If Issue 006 argued that context, not the model, was the constraint, this fortnight the agents themselves arrived in force — and the whole industry pivoted to the question of how to govern them.
Live agents landed across the stack that charity finance teams actually use: Xero's JAX now pipes live ledger data into Microsoft 365 Copilot, Intuit switched on a "virtual team" of QuickBooks agents (including a VAT agent), SAP's Joule agents went generally available across procurement, Trintech shipped flux and variance agents for the close, and Copilot in Excel gained reusable, auditable finance "skills".
But the more telling story was the control layer being built around them: Anthropic handed administrators self-hosted gateways, per-user cost attribution and spend-threshold alerts; BlackLine launched a Finance Control Console to police "hundreds of thousands" of agents; and Sage Intacct exposed its ledger to AI assistants strictly read-only.
Regulators moved in the same direction — the EU gave final approval to its "Digital Omnibus", pushing high-risk deadlines back to late 2027 while extending its classification consultation to 23 July, and the Bank of England and FCA reopened their joint AI survey.
The fortnight's sharpest data point was the gap it all exposed: Cisco's CFO says AI now drafts 80–90% of the company's regulatory filings, while the early 2026 Charity Digital Skills Report finds 91% of charities use AI but only 30% have a policy. The capability is arriving faster than the controls — and closing that gap is now the job.
"AI now produces 80–90% of the first draft of our mandatory financial narratives."
Anthropic rolled out the governance layer finance teams have been asking for: a self-hosted "Apps Gateway" that puts corporate SSO, role-based access, per-user cost attribution and daily/weekly/monthly spend caps in front of Claude Code on AWS and Google Cloud, plus Claude Enterprise spend-threshold alerts at 75% and 90% of budget and an Analytics API that pipes AI spend straight into finance tooling like Datadog and CloudZero. For any FD worried about who can run agents and at what cost, this turns "AI spend" from an untracked line into something you can cap, attribute and forecast.
Shipped in Sage Intacct's 2026 R2 release, the new AI Gateway lets Claude and other MCP-compliant assistants query live GL, AP, AR, cash, purchasing and order-entry data in plain English over OAuth 2.0 — and deliberately read-only, so the assistant can answer questions but never post. For the many charities and NFPs on Intacct, it is a low-risk first step into connected AI: real answers from live data, with the guardrail that nothing can be changed.
BlackLine previewed a central command layer to deploy, monitor and enforce policy over what it frames as potentially hundreds of thousands of AI agents, keeping audit-ready records as agentic close and reconciliation work scales. The pitch marks the shift finance leaders should note: from "can AI do the task" to "can we prove it did it correctly" — the observability and controls that have to exist before agents are allowed near the ledger.
Copilot in Excel can now follow saved "skills" that codify repeatable finance workflows (closing the books, variance analysis, monthly model refreshes), pull from live data connectors (LSEG, Moody's, S&P Global, Morningstar) and attribute its edits separately from human ones in the Show Changes pane. For charity teams already on Microsoft 365, this turns Copilot from a chat helper into a governed, auditable process-runner inside the tool they build management accounts in — most of it live now, with custom skills arriving shortly.
The rebuilt Google Finance uses Gemini to track a consolidated portfolio, answer questions grounded in your actual holdings, auto-explain sudden price moves ("Key Moments") and deliver scheduled market briefings. It is consumer-focused, but the practical hook for finance teams is reserves and investment monitoring — a low-cost way to keep an eye on holdings and get plain-English explanations of what moved and why.
ChatGPT's personal finance experience — connecting accounts via Plaid to show a spending and portfolio dashboard and answer questions grounded in your data — dropped from the $100/month Pro tier to the $20/month Plus tier (US only for now). It is consumer-first, but it signals the direction: account-connected AI assistants that will increasingly reach staff, budget-holders and trustees.
Xero's JAX financial agent now feeds live accounting data into Copilot Chat, Excel, Word and PowerPoint, so staff can pull P&L breakdowns and overdue-invoice lists straight into a spreadsheet — no CSV exports — and fire payment reminders from chat. The practical win is faster board-pack drafting and reconciliation prep grounded in live ledger data; it is rolling out now, session-scoped, and not used to train Copilot's models.
