Issue 009 ended on the gap between how confident finance leaders were about their AI agents and how little accountability sat underneath. This fortnight, both sides of that gap moved — and they moved in the same direction.
On 2 August the European Commission's enforcement powers under the AI Act came into force. The AI Office can now demand information from a model provider, require access to a model for evaluation, order risk mitigations, and fine up to 3% of worldwide annual turnover or €15m, whichever is greater. The same date switched on the Article 50 transparency duties: systems that interact with people have to say they are AI, and generated or altered content has to carry a machine-readable mark. The high-risk obligations are still deferred to December 2027, so nothing in a finance function's core stack changed overnight — but the era in which AI rules existed only on paper ended on a Sunday morning in August.
The vendors moved the same way. Anthropic shipped three governance features in seven days: inference hooks that let a data-loss-prevention server allow or block content in real time, self-hosted environments so sessions run inside a firm's own network, and a Compliance API that now reaches Claude Code and Cowork as well as chat. Sage Intacct's Release 3 landed on 7 August with AI anomaly detection in accounts payable generally available — a control, not a productivity feature — while the headline Finance Intelligence Agent stayed in early adopter. Trustmi launched an AI investigation agent for payment fraud on 3 August, which is AI arriving on the defending side of the AP process rather than the attacking side.
Gartner supplied the shape of what all this leads to. Writing on 3 August, two of its analysts argued that the finance pyramid becomes a diamond: entry-level work shrinks, the middle thickens, and by 2030 around 70% of finance roles will be "shared jobs" in which a person and a machine do the work together.
The uncomfortable part is that almost every control that shipped this fortnight is enterprise-tier. A charity finance team of four is not buying inference hooks. What it can do is the thing that was true before any of this: decide who owns each process the AI touches, and write it down.
"By 2030, 70% of finance roles will be 'shared jobs', where humans and AI collaborate closely."
The Commission's AI Office, working with national authorities, now has active enforcement powers over providers of the most advanced general-purpose models: it can request information, request access to a model for evaluation, require risk mitigation, and ultimately require a provider to restrict, withdraw or recall a model. The penalty ceiling is the greater of €15m or 3% of worldwide annual turnover.
This is the end of the one-year grace period that ran from 2 August 2025, not a new obligation — but grace periods are where compliance projects go to be deferred, and this one has closed.
From 2 August, certain AI systems must tell people they are interacting with AI, and AI-generated or altered content must be marked in a machine-readable way, with deepfakes labelled. These are deployer-facing duties, not just provider-facing ones, and they carry the same penalty tier.
For a UK organisation the question is not whether it is "an AI company" but whether it puts an AI-generated image, an AI chatbot or AI-drafted content in front of people in the EU — which is a live question for any charity with EU beneficiaries, EU donors or an EU-facing campaign. Worth an hour with whoever runs your communications before it becomes a complaint.
Anthropic released inference hooks in beta for Claude Enterprise on 5 August, covering chat, Claude Code and Cowork. The mechanism matters more than the feature: the organisation's own DLP server makes the allow-or-block decision and Claude enforces it in real time, rather than the AI vendor deciding what your data policy is.
For a finance function, that is the first credible answer to "what stops someone pasting the payroll file into a prompt" that does not depend on training people not to.
On 6 August, Claude Code gained public-beta self-hosted environments, letting sessions run on an organisation's own infrastructure with internal network access and its own compliance controls, on Team and Enterprise plans. The same day brought security scanning of skills and plugins on Enterprise plans, to detect malicious content in third-party uploads.
The second one is the quieter risk and the more relevant one: as soon as a finance team starts sharing reusable AI "skills" between people, someone needs to be checking what is inside them, exactly as they would with a macro-enabled workbook from outside the organisation.
From 11 August, compliance and security teams can pull session content and metadata from Cowork and Claude Code through the same Compliance API interface as chat, across desktop, web, mobile and CLI. This is the unglamorous plumbing that turns AI use into something that can be sampled, retained and produced on request — the difference between telling an auditor what your team does with AI and showing them.
Generally available: AI-driven anomaly detection in AP automation, which analyses supplier billing patterns and submission details to flag unusual amounts and suspicious email addresses before payment; automated prepaid expense amortisation; inter-entity bill back; and a relocated journal entry approvals setup.
Still in early adopter: the Finance Intelligence Agent, which answers natural-language questions about financial data, and Smart Excel Reporting, which builds and refreshes Excel reports from live Intacct data under existing permissions.
One note for anyone planning around the pre-release material: the partner listings that circulated before launch also described an AI Gateway and intelligent three-way matching, and the post-release write-ups describe neither. Check anything you are relying on with your partner before it goes in a paper.
Trustmi launched an AI Investigation Agent on 3 August that runs the investigative workflow around a suspicious payment — correlating evidence across email, security tooling, vendor records and financial systems, connecting related events into a single incident with a timeline, answering an investigator's questions and producing a report.
