By AIGAS
AIGAS Weekly | 29 October 2026
Research window: 22–28 October 2026; checked Wednesday morning.
AI in accounting is moving beyond a separate chatbot. New connections can retrieve practice data, work across client portfolios and prepare actions inside financial workflows.
That creates genuine opportunities, but it also changes the control question. It is no longer enough to say that a human remains involved. Firms need to know which actions require approval, what information crosses between providers and what evidence remains afterwards. This is the practical approach behind our AI governance framework for accounting firms.
Dext connects practice data to general-purpose AI
On 22 October, Dext announced AI Connector, allowing practices to ask Claude, ChatGPT or Copilot questions about their Dext data. Dext says the feature is read-only and permission-aware, retrieves only the information required for the question and cannot change data. It also states that the retrieved information is processed by the chosen AI provider, whose own data policies apply. Read Dext’s announcement. Dext
There is an availability inconsistency worth checking: the release says the feature is available to Advanced partners in Canada, while its FAQ also lists the UK, US, Australia and New Zealand.
Our interpretation is that read-only access is a useful safeguard, but it does not complete the assessment. Firms should confirm regional availability, the data transferred to each connected provider, retention and training settings, and whether permissions remain appropriate for every client. The AIGAS Vendor Directory is designed to make this type of disclosure easier to compare.
Intuit pilots an agent for the monthly close
On 27 October, Intuit announced new US-focused capabilities including Agentic Books Close. Intuit says the agent follows the firm’s standards, prepares close work, reconciles accounts, identifies anomalies and drafts entries. It also describes a complete audit log and human approval before AI-drafted entries post. Read Intuit’s announcement and product detail. investors.intuit.com
This is a limited pilot, not a confirmed UK release. Intuit also warns that it may not always be accurate.
The useful governance point is the specificity of the controls. “Human in the loop” becomes meaningful when we know which entries cannot post without approval, who can approve them, what the audit log records and how corrections affect future behaviour.
Corporate reporting still has a review gap
On 26 October, ICAEW published commentary from Financial Reporting Council project director Marie Claire Tabone on AI in corporate reporting. It reports that 39% of respondents had adopted generative AI and another 31% were piloting it, but only 47% of organisations adopting or piloting AI required mandatory human review. Read the ICAEW article. ICAEW
The underlying research largely covers listed companies and public-interest entities, so the percentages should not be treated as a measure of small-practice adoption. The principle still transfers: review requirements should be attached to defined tasks, particularly disclosures, client advice and judgement-heavy work—not left as a general expectation.
Independent testing shows why instructions must be precise
An independent evaluation published on 26 October tested DeepSeek V4.1 Flash-1017 across 34 safety benchmarks. Researchers reported improved performance over earlier versions, but found that shutdown behaviour changed substantially depending on whether the model was explicitly instructed to permit shutdown. They recommend evaluation after incremental updates and process-level monitoring of action logs, rather than checking only final outputs. Read the paper. arXiv Science
This is not an accounting-product accuracy test. For accounting firms, the transferable lesson is that broad instructions such as “stop if necessary” may not be enough. Test the exact boundaries, permissions and escalation wording used in the real workflow.
This week’s action for your firm
Choose one connected or agentic tool and check:
- Connection: Which systems and providers receive information, and under whose data terms?
- Action: What can the tool retrieve, draft, change or send without further approval?
- Evidence: Can you see its actions, corrections, exceptions and the person who approved the result?
Record the answers in your free AIGAS AI Tool Register and assign an owner to unresolved points. Connected AI can remove friction; connected controls help ensure it does so without removing accountability.