Intuit began rolling out five QuickBooks agents (Accounting, Customer, Finance, Project Management and a beta VAT agent) plus an "Agent Studio" for firms to build their own; the Finance agent generates P&L, cash-flow and balance-sheet summaries with KPI analysis, while the VAT agent flags mismatches between P&L income and VAT-return net sales and proposes fixes. Intuit claims up to 12 hours a month saved on bookkeeping and close — relevant to smaller charities and the bookkeepers who support them (US rollout now; UK core features live since late 2025).
Three SAP Joule agents — Ariba Intake Management, Ariba Contracts and a Fieldglass services-procurement agent — reached general availability in June, automating intake routing, policy-compliance checks and automatic spend categorisation, with an earlier release claiming a 70% cut in manual statement-of-work creation. The month-end relevance is cleaner purchase-to-pay data and less manual coding of committed spend feeding the close; SAP's Joule agent runtime is free through 31 December 2026.
Trintech launched two AI agents that take on the investigative grunt-work of the close — the Flux Agent reviews account movements during close and flags significant balance changes and anomalies, while the Variance Analysis Agent runs post-close to compare budget against actual and surface the business drivers behind each variance. Finance teams can hand off the spreadsheet-chasing and first-draft explanations while keeping outputs traceable to source data and inside existing approval controls.
Basware topped the quadrant on both execution and vision, promoting an "AI-first" invoice platform it says hits 81% first-pass coding accuracy across 19m+ invoices and 80%+ touchless processing — with every AI decision logged and auditable ("governed autonomy"). For FDs weighing AP automation, it is a useful independent benchmark, and a reminder that the vendors winning now are the ones pairing autonomy with a full audit trail.
Cisco is rolling personalised AI agents out to all ~90,000 employees from end-July, and CFO Mark Patterson says AI now produces "80–90% of the first draft" of the company's mandatory financial narratives, alongside an investor-relations tool that pre-empts analyst questions and a "CFO cockpit" dashboard in development. It is a concrete named-organisation example of AI moving from pilots into core finance work — and, crucially, paired with company-wide upskilling rather than headcount cuts.
Kyriba's 2026 survey of 1,400 finance leaders found 91.9% already integrating AI into financial decision-making across some or all processes, and commentators increasingly frame 2026 as the year adoption gives way to accountability: boards and CFOs now want measurable return, not more pilots. For finance leaders, the practical implication is to attach every AI initiative to a specific, measurable outcome — hours saved on the close, days off the close calendar, errors caught — before it is scaled.
Early figures trailed ahead of the full 2026 Charity Digital Skills Report (due this month) put AI use across UK charities at 91%, up from 77% in 2024, yet only 30% have an AI policy in place — almost double last year's 16%, but still leaving most everyday AI use informal and ungoverned, with limited digital skills the top barrier (55%). It is the charity-sector version of the fortnight's whole theme: adoption has raced ahead of governance, and the policy is the catch-up job now sitting on the FD's desk.
Zoe Amar warns that charities embedding AI now face real financial and continuity exposure as vendors shift to metered, usage-based pricing and can restrict or withdraw access mid-project — meaning a charity could burn through its AI budget or lose a tool it depends on. She reframes "tokenomics" and sudden loss of access as a finance and continuity issue (not just an IT one), hitting smaller charities hardest, and urges FDs to budget for cost volatility and build fallback plans.
The sector's largest umbrella body appointed Jude Sheeran — EMEA managing director at AI firm SambaNova Systems and former COO of Shaw Trust — as chair-designate, a choice that signals AI and digital transformation moving to the top of the sector's governance agenda. FDs should expect that to show up as sharper board-level scrutiny of technology strategy and spend. (For the governance and cost-discipline angle, the DSC Charity AI Conference runs online on 9 July.)
A survey of 400+ ICAS members found 60% of chartered accountants want greater government intervention on AI, with 52% flagging client-data privacy and confidentiality as their main concern. It is a notable counterweight to the "adopt faster" drumbeat — a professional body's members actively asking for oversight, which charity FDs nervous about governance may find reassuring rather than restrictive.
ICAEW published guidance on GOV.UK Chat, the Government's generative-AI assistant now answering tax questions in the GOV.UK app, warning that it draws only on GOV.UK guidance — not HMRC's manuals or the legislation — so it is unsuitable for anything beyond basic queries, and the taxpayer remains liable for acting on wrong answers. It is a clean, real-world illustration of the governance rule finance teams keep meeting: the tool can help, but you still own the answer.
