KPMG AI audit pilot leaves costs and client savings unresolved | Magica
KPMG’s AI audit pilot found deeper questions, but savings remain unproven
Editorial Team
••📖6 min read
KPMG says AI agents pursued some issues more deeply in a parallel UK audit, though humans still had to correct their work. Client fee savings remain unconfirmed, even as KPMG International has negotiated lower fees from its own auditor by citing AI.
KPMG says AI agents examined some financial issues more deeply than its conventional audit team.
A human had to intervene when an agent failed to connect regulations with company documents.
Client fee savings remain unconfirmed, although KPMG International has negotiated lower fees from its own auditor by citing AI.
KPMG, one of the Big Four accounting firms, says more than 100 AI agents sometimes asked better questions and investigated financial issues more deeply in an experimental audit of an unnamed FTSE 250 company this year. Conducted alongside a conventional audit, the experiment was not relied on when formally approving the company’s accounts, according to a report relaying KPMG’s account to the Financial Times.
The trial tests whether AI can work across an audit rather than support particular procedures. Its significance turns on both reliability and economics: deeper inquiry could improve the work, but the reported results do not establish lower total costs or lower client fees.
Company-reported scope of Project Exponential. The parallel experiment was not relied on for formal approval of the accounts; accuracy and cost savings remain unmeasured. Source: Business Standard.
An experiment across the audit process
Project Exponential tested agents across the audit process rather than assigning them only isolated tasks. They gathered data, analysed and documented it, and produced draft conclusions and observations. Agents working on different parts of the audit interacted autonomously and suggested subsequent steps.
Catherine (Cath) Burnet, KPMG's UK head of audit and previously its regional chair for Scotland and a lead partner on company audits, said that in several cases the AI investigated more deeply than the regular team and sometimes asked better questions. Her appointment in 2021 followed more than 25 years at the firm.
Her assessment also came with a concrete failure: an auditor had to step in after an agent failed to match a set of regulations to the company's documents. The present KPMG's assessment of its experiment. The account does not supply an independently verified performance comparison, a success rate or a count of additional material problems found.
Binance Wallet says subscriptions for its Anthropic IPO xStocks campaign reached the 500 million USDC limit. Final allocations, token delivery and applicable refunds remain pending; the cap does not establish proceeds received by the AI company.
The agents were not operating without human direction. Staff set their objectives, supplied information about the company and previous audits, checked their work and challenged their conclusions. An audit partner signed off the experimental exercise separately from the formal approval of the company's accounts.
Broader scope, uncertain savings
KPMG's prior work helps define what is new. In an April 23, 2025 announcement, its US firm described deploying agents in Clara, KPMG's global audit platform, to support expense checks and searches for unrecorded liabilities and accrued expenses. The announcement also described generative AI integrated in 2024 to help refine risk assessments, develop substantive testing procedures and improve audit documentation. The earlier tools therefore supported more than routine processing. Project Exponential extends the test across the workflow; it is not the firm's first use of agents in auditing.
Nor is KPMG alone in testing that broader approach. Rival Big Four firm PwC has also conducted parallel projects to assess AI across the audit process without relying on it, according to the report describing KPMG's trial.
The economic question is how much human work can be saved once building and supervision are counted. Audit firms have traditionally used large teams of junior accountants under senior supervision. KPMG acknowledged time-saving opportunities but did not say whether these would reduce clients' fees.
A team built the agents over four months. KPMG declined to disclose its exact size, but said two partners and one director worked with senior managers and junior staff to build and run them. Those details matter: a changed staffing mix is not yet a measured reduction in the total cost of an audit. The reports provide no like-for-like figures for hours, computing costs or fees.
There is a pointed contrast within KPMG’s global organization of legally separate firms. KPMG International has itself negotiated lower fees from its auditor by arguing that AI would make the work cheaper. That does not establish the savings from this UK pilot, but it makes the question of who receives any efficiencies concrete.
CPIA’s Censuswide survey found greater confidence in AI for reviewing large volumes of data or contracts. These are stakeholder attitudes, not measured performance; the two categories do not cover every response. Source: Centre for Public Interest Audit.
Confidence depends on the task
Evidence about audit users' attitudes points to a distinction between processing information and exercising judgment. In an online survey conducted by Censuswide from August 3 to 25, 2026, 67% of respondents had more confidence in work carried out primarily by AI for reviewing large volumes of financial data or contracts; 14% had more confidence in work carried out primarily by an auditor. Respondents favoured auditors on judgment tasks, including challenging management assumptions and forming the overall audit opinion.
The survey, published by the Centre for Public Interest Audit, a policy and research group for large audit firms, covered 257 senior finance professionals in the UK: 100 FTSE 350 chief financial officers or finance directors, 57 audit committee chairs and 100 professional UK-based equity investors.
Overall, 79% were happy with audit firms’ deployment of AI and technology. But 46% identified over-reliance on AI and weakened professional scepticism as a key risk. Among the 57 chairs, who appoint and oversee auditors, 23% had limited confidence in AI's use in their company's financial-statement audit. These are attitudes, not evidence that either humans or agents perform better.
