Issue 007 argued that the agents had arrived and governing them was the job. This fortnight the bill for that governance arrived — and it has a name.
Sage and IDC surveyed 2,275 senior finance decision-makers and found that finance teams now spend close to 13 hours a week reconstructing, validating and defending what their AI produced, losing roughly 26% of the expected productivity gain in the process. They called it the verification tax, and it reframes every AI business case: the gains are net, not gross. The same study found 71% would reject a tool that was 99% accurate if it could not explain how it got there.
The trust data lined up behind it. ACCA and CA ANZ found 93% of finance professionals worried about the integrity of AI-generated insights, while 72% rate their own generative AI skills as basic or non-existent. The FRC reported that corporate reporting "remains human-led", with trust and data quality the stated brakes. And the full Charity Digital Skills Report 2026 landed with a finding that deserves attention: charity distrust of AI has more than doubled in a year, from 15% to 35%.
Meanwhile the plumbing kept moving. Xerocon London brought a natural-language agent builder, the MCP specification ships its largest revision yet on 28 July, Microsoft set an October deadline to migrate off its old Dynamics 365 MCP server, and the FCA published the Mills Review — its blueprint for a world where AI initiates and executes transactions.
The pattern is not that AI stopped working. It is that the burden of proof has landed on finance, and nobody has budgeted for it.
"Finance professionals spend nearly 13 hours every week reconstructing, validating and defending AI outputs — around 26% of the expected productivity gain, lost to proving the work."
The complete report published this month on a record 807 responses (71% from organisations under £1m income), reporting that 79% of voluntary organisations now routinely use AI and 60% have an AI policy or are developing one. Those figures differ from the early findings we carried in Issue 007 (91% using, 30% with a policy) because the full report measures routine use and counts policies in development — a reminder to wait for the full dataset before quoting a headline to your board.
Distrust of AI tools rose from 15% to 35% in a year, running highest among large charities (44%) and those working with marginalised communities, while 46% of charities take no action at all to manage AI risk. The governance numbers are the ones for the audit committee: a third of boards and a quarter of CEOs are rated poor on AI skills, and 46% have no trustee with digital competence.
The report found 45% of charities using AI for grant fundraising, while 58% of funders say AI is already changing the applications they receive and 55% have no visibility over whether an applicant used it. Meanwhile digital funding from trusts and foundations has fallen to 17%, and 63% of charities name squeezed finances as their single biggest barrier to digital progress.
July's Governance & Leadership makes AI its cover theme, with case studies from Myton Hospices and Citizens Advice, and one warning worth carrying into your next trustee discussion: AI tools change substantially after boards have approved them, opening a gap between what trustees signed off and what is actually running. The practical answer is a version-aware AI policy with periodic re-approval, rather than one-off sign-off.
The 2026-07-28 specification makes MCP stateless at the protocol layer and adds two things finance teams should care about: a Tasks extension for long-running work that can be started, polled and cancelled — the pattern month-end jobs actually need — and authorisation hardening that aligns MCP with OAuth 2.0 and OpenID Connect, including issuer validation. If you are running MCP connectors into finance systems, ask your provider when they will be on the new specification.
Message centre notice MC1420830 confirms the preview static MCP server for Dynamics 365 finance and operations is retired on 1 October 2026, replaced by a dynamic server already generally available in versions 10.0.47 and 10.0.46 PQU-2. The new server can create, read, update and delete against data entities directly, so Microsoft explicitly warns admins to re-verify user permissions before switching — access is bounded only by what the calling user is granted.
Oracle added a pro-code builder to AI Agent Studio for Fusion Applications, letting teams create and run agentic applications from VS Code, Git workflows and AI coding assistants including Claude Code. The agents target named finance outcomes — Ledger, Payables, Expenses and Payment agents aimed at close acceleration and collections — and run inside Fusion, inheriting its security, approvals and audit trail; it is live now at no additional cost.
Entrust launched a co-development programme building an identity and authorisation layer for production AI agents, covering verifiable agent identity traceable to a confirmed human principal, real-time authorisation within delegated policy, cryptographic assurance of agent credentials, and accountability records for regulators. Financial institutions are named as target participants — this is the emerging answer to the audit question of who let an agent near the ledger.
At Xerocon London on 8–9 July, Xero announced XeroForce, which lets accountants create custom AI agents through prompts rather than code, connecting Xero to third-party tools already in the practice (early access now, general availability later this year). Alongside it, Bill Protection inspects bills before payment and flags unusual amounts, altered bank details and unfamiliar suppliers — putting a mandate-fraud control at the point of payment rather than in a periodic supplier-master review.
The 2608 release introduces an assistant that lets the close accountant orchestrate a team of agents handling receivables clearing, intercompany reconciliation and KPI analysis within set boundaries, plus a Receivables and Payables Clearing Agent that generates balanced clearing proposals with written explanations. The month-end effect is that AR/AP clearing and intercompany reconciliation move from prepare-from-scratch to review-and-approve, with an audit trail attached to each recommendation.
