By AIGAS
AIGAS Weekly | 24 September 2026
Research window: 17–23 September 2026; checked Wednesday morning.
“Human oversight” sounds reassuring. But who is reviewing the work, what can they see, and do they have the knowledge to challenge it?
This week’s AI accounting news brings those questions into focus. Xero has outlined document-processing improvements, EY has explained its approach to audit agents, and OpenAI has proposed principles for independent assessment.
For us, the opportunity is straightforward: use AI to reduce repetitive work while making responsibility easier to understand and demonstrate.
Xero’s document update: check what happens after extraction
On 18 September, Xero announced its move from Files to Documents, alongside Smart Document Capture improvements. These include credit-note support, switchable automatic document matching, upload/view-only permissions and alerts before deleting linked attachments.
Importantly, UK and Australian tax-number validation appears in the planned features, not the current improvements. Firms should confirm access in their own subscription before changing procedures.
Our view: the useful governance question is not simply whether AI reads a document accurately. It is what happens next.
Who checks an extracted amount? Can an incorrect match be spotted and corrected? Does everyone uploading documents need wider financial access?
We would test those steps with representative documents before expanding automation. That is the practical approach behind our AI governance framework for accounting firms: understand the activity, assign responsibility and check the controls.
EY: a human reviewer still needs the fundamentals
In an ICAEW interview published on 22 September, EY UK partner Richard Harrison describes task-specific agents, restricted access, testing and human review. The article discusses deployment earlier in 2026; it is not a launch announcement this week.
Harrison also argues that trainees still need core accounting knowledge to challenge AI output.
We agree. Putting someone between an AI result and a client does not automatically create effective oversight. They need the competence, evidence and time to review it—and permission to reject it.
For smaller firms, the lesson is proportionate. Choose a defined task, agree what a satisfactory result looks like and give the reviewer access to the underlying records. Prompting skills help, but they do not replace professional understanding.
OpenAI: independence needs more than an external name
On 22 September, OpenAI published proposed principles for third-party assessments. They address assessment scope, evidence access, conflicts of interest and assessors’ editorial independence.
These are the developer’s proposals for frontier-AI assessment, not new requirements for UK accounting firms or proof that a particular system is safe.
Our takeaway is a supplier-checking question: when a vendor says its AI has been independently assessed, what exactly does that mean?
Ask what was tested, what was excluded and whether the assessment covers the version and use case you intend to adopt. Ask how relevant commercial relationships were managed.
A scoped assessment can be valuable without answering every question. Understanding its limits helps firms make better decisions.
Xero and Anthropic: supplier safeguards are only part of the picture
In commentary published on 23 September, Xero’s James Bergin discusses human judgement and system boundaries, drawing on an earlier Xerocon US conversation with Anthropic’s Lina Ochman. This is vendor commentary, not an independent evaluation.
We welcome the emphasis on suppliers building safeguards into their products. A small practice should not have to engineer every protection itself.
But our view is that supplier controls and firm governance answer different questions. The supplier controls the product; the firm decides which client work it is suitable for, who may use it and how outputs are reviewed.
Both matter.
This week’s action for your firm
Pick one AI-enabled workflow and make three checks:
- Permissions: Can it access only the information and functions it needs?
- Review: Can the reviewer inspect the evidence, challenge the result and stop the process?
- Assurance: Do supplier claims explain scope, limitations and independence?
Record the workflow and its owner in the free AIGAS AI Tool Register. Good governance should make useful AI easier to adopt—and easier to stand behind.