Oversight involves more than technical accuracy. The UK's audit regulator, the Financial Reporting Council, has warned that sharing technology-generated insights with audit clients can threaten integrity, objectivity or independence, particularly when the work producing them falls outside the financial-statement audit, according to the report. Auditors face restrictions on advising audit clients. The regulator’s position is that auditors remain accountable: they must explain their use of technology, challenge its outputs and stand behind the evidence.
Burnet says KPMG’s system records agents' interactions with documents down to individual characters and timestamps. Such traceability can help reviewers examine the work, but does not establish that the agents’ conclusions are correct. KPMG gave the FRC a high-level overview of the experiment, according to the trial account; that is not evidence of regulatory approval.
The expansion must test repeatability and total cost
KPMG plans to expand Project Exponential to up to 20 larger audits and hopes eventually to offer AI-based audits to clients. Burnet said the technology was not yet available at the scale needed for thousands of audits, few companies currently had sufficiently clean data, and some clients might prefer not to use the service.
The expansion needs to establish whether deeper inquiry recurs across companies, how often people must correct the agents, and how much work remains after building, supervision and review are counted. Comparable evidence on errors, additional findings, staff time and total cost would allow clients and those overseeing audits to assess the claimed benefits. The pilot offers promising examples, with the commercial terms and evidence for wider reliance still unresolved.
Tencent says US restrictions on remote computing will likely slow its next AI model while leaving cloud revenue largely protected. Southeast Asian operators face licensing and financing uncertainty, though US cloud access and domestic demand offer partial alternatives.
Microsoft is rolling out Agent 365 Business Analytics through its Frontier early access program, pairing AI spending with estimates of assisted work. The dashboard expands its existing analytics as agent workloads add usage charges, but its dollar-value estimates do not establish financial returns.
Adobe’s Firefly update saves recurring visual subjects and creative preferences, extending an existing push into repeatable production. Memory has narrower reach than the assistant, while campaign demonstrations leave reliability and savings unmeasured.
Microsoft’s Ignite upgrades package enterprise knowledge, skills and tools into reusable business plugins. The expansion builds on IQ’s June release and competes with existing graph retrieval services; total savings and portability remain unproven.
The European Central Bank warns that debt and interlinked financing could carry an AI investment reversal into credit markets and the European economy. Resilient earnings, limited direct bank lending and stronger bank buffers temper its comparison with past bubbles.
Reported introductions helped Alice Huang secure Alibaba as a chip-rental customer and obtain financing. The relationship raises questions about Nvidia's overseas sales network, but does not itself establish illegal exports.
Payhawk announced a $280 million all-equity Series C led by Growth Equity at Goldman Sachs Alternatives, valuing it at $2 billion. The funding backs an expansion from managing business spending into broader AI finance operations, where competitors already combine automation, payments and controls.
A 16-country survey of 15- to 22-year-olds finds a median of 33% have used AI as a companion or friend. Users have higher adjusted odds of loneliness, but the research cannot establish whether AI helps, harms or replaces human connection.
South Korean President Lee Jae Myung has ordered a review of rewards of at least 100 million won for reporting technology-assisted exam cheating and legislation requiring technology companies to supply evidence. Ahead of the November 19 college entrance exam, the proposals would add informant incentives and post-exam investigations to classroom checks, but implementation rules remain unsettled.
Microsoft is pitching Copilot Workspaces as a shared environment for people and AI agents. Team collaboration is already in early access, but personal Autopilot integration is still coming soon, with final agent pricing unannounced.
Google is rolling out file references and sharing for Gemini skills as adult personal accounts lose the ability to use Gems. Automatic conversion preserves supported material, but file limits and missing tools can interrupt existing workflows.
Muddy Waters disclosed a short position in Nvidia while questioning a possible connection involving senior adviser Ajay K. Puri. Its own account leaves the key identity question unanswered, separating its new claims from an existing server-diversion prosecution.
Nvidia reported $115.8 billion in quarterly revenue in its first earnings release since co-founder Jensen Huang resigned. Strong demand gives interim CEO Colette Kress a financial cushion, while a campaign for Huang’s return collides with unresolved investigations and the board’s reported opposition.
OpenAI launched GPT-6.2 Astra and Sol for professional work, cutting Astra’s standard input and output token prices by 20%. The models offer cheaper paths through selected tasks, but research evaluations, reporting failures and unfinished independent safeguard testing complicate the case for broader autonomy.
Microsoft’s first live speech-to-speech model offers OpenAI API compatibility and planned enterprise integrations. Its published evaluations show a low-cost option, with several rivals completing more agentic tasks.
Canberra-based Trellis Data is offering its 220-billion-parameter Agentaus 1 model to business and government. Its pitch goes beyond keeping data in Australia, but performance, deployment costs and a planned capital raising remain key tests.
Perplexity is adding direct collaboration between frontier models to Computer, letting a lead agent coordinate up to 16 supporting agents. Company-reported benchmarks suggest mixed teams can outperform repeated copies of one model, but leave customer costs and production reliability unresolved.
Marsh reported 105 client cyber incidents in the last two weeks of October, against 48 in the final two weeks of May. The broker links the increase partly to AI, while warning that insurance panel access may not guarantee an available response team.