The Legal edition
This is a real edition from 20 October 2026, at the Practical level, in plain language.
Meta's new model scores highly on financial reporting, but struggles with day-to-day bookkeeping.
Meta's accounting model misses transactions
Muse Wonder scores highly on reporting, but less well on day-to-day bookkeeping
Meta released Muse Wonder on 12 October. Meta calls it its most powerful model so far. It scored 83.2 per cent in an accounting test. That put it joint second out of 57 models. It scored higher than all the OpenAI and Google models tested. Only Grok 4.5 did better, with 84.2 per cent. Its predecessor, Muse Spark 1.1, scored 61.4 per cent.
Behind that strong score is an important difference. Muse Wonder was very reliable at reporting, for example with cash flows and financial statements. It made more mistakes with day-to-day transactions. The model made errors with some supplier payments. Results varied for month-end closing and complex bookkeeping.
In the test, Muse Wonder did the same tasks as 56 other AI models. So a high overall score does not mean the model can handle every financial task. An average score can hide differences between types of work. Its weak points are classifying and recording individual transactions. That can matter in a file containing financial information.
Note: the test covers accounting work only. It says nothing about how well Muse Wonder handles legal work. There was also no safety report when the model was released.
A high score says little about your work. Muse Wonder has only been tested on accounting, not legal tasks. When it comes to evidence, what matters is whether a document is genuine, not whether it sounds plausible. AI can make an invoice or confirmation convincing and still fake. If you use AI for case files, ask exactly what the provider has secured and tested. Microsoft and NautaDutilh have not yet shared many details. The new standard EN 18286:2026 may offer guidance on quality management for high-risk AI.
NautaDutilh is building legal AI with its own data
NautaDutilh is launching Bluefield. On this platform, lawyers and software developers will build AI for legal work together. The firm will use its own data. It says this meets existing security and confidentiality requirements. Bluefield will add to the AI the firm buys, not replace it. Lawyers bring in legal questions from their work. Software developers work with them on solutions. These should fit those questions and the firm's own data. This is different from simply buying a general AI tool. The firm wants to build some of the technology itself, around its own way of working and its clients. It is not yet known which applications will come first. The source gives no specific example yet. Read more →
Microsoft adds protection against hidden instructions
Microsoft says it is adding a new layer of security to Microsoft 365 Copilot. It is designed to protect against indirect prompt injection. These are hidden instructions in documents or emails that Copilot reads, for example. So the instruction is not in what you type yourself. It is hidden in content the system has to process. Microsoft names three measures. Untrusted content is kept separate. Copilot is allowed to do less with it. The system also records which information Copilot can trust. The blog post is only a brief announcement. It gives no test results. It also includes no examples of situations where the protection takes effect. It says nothing about the limits of the security layer. Read more →
EU publishes standard for quality management of high-risk AI
The European Commission published EN 18286:2026. This is the first harmonised standard linked to the AI Act. That is the name of a European standard connected to the law. A standard like this turns legal requirements into practical steps. This one covers quality management throughout the life of an AI system. It describes a quality management system. This is a set way of monitoring quality. It can work alongside existing systems such as ISO 9001. If a provider of high-risk AI follows this standard, it is presumed to meet the requirements in Article 17 that the standard covers. This means the provider is assumed to meet those requirements. It is not general approval of every AI system. The standard also does not replace existing quality arrangements. Read more →
NautaDutilh links its own AI to confidential firm information
With Bluefield, NautaDutilh wants to build AI around specific questions from legal work. The platform uses the firm's own data. According to the source, it must meet existing requirements for information security and confidentiality. The firm is not stopping work with its current AI providers. The first applications have not yet been announced. The firm is combining lawyers' knowledge with that of software developers. This means the firm's own data and ways of working can inform solutions that fit legal practice. A lot remains unclear. The article does not say how the data will be protected technically. It also does not say which data will be used. The source describes no specific safeguards or checks. So it is not yet possible to say what this means in practice for case files or confidentiality. Read more →
AI fraud can make documents convincing, but not genuine
An opinion piece on Accountant.nl warns that AI can make fraud inside an organisation easier. AI can create an invoice or customer confirmation that looks genuine. But it never came from the named sender. This shifts the focus when checking annual accounts. The question is less whether documents match each other. It is more whether they are genuine and where they came from. AI can create several documents that support each other. The fact that they match does not prove that a transaction really happened. The authors call this a shift from consistency to authenticity. In other words, from 'do they match?' to 'are they genuine?' They want more attention paid to independent sources. This affects the basis of the evidence on which an audit relies. Read more →
KPMG finds less confidence in AI's value
KPMG conducts research worldwide every quarter. It found a fall in the share of Dutch organisations that get meaningful business value from AI. It was 82 per cent in the second quarter and 68 per cent in the third quarter. Confidence in good oversight and risk management also fell, from 77 to 62 per cent. The research asks directors and senior managers how they assess the value, oversight and progress of AI. According to KPMG, organisations take a more critical view as they gain experience. There is a wide gap between widespread use and measurable growth. Of the organisations surveyed, 36 per cent say they use and improve AI across the whole organisation. Only 4 per cent report measurable business results and growth from AI. Read more →
AI saves less time in recruitment than expected
According to a ManpowerGroup survey, only a third of participants think AI makes recruitment faster than it was in 2025. Forty-two per cent see no difference. A quarter say recruitment takes longer. In a separate survey, 33 per cent of US recruiters say CVs polished by AI are their biggest source of stress. These are CVs that have been improved with AI. The figures come from different studies. They measure the experiences of recruiters and managers. They do not give a fixed result for every recruitment system. Together, they show that AI does not automatically save time. The sources do not say how many applications are rejected or delayed as a result. Read more →
AI and tech, made simple.
This is a sample of the Legal edition. Subscribers get it in their inbox every week.
Subscribe
inklaretaal · Amsterdam · © 2026
This is the generic Legal edition. Subscribe and you get it made to measure: tuned to your level, your job and the topics that matter to you. That way every week gets a little more personal.
Start for free