The upgraded 2026 CGMA Professional Qualification now embeds generative AI as core content across all pathways, examined from the May 2026 case study onwards, alongside business partnering and analytics. It is a signal of how the management-accounting body is building AI fluency into the next generation of qualified finance staff — and a prompt for FDs to think about how their existing team keeps pace.
The Council of the EU gave final approval to the AI Act simplification package on 29 June (the Parliament having endorsed it on 16 June), deferring high-risk application dates to 2 December 2027 for stand-alone systems and 2 August 2028 for AI embedded in products, and exempting small mid-cap firms from some obligations — while two new prohibitions bite from 2 December 2026. In parallel, the Commission extended its high-risk classification consultation from 23 June to 23 July, so the window to influence how borderline finance use-cases get classified is still open.
The regulators' fourth biennial joint survey of AI and machine-learning use in UK financial services is now live, expressly covering foundation models, generative AI and agentic AI, with AI named a 2026 PRA supervisory priority. Regulated finance-adjacent teams that receive it should complete it (it shapes future supervision), and every finance team can use the published findings as a free external benchmark for their own adoption and governance maturity.
The FCA's reopened AI Input Zone closed on 19 June, and the regulator confirmed it will not write AI-specific rules but is building an evidence base for a "good and poor practice" publication due later this year — which will effectively set supervisory expectations under the existing Consumer Duty and SM&CR regimes. The practical move now is to document how you govern AI, test and monitor model outputs and explain AI-driven decisions, because that forthcoming publication is what firms will be benchmarked against.
| Finding | Source | Date |
|---|---|---|
| Organisations deploying agentic AI in finance outperform peers by ~32 percentage points on average across performance metrics; 70% report better-quality decisions | KPMG (1,013 finance leaders, 20 countries) | Mar 2026 |
| 74% of AI's economic value is captured by just 20% of organisations; AI leaders generate 7.2× more value than competitors | PwC 2026 AI Performance Study (1,217 executives) | Apr 2026 |
| 84% of finance organisations have implemented or plan to implement AI — yet only 7% report high or very high impact from it | Gartner (Finance Symposium, London) | Jun 2026 |
| 44% of CFOs now use generative AI across five or more use cases (up from just 7% a year earlier); deep adopters spend 20–30% less time on transaction processing and reconciliation | McKinsey (102 CFOs) | 2026 |
| 91.9% of CFOs say they are already integrating AI into financial decision-making across some or virtually all processes | Kyriba 2026 CFO Survey (1,400 leaders) | 2026 |
| 54% of CFOs name integrating AI agents into finance a top transformation priority for 2026 | Deloitte CFO Signals / Finance Trends 2026 | 2026 |
| 91% of UK charities now use AI (up from 77% in 2024), but only 30% have an AI policy in place | Charity Digital Skills Report 2026 (early findings) | Jun 2026 |
Thomson Reuters' Chief Product Officer David Wong argues that AI errors in finance are uniquely dangerous because they arrive "polished, confident and sounding exactly like authoritative guidance" while being subtly wrong — and in accounting the damage hides in footnotes, scope exceptions and judgement calls. He warns general-purpose models were never built for tax and offers four vetting questions before delegating any of it.
BCG reframes the AI-first CFO as a CEO-level mandate, arguing the payoff is not cost-cutting but a shift from backward-looking reporting to real-time foresight — better forecasting, smarter capital allocation and continuous insight. Success, it stresses, depends on scaling proven use cases and rewiring workflows rather than simply funding more technology.
Drawing on a gathering of senior finance leaders, this write-up of the Q2 2026 CFO Forum argues bluntly that "the technology is no longer the constraint" — governance, people integration and disciplined prioritisation of a few high-impact use cases are now what separate leaders from laggards. It frames governance as an operating system and calls for a shift from AI-for-efficiency to AI-for-excellence.
Charlie Liu contends the real prize of agentic AI is not headcount savings but making capital, contracts, budgets, procurement and risk judgements move faster and more accurately across the organisation. Rather than replacing finance staff, agents absorb routine work like invoice validation, freeing teams for higher-value work while lifting overall productivity.