reviewed but concerns fiction, not advice; the Deloitte cases and the APS AI Plan details come from press reports. We found no independent study of how generative AI changes the quality of consulting, audit or accounting deliverables, and we grade confidence down for that reason.
Sources
- 1
Australian Securities and Investments Commission, Report 819, "ASIC's oversight of financial reporting and audit 2024-25," 31 October 2025. asic.gov.au
Back to textWhat we took from it
ASIC inspected 10 audit files of companies (eight ASX-listed, two large proprietary) across eight audit firms, selected where it had identified concerns from its financial report reviews and surveillances; one surveillance had no findings, one had one audit finding, six had both audit findings and other findings, and two had only other findings; an audit finding is where, in ASIC's view, auditors did not obtain reasonable assurance that the financial report as a whole was free of material misstatement, and does not necessarily mean the report was misstated; most audit findings related to revenue and receivables, impairment and asset values, and investments and financial instruments; ASIC will increase the audit files reviewed to 25 (compared with 15 in 2024-25, including five superannuation fund files reported separately).
- 2
Carly Page, "Deloitte refunds Aussie gov after AI fabrications slip into $440K welfare report," The Register, 6 October 2025 (press report): theregister.com
Deloitte, "Targeted Compliance Framework Assurance Review: Final Report," published by the Department of Employment and Workplace Relations (primary source, read in full for these passages; the department's page says it was updated on 3 February 2026, replacing the version published on 3 October 2025): dewr.gov.au
Chris Johnson, "Government urged to get Deloitte's total fee returned for delivering a botched AI report," PS News, 12 October 2025 (press report of the Senate estimates hearings of 8 and 9 October 2025; we could not open the Hansard): psnews.com.au
Back to textWhat we took from it
the Department of Employment and Workplace Relations report contained fake citations and a made-up quote from a Federal Court judgment, spotted by University of Sydney academic Dr Christopher Rudge after it was first published in July; the corrected report says Deloitte used "a generative AI large language model (Azure OpenAI GPT-4o) based tool chain"; Deloitte would repay the final instalment of its AU$440,000 agreement; the department said the substance of the review was retained and there were no changes to the recommendations.
the report's update note says it replaces the report dated 4 July 2025 and was updated to correct citations and reference list entries that contained errors and to amend the summary of the Amato proceeding, which contained errors, and that the updates "in no way impact or affect the substantive content, findings and recommendations"; the method describes "a generative AI large language model (Azure OpenAI GPT-4o) based tool chain licensed by DEWR". The report does not state the refund.
officials confirmed $97,587 of the $440,000 fee had been repaid; departmental Secretary Natalie James said "We should not be receiving work that has glaring errors in footnotes and sources."
- 3
Justin Brake, "Accounting watchdog to investigate Deloitte over fake citations in $1.6M healthcare report," The Independent (Newfoundland and Labrador), 7 March 2026 (press report): theindependent.ca
Back to textWhat we took from it
Chartered Professional Accountants Newfoundland and Labrador's Complaints Authorization Committee decided to investigate Deloitte's use of AI in the province's 526-page Health Human Resources Plan after a resident's complaint; the investigation had just begun and no finding had been made; if the committee finds reasonable grounds for sanction it can, among other options, refer the matter to a disciplinary panel; the report "came in at $1.6 million" (currency not stated) and was found by The Independent to cite research articles that do not exist; in an earlier emailed statement Deloitte said it used AI "to support a small number of research citations", that it was making a small number of citation corrections "which do not impact the report findings", and that it "firmly stands behind the recommendations".
- 4
Fedyk, A., Hodson, J., Khimich, N. and Fedyk, T., "Is artificial intelligence improving the audit process?", Review of Accounting Studies 27, pp. 938 to 985, published online 7 July 2022 (peer-reviewed, open access; abstract, introduction and fee results read on the publisher page). doi.org
Back to textWhat we took from it
The authors report associations ("is associated with"), not a trial, and the sample ends before generative AI tools were released. More than 310,000 employee resumes at the 36 largest US public accounting firms with at least 100 employees, 2010 to 2019; AI investment measured by employees with AI skills such as machine learning and neural networks; a one-standard-deviation change in recent AI investments is associated with a 5.0% reduction in the likelihood of an audit restatement and a 0.9% drop in audit fees; interviews with 17 audit partners from the eight largest firms found the primary goal of AI in audit is improved quality, through anomaly and fraud detection, risk assessment and refocusing people on higher-risk areas.
- 5
Doshi, A. R. and Hauser, O. P., "Generative AI enhances individual creativity but reduces the collective diversity of novel content," Science Advances 10(28), eadn5290, 12 July 2024 (peer-reviewed). pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
Online experiment with 293 writers recruited through Prolific, each writing an eight-sentence story with no AI, one GPT-4 idea or up to five ideas, rated by 600 separate evaluators; AI ideas made stories more creative, better written and more enjoyable, especially for less creative writers, whose stories with five ideas were rated 10.7% more novel and 11.5% more useful; AI-enabled stories were more similar to each other than stories by humans alone.
- 6
Financial Reporting Council (UK), "Thematic Review: Certification of Automated Tools and Techniques," June 2025, covering the six largest audit firms: media.frc.org.uk
FRC, "Innovative new guidance supports audit firm adoption of emerging AI technologies," 30 March 2026: frc.org.uk
Back to textWhat we took from it
monitoring was generally restricted to a sub-set of widely used tools and performed to understand how many audit teams used them, typically for licensing; "There was no formal monitoring performed by the firms to quantify the audit quality impact of using ATTs"; generative AI deployments were limited to applications to assist audit teams or aid productivity, such as chat bots; firms acknowledged that AI presented additional risks that existing certification processes may not address.
the generative and agentic AI guidance codifies good practice; "the human auditor is always accountable" and regulatory accountability for the quality of audit outputs remains unchanged.
- 7
Brittney Levinson, "What the public service's new AI plan means for external contractors," The Canberra Times, 13 November 2025 (press report of the APS AI Plan released 12 November 2025): canberratimes.com.au
Back to textWhat we took from it
consultants and contractors will remain responsible for their work regardless of whether generative AI is used in its development or delivery; a new clause will ensure transparency and accountability in external providers' use of generative AI in government work; suppliers on the management advisory and people panels will have to disclose planned AI use when responding to requests for quotes; no implementation dates given.