Issue 011 was about AI features arriving switched on by default. This fortnight the products grew a gate — and the interesting part is where the vendors chose to put it.
Caseware's new agent sits inside a live Excel workbook and will not change a cell without explicit approval. Paystand's agents went generally available writing straight into NetSuite, Sage Intacct and Dynamics, with finance setting each agent's access level. Docusign is opening its contract data to any agent from 30 September.
Meanwhile the defaults kept moving on their own. Claude Cowork's built-in browser turns on by default for Enterprise organisations from 10 September unless somebody turned it off, Excel's =COPILOT() worksheet function is withdrawn on 14 September, and Microsoft has stopped publishing Dynamics 365 release plans altogether.
Against that, PEX found 66% of finance leaders interested in AI and 31% using it, with only 28% comfortable letting it decide anything routine. The capability question is settled. The unresolved one is who signs the work off — and this fortnight the lawyers answered it in a way that should worry anyone whose AI policy is a prohibition.
"61% of finance and operations leaders are interested in AI-generated financial reports. 14% actually use them."
Verity for Excel is a side-panel add-in that takes plain-language instructions inside a live workbook — pull these balances into a summary table, chart the year-on-year movement — and works from actual engagement data rather than generating assumptions.
It requires explicit approval before it changes anything in the sheet. That is the design pattern worth copying for any team letting AI near a working paper, and it is a noticeably different instinct from the one that shipped a year ago.
A Reporting Agent that prioritises aged receivables and maintains a rolling cash forecast, and a Spend Agent that handles employee spend requests inside Slack and Teams — routing them, applying policy, capturing receipts and posting approved expenses. A Collections Agent remains in private beta.
It integrates natively with NetSuite, Sage Intacct, Microsoft Dynamics and Acumatica, writing to the accounting system rather than dumping generic journals, and finance controls each agent's access level. An agent that posts to the ledger sits inside your segregation of duties whether or not the documented version says so.
Microsoft published the update 29.0 preview on 7 September, with general availability expected in the first week of October. For month-end the useful items are the Payables Agent matching purchase orders on line amounts and expected receipt dates, and an Expense Agent running AI policy validation to flag non-compliant or suspicious lines before approval.
The least glamorous items are the most useful: scheduled report packs that render multiple financial reports into a single PDF, G/L account usage tracing, and automatic change-logging on financial report definitions. That last one is an audit trail you previously had to keep by hand.
Anthropic's own documentation is unambiguous: on Enterprise plans the built-in browser is "off by default at launch and turns on by default starting September 10, 2026, unless you've turned it off". Any organisation that deployed Cowork and never made an explicit choice has a new default, controlled from Organization settings.
The same page carries a second line worth reading twice: network egress permissions "don't apply to the web fetch or web search tools or MCPs, including Claude in Chrome". If you assumed egress controls covered connector traffic, they do not.
=COPILOT() Function Is Withdrawn on 14 SeptemberMicrosoft's message centre notice MC1454373 confirms the =COPILOT() worksheet function stops working from 14 September 2026, with users directed to the Copilot side pane instead.
If anyone in your team embedded it in a shared workbook during the preview, that workbook breaks. A five-minute search of your finance drive is worth doing before Monday.
From September 2026 there are no more Release Plans on Microsoft Learn. New Business Central, Power Platform and Dataverse capabilities go to a rolling "AI at Work" roadmap instead, and Release Planner itself retires on 15 November 2026.
This answers the question Issue 011 left open: there is no 2026 release wave 2 plan and there will not be one. If you plan system upgrades, testing windows or user training around the twice-yearly wave documents, that planning input has gone and something has to replace it.
Claude Fable 5.1 arrived on 1 September with headline pricing unchanged at $10 per million input tokens and $50 per million output, but cache reads down to $0.25 per million — a 75% reduction.
