Newsletter

The AI assistant Moneybird is testing has the same permissions as the user

Moneybird's AI assistant uses your permissions, DPG tests GPT-NL in the newsroom, and an Anthropic customer says the model admits what it doesn't know.

By Piet Baudoin · 11 November 2026

The AI assistant Moneybird is testing can do exactly what the user can do: if you can only look, you cannot ask it anything. Also: DPG Media is testing a Dutch language model in the newsroom, and Anthropic’s new model comes with a customer who says it admits what it doesn’t know.

What can an AI assistant do in my bookkeeping?

Moneybird, a Dutch bookkeeping package, is testing an AI assistant that you can ask about your invoices, transactions, contacts and VAT from the bottom of every page. The help article keeps it short: “The feature has the same permissions as you” (translated from Dutch). If you only have read access, you are out: “Then you cannot start a conversation” (translated from Dutch). The assistant never makes a change by itself. It prepares a proposal and you click “Allow” or “Decline”.

Do you still know who has which permissions in your bookkeeping package? With an assistant like this, those permissions also decide what the AI may look up and prepare for someone. The beta is switched on for a limited number of company administrations that Moneybird picks at random; you cannot switch it on yourself, not even through support. The help article gives no publication date.

Source: Moneybird, help article “Vraag het Moneybird: de AI-assistent (bèta)”, read on 11 November 2026

Does AI write in your company’s voice by itself?

DPG Media, a Dutch and Belgian media company, is testing with the research organization TNO whether GPT-NL, the Dutch language model, can help newsrooms: first with headlines, then as a digital copy editor that reads along while you write. A piece for de Volkskrant follows different style rules than a Flemish title, so the model also has to understand “the editorial identity of a title” (translated from Dutch). That does not happen by itself: “The first tests show that a usable model does not come about by itself” (translated from Dutch). At the end of this phase, editors from two DPG titles will review the headlines; the whole study runs until the end of January 2027.

The model learned from the archives of more than thirty news media. “Part of any revenues flows back to the copyright holders” (translated from Dutch). No amount is given. TNO helps develop GPT-NL itself, so it is not an outsider.

Source: TNO, 10 November 2026

Does an AI model say when it doesn’t know something?

On 9 November, Anthropic released Claude Fable 5.2, the newest model from the maker of Claude. The announcement quotes a customer, Adhyyan Sekhsaria of the company Cognition: “Claude Fable 5.2 says it doesn’t know things instead of hallucinating.” He adds that the model called one of its own earlier estimates wrong and only wanted to give a number once it had measured the change.

On 9 November I wrote here that within two years you will no longer choose an AI agent by how smart it sounds, but by how honestly it says what it isn’t sure about. The same day, a maker put exactly that in its announcement. It is still one customer’s quote on the seller’s own page, not a measurement. Whether it holds up in your work, you only see once you put the model on your own data.

Before it goes online, this newsletter is reviewed independently by a second AI model that did not write it. This morning that reader stopped this edition: it said Anthropic wrote this two days after my prediction, when it was the same day. The model that wrote the sentence had no doubt about it.

Source: Anthropic, 9 November 2026

Where are we heading?

For GPT-NL, more than thirty news media opened their archives to train the model, and part of any revenues flows back to the copyright holders. At GPT-NL, whoever contributes knowledge therefore shares in whatever the model may earn. I expect ordinary companies to make agreements like that within five years. An installation company that writes down well how it fixes a rare fault makes that knowledge available to AI on its own terms, and earns from it when a company on the other side of the country is helped with the same fault. Know-how then becomes a second product next to the trade itself. I also expect models to learn to recognize where an answer really comes from; then customers find the company that has written down its trade best, even if they never knew it. What a company shares, and for what payment, it sets down itself, as the news media did with GPT-NL.

What does this mean for your business?

The thread today: an AI assistant works within what you give it. In Moneybird, that is your permissions and your click on “Allow”. At DPG, it is the newsroom’s voice, which the model first has to learn. The better your own way of working is written down, the more the AI’s work looks like how it is done at your company. Read how to give AI agents your company context.

Bombos works with the same boundary: the AI agent prepares, a person approves. Take payment reminders: they are ready on time, in your voice and with the current balance, and you check them and send them. That is the first task; your team teaches Bombos the next one itself, and the free scan shows where to start.

Sources

Every source was opened on 11 November 2026; every quote appears in it word for word, quotes from Dutch sources in our translation.

Moneybird: Vraag het Moneybird: de AI-assistent (bèta) (undated) · TNO: AI met gevoel voor onze taal en context: DPG Media test GPT-NL op de redactie (10 November 2026) · Anthropic: Claude Fable 5.2 and Mythos 5.2 (9 November 2026)

The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.