Issue 011 said the AI arrived switched on. Issue 012 said the vendors had built an approval gate in front of it. This fortnight the connectors stopped reading and started writing, which moves the question from the gate to the permissions behind it.
A free third-party connector now lets Claude and ChatGPT post journals into Xero, with no review step of its own. Microsoft's Dynamics 365 ERP MCP server is generally available and can create, update and delete records within the user's security role, and the older version retires on 1 October. Xero itself went live inside ChatGPT.
The cautionary tale arrived in the same week: a CFO roundtable heard how a Claude-to-NetSuite cash forecast kept running on wrong numbers after its connection dropped, and nobody noticed until someone went looking. Deloitte found that 31% of UK workers who use generative AI at work do so without their employer knowing.
So the useful control this fortnight isn't a policy. It's a list of what can write to your ledger, whose credentials it uses, and who would notice if it stopped working.
"31% of UK workers who use generative AI for work do so without their employer's knowledge."
Meridian, built by the US firm Pilot, launched a free hosted connector that lets Claude, ChatGPT and Codex "create and modify supported Xero transactions, including journal entries, invoices, bills, payments, purchases, and transfers", with each connected organisation set to read-only or read-write. The vendor is candid that it "does not provide shared access, workflow management, review, or an end-to-end close process".
Read that as a controls warning rather than a recommendation: anyone in your team with Xero access could connect it this afternoon, and nothing in it would stop an unapproved journal. Live. It is a third party, not Xero, and UK support is not stated.
Microsoft's Sameer Verma wrote that "the Dynamics 365 ERP plugin for Copilot Cowork and the Dynamics 365 ERP MCP server, both generally available, allow agents to reason over and work with ERP data, actions, and attachments", and that the Account Reconciliation Agent "is also evolving to resolve exceptions in bulk".
The dynamic server can change data as well as read it, bounded by the user's security role, which makes role design a control question rather than an IT one. The older static 13-tool server retires on 1 October. This is Finance & Operations, not Business Central.
Xero's plugin in ChatGPT "is now live", reading real-time profit and loss, balance sheet, invoices and bills, and its Claude integration now covers business pulse, overdue invoice chasing, cash flow snapshots, AP preparation, month-end preparation and recurring reports. Xero frames the output as "a first pass to review rather than a blank page", with "the review step and final judgement" staying with the accountant.
Connections to Excel, Word, PowerPoint and Copilot Cowork are coming soon. Separately, JAX's missing-receipt chasing, promised "from October", has not shipped.
At a CFO Alliance roundtable in Toronto, a century-old manufacturer described connecting Claude to Oracle NetSuite for KPIs and an AI cash forecast. The ERP connection dropped without anyone being told, and the tool rebuilt the forecast on incorrect data — "Nobody caught it until they went looking."
The company's rule now is "AI is for design, and systems are for production": finance uses AI to specify the report, and developers build it inside the system. The same discussion surfaced an engineer's AI spend reaching $5,000 in 30 days with no clear project attribution.
Of 25,000 UK workers, 63% knowingly use generative AI for work, and 31% of those do so without their employer knowing. Around half have had no formal training, and workers spend an estimated £958m a year of their own money on the tools.
"The leak isn't an enterprise tool. It's the personal account," as the Toronto roundtable put it. A licensed tool with a data agreement protects nothing if the work is going through someone's personal subscription instead.
Gartner found 55% of CFOs reported positive overall returns from 2025 AI initiatives, but 57% said returns were unclear when asked about individual use cases.
For data extraction, AP/AR automation and report creation, Marco Steecker's advice is blunt: "if you're not seeing value within the first nine months, 10 months of deployment, then you need to go and actually take a hard look at that investment." That is a review date you can put in the project plan on day one.
In a survey of 338 finance leaders across the US, UK, EU, Australia and Canada, 72% said their organisation spent more than planned on AI initiatives over the past year, and 66% know or suspect finance staff are using unapproved AI tools.
Data quality (48%) and system integration (42%) were the main limits on going further. Esker sells AP automation, so treat this as vendor research.
Anthropic announced that "Claude Cowork and chat are merging into one Claude", with Claude Docs and Claude Slides launched in beta and scheduled tasks available from any conversation. It is "rolling out on Pro and Max plans over the next few weeks"; Team and Free "will follow soon", and "Enterprise admins will hear from us at least 30 days before anything changes for their organizations".
Most charities sit on Team or Enterprise, so nothing changes today — but any staff guide or training material that describes Cowork as a separate app will need rewording.
Anthropic says Opus 5.5 costs "40% less to run than Opus 5" and performs at the level of Fable 5.1 on most work. Its own finance example — a merger model and executive presentation — took 63 minutes rather than 93, at half the cost.
The benchmarks are Anthropic's, but for anyone paying for heavy spreadsheet and analysis work by usage, the price change is real.
Microsoft's Show Changes pane in Excel "now includes on-card attribution showing when Copilot made edits or suggestions", and Copilot's replies link straight to the cells, tables and charts it changed. Generally available on Windows, Mac and web.
For a management-accounts workbook that several people touch, this is the first built-in way to evidence which numbers a person entered and which the AI did.
Sixthfin's survey of 303 UK CFOs at private companies with 250+ staff found 67% name reliability of the accounts as their top close priority, ahead of speed (57%), and only 42% describe their close management as "very satisfactory". 67% use Excel for account analysis and justification, and 3% use dedicated reconciliation software.
It is vendor-sponsored and the fieldwork was in March, but it supports a point worth making to any board: the bottleneck is the review, not the posting.