Placed next to Sage's anomaly detection reaching general availability the same week, the direction is clear enough: the first genuinely mainstream AI use cases in AP are not paying invoices faster, they are catching the ones that should not be paid at all.
An automation that pays invoices carries the full weight of the control environment from day one. An automation that reads the same invoices and raises a hand carries almost none, because a person still makes every decision — and it earns the trust that the first kind needs.
If you want one AI project in the close this autumn that will survive contact with your auditor, make it one that produces an exception list rather than a posted transaction.
Writing in Corporate Compliance Insights on 3 August, Gartner's Markus Hofbauer and Alissa Lugo argue that AI does not shrink finance evenly — it hollows out the base. Entry-level roles contract as automation absorbs routine knowledge work, while mid-level roles expand because someone has to supervise both the people and the machines. By 2030 they expect around 70% of finance roles to be shared human-AI jobs.
Their two named risks are the ones a finance director should recognise: data flows become less visible, which makes privacy and security compliance harder to evidence, and AI-generated results may not be fully traceable, which challenges conventional audit and approval processes.
CFO Dive reported on 6 August that requests for interim C-suite roles have risen 151% since 2021, with demand for interim CFOs up 14% year on year, and that digital and AI initiatives now touch a quarter of all interim requests. Financial controls, accounting and audit remain the most in-demand competencies.
Read alongside the Gartner piece, it suggests the mid-market answer to the AI transition is not a permanent hire but borrowed capacity — which is a reasonable answer, provided somebody permanent owns the outcome after the interim leaves.
Inference hooks, self-hosted environments and compliance APIs are enterprise-plan features. A charity running AI on individual subscriptions gets none of them, and will not get them at this year's budget.
That is worth saying plainly rather than pretending the governance gap closes by buying something, because the practical consequence is the opposite of discouraging: if the tooling is out of reach, the controls have to be procedural, and procedural controls are free.
Which processes does AI currently touch, and who owns each one by name? What does the AI decide, and what does it only suggest? Who reviews the exceptions, and where is that written down? And if a trustee asked tomorrow what AI the charity uses, could you answer from a document rather than from memory?
The November 2025 refresh of the Charity Governance Code added a recommendation that charities have a policy for the use of technology and AI tools; those four answers are most of the policy, and the rest is formatting.
The Consultative Committee of Accountancy Bodies — ICAEW, ACCA, ICAS, CIPFA and Chartered Accountants Ireland, covering more than 285,000 accountants across the UK and Ireland — launched a dedicated AI Hub in late July, pulling together guidance, checklists and ethical leadership podcasts alongside the draft Statement to the Profession on the Ethical Use of Artificial Intelligence and its six case studies. Comments are invited through a survey on the Hub, and the draft and the responses will be discussed at an online event on 25 September.
This is the third issue running in which we have flagged it, for the same reason: the profession's AI guidance is being written now, and charity finance is materially under-represented in the responses.
| Finding | Source | Date |
|---|---|---|
| EU AI Act penalties for GPAI and transparency breaches reach the greater of €15m or 3% of worldwide annual turnover | European Commission | Aug 2026 |
| By 2030, around 70% of finance roles will be "shared jobs" in which humans and AI work together | Gartner, via Corporate Compliance Insights | Aug 2026 |
| Requests for interim C-suite roles are up 151% since 2021; interim CFO demand up 14% year on year | Business Talent Group, via CFO Dive | Aug 2026 |
| Digital and AI initiatives now feature in 25% of all interim leadership requests | Business Talent Group, via CFO Dive | Aug 2026 |
| CCAB's AI guidance covers more than 285,000 accountants across the UK and Ireland | CCAB | Jul–Aug 2026 |
| 79% of UK charities routinely use AI, while distrust of AI tools has more than doubled to 35% | Charity Digital Skills Report 2026 | Jul 2026 |
| 46% of charities take no action at all to manage AI risks | Charity Digital Skills Report 2026 | Jul 2026 |
The most useful line in a fortnight of governance news comes from the CFO of the company with the least reason to undersell AI: redesign workflows around the decision, not the task, and pair speed with accountability and controls. Her framing of what an AI-native finance function actually delivers is worth borrowing for a board paper — not fewer people, but a team that understands what is happening as it happens and gives leadership more time to act.
The clearest structural argument published this fortnight, and the one most likely to change how you think about next year's establishment. The diamond-shaped finance function is a more honest description of what AI does to a team than either "no job losses" or "half the headcount", and the compliance section is unusually candid about traceability being the unsolved problem.
Published in June rather than this fortnight, but it is the counterweight to everything above. The research finds proofs of concept that never leave the sandbox and dashboards that never shape a decision, and argues the cause is not data quality or vendor overpromising — because the same technology produces very different results in other functions. If you are being asked why the pilot did not scale, read this before you answer.