Oracle started the 2026.2 rollout on 15 July, adding Intelligent Flux Analysis that drafts explanations of why account balances moved between periods, an AI assistant that recommends a single likely GL match where reconciliation suggestions are ambiguous, and narrative summaries of unmatched items and ageing exceptions. It targets the two slowest jobs in a small finance team's close: writing variance commentary and clearing the bank rec exception queue.
Unit4, which has one of the larger UK charity and not-for-profit footprints, is offering commitment-free access to its AI capabilities until 31 August 2027 for customers registering before 31 December 2026, including Ava, which orchestrates agents across finance, projects and people through natural language in Microsoft Teams. Pricing afterwards is a 20% premium for ERPx and 5% for Azure-hosted ERP CR, which removes the usual business-case barrier to a controlled pilot.
Puzzle's AI Suite went live on 10 July with a Close module where accountants describe their month-end process in natural language and agents draft bank reconciliations, revenue recognition, payroll, depreciation and amortisation schedules, then flag journal anomalies with plain-language variance explanations. The shift worth noting is that the process description itself becomes the automation — no rules-engine configuration.
BlackLine published a seven-metric benchmark guide arguing teams should measure record-to-report health — journal automation rate, hours spent cleaning bank data, reconciliation rates, audit-file completeness — before deploying agents. Its core argument is the right one: AI may accelerate the close, but it will not fix a process still built on manual journals, spreadsheet reconciliations and audit documentation assembled after the fact.
Payouts.com launched role-based agents covering accounts payable, collections, treasury and month-end close, handling invoice OCR, PO matching, vendor onboarding, approvals and reconciliation. The governance design is the transferable part regardless of vendor: every new agent runs in shadow mode for two weeks before earning autonomy, with hard spend limits, plain-English policy rules and weekly reports to the finance lead.
ChatGPT Work takes a goal and produces spreadsheets, slide decks, documents and interactive dashboards, working autonomously across connected apps for extended periods rather than answering in a chat window. For finance teams the practical hook is scheduled recurring tasks — a monthly variance pack or board dashboard set to assemble itself on a calendar.
Anthropic moved Cowork beyond the desktop to iPhone, iPad, Android and claude.ai, running sessions remotely so work continues with the laptop shut and scheduled jobs run with no device online. The finance-relevant mechanic is the approval checkpoint — Claude pauses and notifies you before consequential steps, which makes it usable where human sign-off is a control requirement rather than a courtesy.
Claude models are now generally available in Microsoft Foundry on Azure, running under existing Azure authentication, billing and governance. The detail that matters for UK and EU finance teams: there is no European data zone for Claude on Foundry, inference under "Global Standard" can route anywhere including US infrastructure, and as a third-party marketplace offering it falls under the US CLOUD Act — so the Azure wrapper does not deliver the data residency that Bedrock or Vertex AI do.
Three major UK banks signed memorandums of understanding to co-design "Lumen Sovereign", a large language model led by British startup Cosine, intended to run without reliance on US cloud infrastructure or US AI providers. It is a data-sovereignty play for regulated financial services, and a signal worth watching for any finance function that has been asked by trustees or auditors exactly where its AI processing physically happens.
A CFO Dive synthesis of PwC, EY, McKinsey, Protiviti and Cloud Security Alliance research set out seven risks CFOs are underweighting, led by ungoverned "shadow agent" deployment and absent ROI: 82% discovered internal AI agents they were unaware of, 56% of CEOs reported no revenue gain or cost reduction from AI in twelve months, and 71% of CFOs say their current ROI metrics are unfit for measuring AI gains. It is the counterweight document for anyone about to sign an agentic finance business case.
Sarah Friar has pitched a four-part scorecard — whether the technology does meaningful work, cost per successful task, output reliability, and whether each dollar produces more value as usage scales — arguing that "AI is different: it needs to be measured by work accomplished." Set against a January PwC study in which only 12% of CEOs reported AI delivering both cost and revenue benefits, it is a more honest frame than headline time-saved claims.
Federal Reserve Bank of Chicago research found employment in the occupations most exposed to AI rose 4.05% between 2019 and 2024, while less vulnerable occupations saw payroll reductions, with wages rising across all exposure levels. The researchers concluded that exposure "does not map mechanically onto job loss, at least in the short run" — a restructuring-of-tasks story rather than an elimination one, and useful evidence when reassuring a finance team.
A global survey of 1,600 finance professionals found 93% concerned about the integrity and verifiability of AI outputs, citing hallucinations, incomplete datasets, opacity and bias. The skills picture underneath is the actionable part: 72% rate their generative AI skills as basic or non-existent, 41% are funding their own training, and data quality and skills shortages are each named as a barrier by 42%.
The Consultative Committee of Accountancy Bodies launched an AI Hub with a draft Statement to the Profession on the Ethical Use of Artificial Intelligence and six draft case studies spanning firms, industry, public sector and non-executive roles. Comments are open ahead of an online event on 25 September, and this is a genuine opportunity to shape guidance rather than receive it.