That matters more than it sounds for finance work, because the workloads that repeat are exactly the ones that re-read the same large context: a month-end pack, a policy suite, a ledger extract. Anthropic also set out Enterprise Frontier Safeguards, under which activity data used for misuse monitoring can sit in the customer's own cloud account under their own encryption keys, at no charge from Anthropic. Read the availability line before it reaches a DPIA — "EFS will be rolling out to customers in phases, starting later this fall". It is not something you can specify today.
Docusign will expose its agreement-intelligence engine to Claude, ChatGPT, Gemini, Copilot, Slack, Salesforce Agentforce and Perplexity. For a finance team this is the contract layer becoming queryable — renewal dates, notice periods, payment terms, indexation clauses — rather than sitting in a document store nobody opens between audits.
Announced, not live. The stated launch date is 30 September 2026.
ReInvestWealth built an MCP server into its bookkeeping product from 2 September, so a user connects from their own Claude or ChatGPT settings rather than running a configuration file. It is read and write — categorisation, invoicing, reporting — and the vendor's stated rule is that the assistant prepares a summary of what it is about to do and asks before changing anything. Live.
It is a small North American SMB platform rather than anything a UK charity would adopt. The reason to read it is the pattern, which is the same gate Caseware chose, arrived at independently.
The older static 13-tool D365 MCP server retires on 1 October 2026. If anything in your environment points at it, that connection stops working in three weeks.
A survey of 687 finance and operations leaders found 31% using AI today against 66% interested, with accuracy the top barrier at 36% and only 28% comfortable letting AI decide routine finance matters. Among the 340 already piloting or using it, 69% reported less time on manual review and 51% a shorter close.
The sector split is the line charity finance directors should note. PEX's own release puts it as 48% of professional and managerial organisations using AI, while "nonprofits and media & entertainment trail at only 23%" — the two sectors are reported together, so read 23% as the band nonprofits sit in rather than a nonprofit-only figure. PEX sells spend management software, so treat this as vendor-sponsored research — but the direction is consistent with everything else this fortnight.
Further detail from the survey Issue 011 reported puts AI sixth on the finance priority list at 6.9 out of 10, up from thirteenth a year ago, with data security still first at 7.6.
The number underneath it is the one to keep: 1% of the 902 finance executives surveyed said AI is driving significant business transformation. Near-universal adoption, almost no transformation, and a priority ranking that has jumped seven places. Those three facts belong in the same sentence.
Kyle Kober, Director of Product Finance, used the Codex coding agent to automate reconciliation of product-usage data against accounting records for compute-cost reporting. The detail that matters is not the ratio but who did it: the finance team built it, not IT — "there's now the opportunity to just do it yourself."
Treat the figure as self-reported and unaudited, and note that the zero-day close his CFO has talked about is an ambition rather than an achievement.
Casey Janick of The CFO Alliance points out that boards and investors now use AI to generate questions about financial data, which pushes work back onto the reporting function rather than removing it.
His underlying point is a data-quality one. AI cannot tell whether figures pulled from the ERP, the till system and payroll are actually comparable, so it produces confident questions about the wrong metrics — "there are more tools to pretend you know why than ever have been."
TLT's September AI Brief sets out the UK Jurisdiction Taskforce view that professionals face negligence exposure both for using AI badly — unsuitable model, no due diligence, unvalidated output — and for failing to use it where a competent professional would.
If your organisation's AI position is a prohibition, that is now a risk in its own right rather than the safe option. The practical response is a documented human-validation step on any AI-produced number or narrative, which is a policy you can actually keep, unlike a ban.
The FCA's own approach page carries nothing on it and was last updated in February. The AI Input Zone that gathered the evidence closed in June, and the publication is now signalled for late 2026.
Stop waiting for it. The FCA's stated position is that Consumer Duty and senior-manager accountability already apply to AI-assisted work, so document your controls against those rather than deferring the governance until guidance lands.
The government's consultation on regulating workplace monitoring technologies closes 30 September. Its eight principles reach further than attendance and email monitoring, into algorithmic management and automated decisions about workers, so check whether anything finance or shared services runs — productivity dashboards, automated expense or timesheet flagging — sits inside scope.
