Three Quarters of Finance Teams Use AI. Fewer Than Half Can Prove It Works

KPMG's Global AI in Finance survey, conducted in March 2026 among 1,013 senior finance leaders, found active use of AI in finance at 75 percent, up from 30 percent in 2024 ^[1][2]. In the same survey, only 42 percent said their organizations were fully assurance ready ^[3]. The 33-point spread between those two figures is not a technology gap, it is a paperwork gap ^[4], and paperwork is what auditors test. The tools are past the demo stage. BlackLine made its Verity Prepare agents generally available in July 2026 ^[5], FloQast launched AI agents for close workflows in March 2025 ^[6], Microsoft has an account reconciliation agent in preview for Dynamics 365 ^[7], and Workday has announced a financial close agent slated for 2026 ^[8]. Adoption has grown two and a half times in two years ^[9]. Assurance readiness, on the same survey's evidence, has not followed ^[3]. What the undocumented version looks like in production is visible in Workiva's 2026 Midyear Executive Benchmark Survey, which polled 2,272 finance, risk and sustainability professionals, including 847 C-level executives, plus 367 institutional investors ^[10]. Eighty-four percent of executives said they were at least somewhat confident in the accuracy of AI output without human review ^[11]. Twenty-six percent said internal audits had detected AI errors that reached external audiences or board members ^[12]. Only 11 percent believed their data quality was sufficient for AI use ^[13], while 71 percent said poor data quality had at least moderately affected AI use in financial and sustainability reporting ^[14]. Among the investors polled, 89 percent said they were concerned about AI accuracy in corporate disclosures ^[15]. The remedy, per CPA Practice Advisor, is five controls written down before go-live rather than a pause ^[16]. First, an agent inventory: task, accounts and assertions touched, read access, write access, and one named human who owns the output ^[17]. COSO published guidance in February 2026 on internal control over generative AI ^[18], and its executive director Lucia Wind said such tools "can be confidently wrong, easily manipulated, or deployed outside formal oversight channels" ^[19]. Second, evidence a reviewer can replay after the fact: inputs used, steps taken, exceptions raised, and the basis for each match ^[20]. Vendor claims of 90 percent match rates and 90 percent preparation time savings are marketing figures until a client's own data reproduces them, which argues for a validation period with retained results ^[21]. Third, human sign-off gates scaled to materiality and judgment, written into the close checklist with names ^[22]. Deloitte Australia agreed in October 2025 to refund part of its fee to an Australian government department after a delivered report was found to contain fabricated citations ^[23]; the reading offered is that unreviewed output is a professional liability in any workflow ^[24]. Fourth, third-party assurance read closely: most agentic close tools are subservice organizations in SOC terms, and the AICPA has not yet published guidance specific to AI use in SOC reporting, so existing SOC 1 and SOC 2 reports may say little about model behavior, retraining or prompt controls ^[25]. Where the report is silent, the gap gets covered by contract terms, complementary user entity controls and the user's own testing ^[26]. Fifth, baseline metrics set before go-live, tracked every period, with a defined threshold at which the agent is pulled and a named person authorized to pull it ^[27]. The deliverable is a one-page memo covering inventory and owners, evidence standards, sign-off gates, assurance mapping and monitoring with an exit threshold, dated, signed by the controller, and copied to the auditors ^[28]. Worth watching: whether Workday's announced close agent ships in 2026 ^[8] and whether Microsoft's reconciliation agent leaves preview ^[7]; whether the AICPA fills the SOC reporting gap ^[25]; and whether validation periods reproduce the 90 percent numbers on real client ledgers ^[21].