If Issue 004 was about the autonomous close arriving inside the big ERPs, this fortnight it became a product you can buy — and one that finally reaches charity scale. Ramp launched Stack, an "AI operating system" that runs month-end, codes transactions and reconciles cash (up to 50% faster on some books); OpenAI and PwC partnered to build agents across the finance rhythm; and Anthropic shipped Claude for Small Business with native QuickBooks and PayPal connectors for close, reconciliation and cash-flow — putting agent-driven month-end within reach of a two-person finance team. But the centre of gravity has shifted from "can AI do it?" to "can you trust it?". Sage used its conference to argue for "glass box, not black box" AI, citing research that 70%+ of finance leaders would reject an output they can't explain; the new GOV.UK tax chatbot launched at 90% accuracy — wrong one time in ten; and Gartner warned CFOs they are confusing deployment with value creation. Governance is hardening to match: Anthropic added a Compliance API, 28 security integrations and self-hosted sandboxes, and the ICO's statutory AI duty is now live with an agentic-AI code on the way. The one piece of breathing room: the EU AI Act's high-risk deadline has slipped from 2 August 2026 to December 2027. For charity finance teams, usage is now near-universal — but quality, not access, is the brake, and the person signing the accounts still owns the number.
"More than 70% of finance leaders say they would reject an AI output they cannot explain — and teams lose roughly a quarter of their AI time savings re-validating results."
On 3 June, Ramp unveiled Stack, an AI system that performs month-end close, cash reconciliation, transaction coding and journal posting — reporting up to a 50% cut in close time on some client books and outperforming general-purpose models on 200+ accounting tasks. The signal for finance teams: close-and-reconcile automation is now a packaged product with built-in auditability, not a DIY project.
OpenAI and PwC announced a partnership to develop AI agents across forecasting, reporting, procurement, payments, treasury, tax and month-end close, with OpenAI's own finance team acting as "customer zero" for governance testing. It shows the Big Four productising supervised agent workflows — raising the bar for what even small teams are expected to deploy.
LiveFlow research (5 May) found nearly 80% of corporate finance professionals blame month-end delays on waiting for data from other systems or departments — despite heavy AI investment. The practical lesson before buying anything: fix the fragmented data flows and chase the late inputs first, or you just accelerate the wrong workflow.
On 12 May, Xero's live Claude integration began letting users ask about cash position, overdue invoices and revenue or profit trends straight from their Xero ledger inside a Claude session, using the same engine behind its JAX superagent — with business data excluded from model training. For the many charities on Xero, month-end-style questions can now be answered conversationally without exporting data.
SAP confirmed (13 May) a U-turn extending selected AI capabilities to ECC and on-premise S/4HANA customers, not just RISE/cloud subscribers, after pressure from its large on-prem base. For finance teams on older or self-hosted SAP — common where cloud migration has stalled — it widens access to AI tooling previously gated behind a cloud move.
UK cloud-finance vendor iplicit released iplicit Insights on 12 May, letting users ask natural-language questions like "what are the key revenue trends?" and get instant answers without building a report. For lean charity finance teams, it lowers the barrier to self-serve management information without an analyst or manual exports.
On 26 May, Anthropic announced integrations with 28 security and compliance tools (CrowdStrike, Microsoft, Okta, Palo Alto, Varonis, IBM and others) plus a Claude Compliance API giving programmatic access to conversation content and activity logs. This lets finance and IT teams govern, monitor and audit staff use of Claude inside existing security tooling — a key prerequisite before any charity puts financial data near AI.
At Code w/ Claude London (26 May), Anthropic announced self-hosted sandboxes (public beta) that run agent tool execution on your own infrastructure, and MCP tunnels (research preview) that let agents reach private MCP servers without exposing them to the public internet. Finance agents can now reach internal ledger and CRM systems while keeping data, audit logging and network policy inside the organisation's perimeter.
Alongside the Opus 4.8 launch on 28 May, Anthropic added connector permissions to its Enterprise custom-roles framework — letting administrators control exactly which connectors and individual tools each role can use (for example, read-only ledger access for one team, write access for another). It is the kind of granular control finance leaders need as MCP connectors multiply.
A controllers' guide confirms Intuit's official QuickBooks Online MCP server (read chart of accounts, pull transactions, generate P&L/balance sheet, write transactions) remains an early preview running locally — so it works with Claude Desktop but not Claude.ai, and production teams route through layered services instead. QuickBooks-to-Claude is live, but DIY and desktop-only today, with turnkey web access still coming.
On 28 May, Workday and Google Cloud expanded their partnership so finance teams can query expense and travel policies, check corporate-card eligibility and raise requests conversationally via the Sana Self-Service Agent in Gemini Enterprise, with zero-copy data sharing between the two platforms. For teams already on Workday and Google, policy guidance and spend controls move into everyday chat rather than separate portals.
Fortune reported (29 May) that Microsoft is building a single Copilot interface combining GitHub Copilot, Copilot chat, Cowork and an internal "Autopilot" project, targeted for launch by end of summer. The striking context for buyers: fewer than 4.5% of Microsoft 365's 450 million users currently pay for Copilot — a sign that licence cost and unclear ROI are still holding finance teams back.
On 15 May, OpenAI launched personal-finance features in preview for ChatGPT Pro users in the US, using Plaid to connect 12,000+ institutions and surface portfolio performance, spending, subscriptions and upcoming payments. Consumer-focused for now — but it shows the direct-to-ledger account-connection model edging toward the cash-flow and spend analysis finance teams will eventually want.
Consero's May survey of 102 PE/VC-backed finance leaders found AI adoption jumped to 97% (from 76% a year earlier), 42% now have it broadly embedded, and most investments are returning within 12 months — yet 33% still name data quality as the top blocker, and 87% are still expanding headcount. The takeaway: treat AI as a capacity multiplier alongside hiring, not a substitute, and fix data foundations before scaling.
