Issue 008 put a price on proving AI right — the verification tax. This fortnight the research turned to a harder question: who is actually accountable when the AI is wrong, and does anyone in the finance team know how it works?
Avalara surveyed more than 1,500 CFOs and senior finance leaders already running agentic AI and found a contradiction worth sitting with. 97% said they were confident they could explain their AI agents' actions to a regulator — while 36% had nobody responsible for understanding how those agents operate and 23% had no clear accountability for a significant AI error. Only 7% said they prioritised governance over speed. Forrester, in research for Basware, found the same shape in accounts payable: 67% of finance teams already deploy AI there, but only 39% have scaled the operating model around it.
Gartner supplied the third angle. Across 204 finance leaders, 45% of finance AI investment leans towards productivity and only 20% towards decision quality — which is precisely the mismatch that leaves a CFO reporting good progress to a board that cannot see the strategic return.
Underneath the survey data, the plumbing had its biggest week of the year. The MCP specification shipped its 2026-07-28 revision — a stateless core, authorisation aligned to OAuth 2.0 and OpenID Connect, and a Tasks extension for long-running work. Claude shipped support the same day and released Opus 5 four days earlier. Bitwave open-sourced an agentic finance toolchain. And Sage Intacct's Finance Intelligence Agent arrives on 7 August.
The pattern is not that finance is moving too slowly. It is that the confidence is running ahead of the evidence, and the gap between the two is where the audit question will land.
"97% of finance leaders are confident they could explain their AI agents' actions to a regulator. 36% have nobody responsible for understanding how those agents operate, and 23% have no clear accountability when an agent makes a significant error."
The 2026-07-28 specification landed on schedule, making the protocol stateless at its core so servers run behind ordinary load balancers rather than requiring sticky sessions, and moving long-running work into a formal Tasks extension — the pattern a month-end job actually needs, where work is started, polled and cancelled rather than held open. Authorisation now aligns with OAuth 2.0 and OpenID Connect, Dynamic Client Registration is deprecated in favour of CIMD, and Roots, Sampling and Logging are deprecated with at least a twelve-month runway.
If a supplier has built you an MCP connector into a finance system, the question to ask this month is when they will be on the new specification and what breaks when they move.
Anthropic shipped support for 2026-07-28 on release day, including the stateless core, the stronger OAuth and OIDC authorisation, and versioned extensions for Apps and Tasks, alongside connector features covering enterprise-managed authentication, observability and private network tunnels. Observability and enterprise-managed auth are the two that matter for a finance function — they are what turn "we connected the AI to the ledger" into something an auditor can look at.
Bitwave launched Bitwave Agentic on 23 July, a model-agnostic set of tools — a command-line interface, an MCP server, a data and analytics layer and an orchestration component — intended to make financial data and accounting operations directly accessible to AI agents rather than only to people. The significant part is not the product but the direction: finance infrastructure is being rebuilt so that an agent is a first-class user of the ledger, and that is a control question before it is a technology one.
The two extended their partnership on 27 July to put Claude in front of Cognizant's enterprise client base, continuing the pattern of frontier models reaching large finance functions through systems integrators rather than through direct procurement. For a finance leader, that changes who is in the room when the deployment is designed — and it is worth asking whose governance framework travels with the model.
Avalara's survey of 1,500-plus finance leaders already running agentic AI found 29% prioritising speed entirely and a further 41% prioritising it mostly, against 7% who put governance first, with more than 90% reporting moderate to significant career pressure to demonstrate ROI. Nearly 90% report some return, but only 38% describe it as at scale and 50% call it limited — so the pressure is real and the evidence is thin, which is the condition under which controls get skipped.
Research commissioned by Basware found 67% of finance teams already deploying AI in accounts payable but only 39% with a scaled operating model around it, while 76% plan to increase AI investment over the next 12 to 24 months and 68% now require demonstrable ROI before further spend. The framing in the report is the useful one for a business case: success in the next phase of AP will not belong to the teams using the most AI, but to the teams that govern it best.
Across 204 finance leaders, 45% said their AI investment leans towards productivity and efficiency against 20% leaning towards decision quality — and Gartner's point is that boards weight the second far more heavily than the first, producing a perception gap where finance reports adoption progress and the board sees limited strategic impact. Functions investing in initiatives that create genuinely new value were more than twice as likely to report high realised value.
Anthropic released Opus 5 on 24 July, positioned as coming close to the frontier intelligence of Fable 5 at half the cost per token. The practical read for finance is the one Issue 008 raised through the cost-per-task lens: a cheaper model that needs fewer correction cycles changes the economics of running a process monthly rather than once as a pilot.
Microsoft's July update makes two frontier models selectable inside Copilot in Excel — OpenAI's GPT-5.6 and Anthropic's Claude Opus 5 — and adds inline citations with deeper links to the exact content behind a response, alongside improved source attribution. Synced Copilot connectors also reached general availability, bringing organisational content such as tickets, database rows and CRM records straight into the grid.
Citations are the finance-relevant part: a figure you can trace to its source is reviewable, and one you cannot is a reconciling item waiting to happen.