ICAEW published five practical ethical scenarios to test professional judgement, alongside a governance checklist of ten essential questions for responsible adoption. One scenario is worth putting in front of your audit or finance committee: a non-executive director at a company that has automated revenue recognition and asset valuations, where neither the CFO nor the finance team can explain how the AI arrived at key figures.
The FCA published the final Mills Review on 6 July, making seven priority recommendations to its Board on adapting the regulatory perimeter for autonomous "agentic finance", where AI systems initiate and execute transactions within pre-set parameters. The FCA confirmed its "good and poor practice" AI publication is still coming later in 2026 — so the practical move now is documenting where AI already touches your decision-making, because evidence of governance in practice is what you will be asked for.
The FPC's July Financial Stability Report flagged that frontier AI advances have increased cyber and operational resilience risks, alongside concerns about leveraged AI-related lending and stretched valuations. With AI in the PRA's 2026 supervisory priorities and a Bank/PRA consultation on cyber and ICT risk coming, the sensible response is to revisit operational resilience and map your third-party AI dependencies now.
The European Commission's targeted consultation on the draft Article 6 guidelines closes on 23 July, with final guidelines due for adoption by end-2026. Separately, the Digital Omnibus was signed into law on 8 July, confirming that high-risk obligations defer to 2 December 2027 and embedded systems to 2 August 2028 — real breathing room, but the AI Office gains significantly expanded supervisory powers and new prohibitions still bite from December 2026.
Research commissioned by the Financial Reporting Council found companies deploying generative AI mainly for narrative reporting and lower-risk, task-specific work, with very limited use in the financial statements themselves, citing trust, data quality, governance gaps and legal risk as the brakes. The FRC's Audit Tech and AI Sandbox closed to applications on 17 July, with findings from both feeding its future guidance.
| Finding | Source | Date |
|---|---|---|
| Finance professionals spend nearly 13 hours a week validating and defending AI outputs — 26% of expected productivity gains lost to verification | IDC for Sage (2,275 finance decision-makers) | Jul 2026 |
| 71% would reject a 99%-accurate AI tool that cannot explain how it reached its answer; 54% would pay a premium for greater visibility | IDC for Sage (2,275 finance decision-makers) | Jul 2026 |
| 93% of finance professionals are concerned about the integrity and verifiability of AI-generated insights; 72% rate their generative AI skills basic or non-existent | ACCA / CA ANZ (1,600 finance professionals) | Jul 2026 |
| Organisations with AI embedded in everyday work rose from 13% to 22% in a single quarter — but only 7% have established ROI | KPMG Global AI Pulse Q2 (2,145 leaders, 20 countries) | Jun 2026 |
| Where the CEO is accountable for AI outcomes, 57% realise meaningful value — against 21% where they are not | KPMG Global AI Pulse Q2 (2,145 leaders) | Jun 2026 |
| 82% of organisations discovered internal AI agents they did not know existed; 71% of CFOs say their ROI metrics are unfit for measuring AI | CFO Dive synthesis (PwC, EY, McKinsey, Protiviti, CSA) | Jul 2026 |
| 79% of UK charities routinely use AI, but distrust of AI tools has more than doubled to 35%, and 46% take no action to manage AI risk | Charity Digital Skills Report 2026 (807 responses) | Jul 2026 |
| Gartner predicts 20% of finance organisations will stop hiring talent without digital literacy by 2028; only 15% of a typical finance team are digital creators today | Gartner, Predicts 2026: Toward an AI-First Finance Function | Jul 2026 |
Jim Tyson offers a genuine analysis of the gap between record commitment — $2.6tn globally in 2026 — and how thin the evidence of mature implementation still is, with leaders reporting a 20% EBITDA lift while most organisations sit in unproven pilots. Nintex CFO Burt Chao supplies the line that carries it: "If you don't do the crawl, walk, run but just go sprinting off the couch — you're going to fail."
Anders Liu-Lindberg identifies a confidence gap that should give any finance leader pause: nearly 60% of CFOs plan to raise AI spending by 10% or more annually, yet only 36% feel confident of driving meaningful impact from it. His diagnosis is that adoption has outrun the operating models, data foundations and skills needed to absorb it — so bolting AI onto unchanged processes is the failure mode.
Mike Nader argues finance's mandate is shifting from reporting what happened to explaining why and what to do next, and that AI shortens the path to an answer without manufacturing the insight itself. His sharpest point: the most valuable forecaster is not the Excel expert but the person who reads operational signals before they surface in the ledger.
Vivian S. Lee, Linda Mantia and Jon McNeill describe "a second great compression of entrepreneurship", as mature models, multi-agent coordination, cheap APIs and falling cloud costs collapse the time, capital and headcount needed to start a business. It is not a finance-function piece, but the implications for capital allocation and competitive positioning are direct, and it is the strongest strategic essay published in the fortnight.