A May analysis of Netskope and IBM data found 40–65% of workers use unapproved AI tools (47% via personal accounts), while only 37% of organisations have any policy to detect or manage it — and shadow-AI breaches cost an average $670k more. For teams handling donor, payroll and supplier data, the practical move is a clear, role-specific AI use policy and an approved toolset now, because staff are already pasting sensitive data into ungoverned tools.
Microsoft's 2026 Work Trend Index (analysed for CFOs by Fortune) found organisational factors — culture, management support, talent practices — account for 67% of AI's impact versus just 32% from individual mindset, while only 26% of users report clear leadership alignment on strategy. The message for finance leaders: change management and workflow redesign, not licence counts, decide whether AI spend pays back.
Interim findings show 88% of charities now use AI day-to-day (up from 76% in 2025) — but limited skills hold back 55%, and accuracy/quality concerns now block 49%, up sharply from 33% a year earlier (full report due July 2026). For finance teams, this confirms that upskilling and output-verification controls, not access to tools, are the binding constraint on safe adoption.
On 13 May, Anthropic released Claude for Small Business: 15 agentic workflows and 15 task skills with native connectors to QuickBooks and PayPal, covering monthly close, book reconciliation, cash-flow forecasting, invoice chasing and plain-English P&L generation. For smaller charities, it puts agent-driven month-end within reach without bespoke integration work — provided data-handling and sign-off controls are set first.
A BambooHR survey of 1,248 US workers (24 Mar–9 Apr) found one in three new accounting and finance hires quit within their first year, with firms now hiring senior to entry-level staff at roughly 3:1 as AI absorbs the data-entry work that used to train juniors. Finance teams need to redesign trainee roles around exception-handling, reviewing AI output and analysis — or risk a broken talent pipeline.
The UK government's AI chatbot, drawing on 80,000+ pages of GOV.UK guidance, went live in the GOV.UK app on 14 May but was quickly shown to mislead on tax-free childcare cliff edges, landlord allowances and Making Tax Digital thresholds. Finance teams should treat it as a signposting aid only — the user remains liable for getting the tax right, so it is no substitute for professional review.
EU negotiators agreed (7 May) to defer high-risk AI obligations for standalone systems from 2 August 2026 to 2 December 2027 (embedded systems to 2028) in the "Digital Omnibus" package, pending formal Parliament and Council adoption. Finance teams using AI for credit scoring, insurance pricing or assessing individuals' financial standing get breathing room — but should keep building risk-management, data-governance and documentation evidence now rather than waiting.
In its 1 June statement, the ICO confirmed its 2026/27 plan includes a statutory AI code of practice (the duty took effect 12 May), dedicated guidance on agentic AI, and more consumer support. Finance teams deploying AI agents should note that controller responsibility and automated-decision-making obligations under UK GDPR remain firmly with the organisation — with clearer rules now coming.
| Finding | Source | Date |
|---|---|---|
| 97% of finance functions now use AI (up from 76% a year ago); only 3% are sceptical of the payoff; data quality still the top blocker | Consero (102 finance leaders) | May 2026 |
| 63% of finance teams found their AI rollout slower than expected — even though 66% reported efficiency gains | Gartner (204 finance leaders) | May 2026 |
| Finance AI adoption has plateaued at 59% (vs 58% the prior year), with data literacy and quality the biggest brakes | Gartner / CFO Dive | May 2026 |
| 87% of large-cap CFOs say AI is critical to finance in 2026; 54% name integrating AI agents a top priority | Deloitte CFO Signals (200 CFOs) | Jan 2026 |
| 70%+ of finance leaders would reject an AI output they can't explain; teams spend ~12.9 hrs/week re-validating AI work | Sage / IDC | May 2026 |
| 40–65% of workers use unapproved (shadow) AI; only 37% have a detection policy; shadow-AI breaches cost +$670k | Netskope / IBM (via MarkTechPost) | May 2026 |
| 1 in 3 new finance hires quit within a year as AI reshapes junior roles; senior-to-entry hiring runs 3:1 | BambooHR (1,248 workers) | May 2026 |
ACCA's largest-ever talent survey — 11,389 professionals across 160 countries — finds 52% now use AI regularly while 51% worry about its impact on jobs (up from 44%), and the charity/NFP sector is the most sceptical of any on AI in recruitment (63% not confident). The appetite to learn is there (82% are confident they can pick up AI skills); the open question is whether organisations make room for people to grow into it.
This Accountancy Age analysis argues the UK profession is splitting into PE-backed AI-native mega-firms and specialist boutiques, squeezing the mid-market — and notes that despite 91% of accountants adopting AI, 95% of AI pilots still fail. The takeaway: adoption intent is not implementation, and governance and change management are where most AI projects actually break.
CFO Dive reports Uber's engineering team burned through the company's full-year AI budget within four months using agentic coding tools, with some engineers spending $2,000/month, while finance struggles to prove ROI as token costs climb. The lesson for finance teams: agentic-AI cost control and ROI measurement are now a core finance discipline, not an IT footnote.
Aqilla user-conference research found three-quarters of finance leaders want to expand automation, 83% are optimistic about AI in finance, and only 8% now call themselves fearful — down from 20% a year ago. The signal for charity finance teams: the debate has moved from whether to how to deploy AI while preserving human accountability and visibility.
Sage launched its "Beyond the Black Box" initiative with PwC and a new AI Trust Label, after research found 70%+ of finance leaders would reject an AI output they can't explain and teams lose around a quarter of their AI time savings re-validating results. It reframes AI buying around auditability — the person signing the accounts cannot outsource accountability to a machine.