The same release lets you attach a Power BI report and ask Copilot to analyse performance, identify trends and build supporting calculations in Excel, working against the underlying report data while respecting existing row-level security. That combination is unusual and worth noting: it removes the export-and-reconcile step that normally breaks the link between a governed metric and the spreadsheet version of it, without quietly handing everyone access to data they should not see. It is rolling out to production for Windows, Mac and web.
Release 3 brings the Finance Intelligence Agent, which answers natural-language questions about financial data without building a report first; an AI Gateway for third-party integrations; and intelligent three-way matching that reconciles line-item detail across purchase orders, receiving documents and vendor invoices. For a charity finance team on Intacct, the three-way matching is the one with a measurable month-end effect, and the AI Gateway is the one to ask your partner about before anyone connects anything to it.
Microsoft's 2026 release wave 2 plans put Business Central 29 into general availability from October 2026, with features rolling out through April 2027, and the finance headline is agents that handle bookkeeping tasks based on previous patterns and established business practice. Two practical notes for anyone on Business Central: the agents learn from what your team already does, so poor coding habits propagate, and SOAP web services retire permanently in favour of OData v4 REST APIs, which is an integration project someone needs to own.
The Avalara finding that 36% of finance functions have nobody responsible for understanding how their AI agents operate describes organisations with dedicated finance technology resource. A charity finance team of four, running AI through a departmental subscription that nobody has reviewed since it was approved, is not better placed — it is the same gap without the budget to close it. The practical minimum is unchanged and unglamorous: name an owner for every process AI touches, record what it does and does not decide, and put the tool on the contract register with a review date.
The most useful reframing to come out of this fortnight is that the risk in an automated process is not the data being messy — it is an exception that nobody owns. Automation that surfaces what did not match and routes it to a named person before sign-off is safer than a tidy dataset with no accountable reviewer, and it is a far easier thing to explain to an auditor or a trustee than a model's internal workings. If you are piloting anything in the close this autumn, agree who works the exception list before you agree what the tool does.
The Consultative Committee of Accountancy Bodies' draft Statement to the Profession on the Ethical Use of Artificial Intelligence, published with six case studies covering firms, industry, public sector and non-executive roles, remains open for comment ahead of its online event on 25 September. It is carried forward from Issue 008 deliberately: this is a short window in which the profession's AI guidance can be shaped rather than received, and charity finance is materially under-represented in these consultations.
The European Commission's targeted consultation on the draft Article 6 guidelines for classifying high-risk AI systems closed on 23 July, with final guidelines due for adoption by the end of 2026. Nothing has been published since, so the position for planning purposes is unchanged: high-risk obligations defer to 2 December 2027, embedded systems to 2 August 2028, and the new prohibitions still bite from December 2026.
A useful legal roundup published on 23 July sets out the timetable now running: the FCA is to decide within three to six months whether general-purpose AI tools that help consumers with savings, pensions or borrowing fall inside its perimeter, and its examples of good and poor AI practice are still expected later in 2026. The Review recommends no new AI-specific regulation, which means existing rules — and the evidence that you follow them — carry the whole load.
| Finding | Source | Date |
|---|---|---|
| 97% of finance leaders are confident they could explain their AI agents' actions to a regulator | Avalara, Agents of Change (1,500+ finance leaders) | Jul 2026 |
| 36% have nobody responsible for understanding how their AI agents operate; 23% have unclear accountability for a significant AI error | Avalara, Agents of Change | Jul 2026 |
| Only 7% of finance leaders prioritise governance over speed; 29% prioritise speed entirely and 41% mostly | Avalara, Agents of Change | Jul 2026 |
| Nearly 90% report some agentic AI return, but only 38% describe it as at scale — 50% call it limited | Avalara, Agents of Change | Jul 2026 |
| 67% of finance teams deploy AI in accounts payable, but only 39% have a scaled operating model for it | Forrester Consulting for Basware | Jul 2026 |
| 76% plan to increase AI investment over the next 12–24 months; 68% require demonstrable ROI before further spend | Forrester Consulting for Basware | Jul 2026 |
| 45% of finance AI investment leans towards productivity; just 20% towards decision quality | Gartner (204 finance leaders) | Jul 2026 |
| Finance functions investing in genuinely new value propositions were more than twice as likely to report high realised AI value | Gartner (204 finance leaders) | Jul 2026 |
The clearest single summary of where UK and EU AI regulation actually stands, covering the FCA's post-Mills timetable, the EU cybersecurity and AI action plan, the AI Omnibus deadline changes and the transparency code for AI-generated content. If you read one governance piece before your next audit committee, this is the one that will save you assembling the timeline yourself.
A practical migration guide to what the new specification changes and what it breaks, written for people who have to move something rather than for people reading about it. Worth forwarding to whoever maintains your integrations, because the deprecations have a clock on them even though nothing stops working immediately.
BCG's executive perspective on agentic AI in finance is the most structured treatment of the deploy-versus-govern question published recently, and it is honest that the constraint is operating-model design rather than model capability. Published in June rather than this fortnight, but it is the document the Avalara and Forrester findings are effectively describing from the other end.