By AIGAS
AIGAS Weekly | 19 November 2026
Research window: 12–18 November 2026
A convincing demonstration is useful. It is not the same as evidence that an AI feature works reliably with our clients, permissions and review arrangements.
This week’s developments bring that distinction into focus. Anthropic discusses stronger testing, IRIS explains where AI differs from ordinary automation, and an accounting practitioner challenges the push towards autonomous workflows. Meanwhile, an EU consultation puts impact assessment back on the agenda.
Our starting point is practical: use our AI governance framework for accounting firms to make adoption more deliberate, without making it unnecessarily complicated.
Anthropic highlights why the setup matters
On 13 November, Anthropic published guidance on building stronger safety evaluations. It reports that different configurations of one model produced differences of up to 34 percentage points in appropriate responses during its multi-turn safety tests.
That is Anthropic’s own research, not an accounting accuracy benchmark or an independent assessment of Claude.
For us, the practical implication is to test the system staff will actually use. A model accessed through a browser, an accounting connector or another application may operate with different instructions and permissions.
Record the product, settings, connected data and reviewer arrangements. A result without that context tells us less than it appears to. anthropic.com
IRIS separates AI from rule-based automation
In a 16 November article, IRIS HCM managing director Stephanie Coward distinguishes rule-based HR workflows from AI assistance. IRIS describes Staffology’s absence and onboarding workflows separately from its AI-powered support assistant.
This is a supplier explanation, not a new release or independent assurance report.
We would apply the distinction to accounting software too. Ask which step extracts information, which applies a fixed rule, which generates a suggestion and which changes a record.
Those steps need different checks. A calculation may need reconciliation; a generated explanation may need source checking; a posting action may need approval. Calling the whole product “AI-powered” does not answer those questions. IRIS
A practitioner argues for human-led AI
In a session broadcast on 12 November, Sidgrove founder Dave Sellick argued for using generative AI to support people and build repeatable tools, rather than allowing it to run mission-critical accounting workflows unattended. The organiser provides session notes and a transcript.
Sellick is also a member of Xero’s UK partner council. These are his practitioner views, not an independent product comparison.
We agree that responsibility should be designed into the workflow. However, repeatable code is not automatically correct: a faulty formula can reproduce the same mistake consistently.
Our test would be whether the chosen approach produces useful, checkable work—and whether someone has the time, evidence and authority to challenge it. ai-summit.
An EU consultation offers a structured prompt
On 16 November, the European Commission sought feedback on a draft fundamental-rights impact-assessment template.
The template concerns certain deployers of high-risk AI under Article 27 of the AI Act. It is not a universal checklist that every UK accounting firm must complete.
For firms and vendors with relevant EU activities, scope needs specific assessment. More broadly, we see a useful discipline: identify who could be affected, what could go wrong and what safeguards would change the outcome before deployment—not after a complaint. Shaping Europe’s digital future
This week’s action for your firm
Choose one AI-enabled workflow and check:
- Purpose: What may it do, and what remains outside its authority?
- Evidence: Has it been tested on representative work, including exceptions?
- Ownership: Who reviews the output and decides when testing must be repeated?
Record the tool and its owner in your free AIGAS AI Register, with a reference to the supporting test record.
Good governance gives us a clearer basis for saying yes to useful AI.