FloQast announced a no-code builder that turns existing close files into agents, and a module that applies COSO's framework — risk assessment, control activities, monitoring — to those agents. It is a vendor announcement with large US customers and no UK pricing.
The takeaway travels anyway: a well-documented close checklist is the raw material for automation, and governing an agent against a recognised controls framework is becoming a normal expectation.
PwC set out how its auditors use AI — human review of every output before use, continuous accuracy measurement, and different vendors' models for different tasks.
If your auditor works that way, expect them to ask you for the same evidence of human review on AI-prepared reconciliations, accruals and journals.
HMRC's algorithmic transparency record describes a "propensity to pay" model that uses machine learning on past debt-repayment behaviour to segment all four million UK-registered businesses, re-scored monthly, and shapes "the frequency and content of letters sent by HMRCs Debt Management function".
HMRC chose decision trees "over neural networks due to being explainable" — a useful precedent to quote when your own board asks why an AI tool should be explainable.
A separate record describes HMRC Assist, a rule-based model — not AI — that compares a draft VAT return passing through your software's API against data HMRC already holds, and sends back a plain-English nudge where it spots a likely error. Acting on the nudge is voluntary.
Gaps between your ledger and HMRC's third-party data will surface at the point of filing, so ask your VAT software supplier whether they will support it.
From 30 September a board-led Information Commission takes over under the Data (Use and Access) Act 2025; it will "continue to be known as the ICO", with its powers unchanged.
The ICO's agentic AI guidance is still at drafting stage, with its public consultation listed as "due to launch September 2026" on the ICO's own guidance plan as of 24 September — watch for it in the coming weeks.
The accountancy bodies' draft Statement to the Profession on the Ethical Use of AI is discussed at a free CCAB online event today. The draft says accountants "cannot abdicate responsibility for outputs produced by their use of AI tools and systems", and comments are invited through the CCAB AI Hub.
The government's workplace monitoring technologies consultation closes at 11:59pm on 30 September. So does the Pro-Worker AI Adoption Prize, below.
An AI assistant explained Trussell's work, took the donor's bank details and set up a Direct Debit mandate inside the conversation, with no separate form. GoCardless built it within the FCA's AI Live Testing programme, and the Chartered Institute of Fundraising welcomed it while stressing transparency, informed consent and donor control.
Before your fundraising team signs up to anything similar, ask how consent, the donor record and the Gift Aid declaration are captured when an agent, not a form, takes the mandate.
The Civil Society Group and Charity Finance Group's Autumn Budget submission asks government to "co-design a Civil Society AI Adoption and Resilience Programme and ensure charities have equal access to public digital adoption schemes that are often only available to businesses".
It is an ask, not a commitment, and AI sits alongside full cost recovery and Gift Aid. Watch the Budget for any business digital-adoption scheme that opens to charities.
The government prize is open to "UK businesses, charities and public sector organisations of all sizes" employing at least five people in the UK, judged by a panel chaired by Nobel economist Simon Johnson, with case studies written on the three winners.
If you have a well-governed AI rollout that made jobs better rather than fewer, it is a low-cost way to have that recognised.
| Finding | Source | Date |
|---|---|---|
| 63% of UK workers knowingly use generative AI for work; 31% of those users do so without their employer knowing | Deloitte UK, 25,000 UK workers (Ipsos UK) | 16 Sep 2026 |
| 95% of finance leaders are comfortable with agentic workflows; only 14% support full autonomy for critical decisions | Deloitte Finance Trends 2027, 1,434 finance leaders at $1bn+ companies | 9 Sep 2026 |
| Only 25% of organisations using internal approval require quantified ROI and a formal business case for AI spend | Deloitte Finance Trends 2027, as above | 9 Sep 2026 |
| 55% of CFOs report positive overall returns from 2025 AI initiatives; 57% say returns on individual use cases are unclear | Gartner, via CFO Dive | 24 Sep 2026 |
| 72% of finance leaders say their organisation spent more than planned on AI in the past year | Esker, 338 finance leaders, US/UK/EU/Aus/Can (vendor-sponsored) | 24 Sep 2026 |
| 67% of UK CFOs use Excel for account analysis and justification; 3% use dedicated reconciliation software | Sixthfin, 303 UK CFOs, fieldwork March 2026 (vendor-sponsored) | 14 Sep 2026 |
| 8% of UK charities use AI for financial management, up from 4% in 2025, against 79% using AI in some way | Charity Digital Skills Report 2026, 807 responses | Jul 2026 |
Boucher's five steps — a checklist with a log, a task board, tools for the agent, a test in chat, then a scheduled daily run — turn an agent into a close coordinator that assigns and chases. The line worth keeping is the one about the checklist itself: "If a line has no name, you have just found a gap to be addressed."
Across 51 AI deployments his team studied, "77% of the hardest challenges weren't technical but rather were change management, data quality, and process redesign", and headcount reduction featured in only 45% of them. A useful corrective for any business case that treats the model as the hard part.
Staff are increasingly using AI to raise concerns they don't feel safe raising directly — asking a tool to make the case against a proposal, then circulating it. "AI's perceived neutrality plays a role in reshaping how psychological safety is experienced in organizations." Worth reading if an anonymous AI-generated critique has ever landed in your inbox before a budget meeting.
The argument is that close and reconciliation AI only works on clean data and embedded controls, and that without reliable supplier master data and consistent approval thresholds, agents scale up the mess. The author is Trintech's Chief Partner Officer, so it is a vendor's view, but the prerequisite list is the right one: supplier master file, delegation thresholds, reconciliation ownership.