Separately, CCAB's draft Statement to the Profession on the Ethical Use of AI is open for comment, with a free online discussion event on 25 September.
The GC Gibson Charitable Trust has permanently shut online applications, saying AI-written submissions "made it too easy to find us" and drove applications past 600 in a year against roughly ten grants awarded. Its site now says there is no point contacting the trust unless it has funded you before.
Whatever you think of the decision, model the consequence. If more trusts retreat to closed networks, trust and foundation income shifts from merit to relationships — and that belongs in your income risk assessment rather than in your AI policy.
Andrew Harper of Epworth Investment Management, speaking at the Faith Charities Forum, argued that charity AI adoption is running ahead of governance, and that AI should handle administration but never decisions affecting vulnerable people.
He also made a point charity finance directors rarely hear from the investment side: charities' own portfolios are part-funding the AI build-out, while the costs land on different people on different timescales.
Organisations that advise, train or sell AI tools into the sector are expected to build the regulator's 2025 AI guidance into what they say and sell, with Paul Winyard of the Fundraising Regulator warning that advice omitting it risks being "incomplete, or even incompatible with the Code".
There is no compliance mechanism for third parties, so the burden lands on charities. Add the question to your procurement checklist: ask any supplier pitching an AI tool how their proposal reflects the guidance, and keep the answer.
| Finding | Source | Date |
|---|---|---|
| 31% of finance and operations leaders use AI today; 66% are interested; only 28% are comfortable letting it decide routine finance matters | PEX State of Finance, 687 finance and operations leaders (vendor-sponsored) | 10 Sep 2026 |
| AI use splits by sector: 48% in professional and managerial organisations, while "nonprofits and media & entertainment trail at only 23%" | PEX press release (not in the CFO Dive write-up) | 10 Sep 2026 |
| Of the 340 respondents already using AI, 69% cut manual review time and 51% shortened the close | PEX — subset of 340, not the full sample | 10 Sep 2026 |
| AI has risen from 13th to 6th on the finance priority list, scoring 6.9 out of 10 against data security at 7.6 | Protiviti Global Finance Trends, 902 finance executives | Aug 2026, detail published 4 Sep |
| Just 1% of finance executives say AI is driving significant business transformation | Protiviti, same survey | Aug 2026 |
| 60% of finance leaders expect AI costs and complexity to rise substantially through 2027 | Deloitte Finance Trends 2027, 1,434 finance leaders, via CFO.com | 9 Sep 2026 |
| 70% of nonprofit leaders and staff believe their organisation is missing meaningful AI opportunities; 8% have a one- to two-year implementation roadmap | Bridgespan Group and NTEN (sample size not disclosed) | 10 Sep 2026 |
Boucher's answer to the Financial Modeling World Cup headline is that the benchmark is the wrong thing to react to, and the material change is computer use — the model driving mouse and keyboard across Excel and BI tools without manual handoffs. His practical advice runs against the marketing: don't switch tools on the back of a launch campaign, get more out of the one you already use, and reserve the expensive model for genuinely complex work.
Anaplan co-founder Michael Gould argues the real hazard is executive overconfidence in slick forecasts built on fragile foundations, and that AI spreads flawed logic across an organisation in seconds where spreadsheet errors used to compound quietly over years — "only faster, at greater scale and with fewer people able to explain what went wrong". If you have ever inherited a model nobody could explain, this is that problem with the brakes off.
As agents begin transacting with other organisations' agents, the open question is authorisation — what an agent may see, decide and do — and the platforms are already splitting into incompatible camps. The finance read-across is direct and uncomfortable: purchasing authority, payment execution, and who carries liability when nobody senior actually pressed the button.
Earlier technology waves eventually settled into something an organisation could plan around, and AI has not, so "move fast or die" produces burnout with no finish line to justify it. Their answer is to shift the strain from individuals onto the organisation — permanent infrastructure, separate cadences for stable and experimental work, learning built into the job. A useful frame for a finance team being asked to absorb something new every quarter.