For accounting, administration and tax firms
Bee Flow is an AI workspace for accounting, administration and tax firms: staff work with AI in one place that belongs to the firm, on a server you choose. On this page each condition from AI-leidraad 2, the AI guide NBA and NOREA published in June 2026, stands next to the setting that answers it, with the licence it needs.
Plan a call Try the live demoNBA and NOREA, AI-leidraad 2: AI Toegepast, June 2026, section 2.1
In English: this data must stay in a secured, closed environment. The sentence is about client and personal data, and it comes from the guide of your own professional bodies. The guide is your rule. This page shows which setting in Bee Flow goes with each condition, which licence it needs and what is missing. Whether your firm meets the guide stays your own judgement.
Read the guide (PDF, Dutch)
78% of Dutch accounting firms use AI, and 49% say they do not know how to get the most out of it in their own operations. Source: Exact MKB Barometer, reported by accountant.nl on 25 June 2026.
Your firm's own AI chatWhen part of that use runs on private accounts and no AI policy is written down, nobody knows which client data went where. The guide says client and personal data may not simply be entered into public AI models. A policy of one page is a start, and one place for AI that belongs to the firm makes it workable.
Starter AI policyThe guide on confidentiality
The guide asks that client and personal data stay in a secured, closed environment (section 2.1). These settings go with that condition.
The personal-data filter (Privacy Shield) detects 21 kinds of personal data, including the Dutch BSN, checked with the elfproef, and IBAN, and can block a message. This is in the Free licence and needs the guard service installed. Detection is machine learning and can miss.
Names, BSNs and IBANs are replaced before the prompt leaves and put back in the answer. The model reads [person_1] where your colleague typed a name. Replacement needs the Pro licence and the guard service.
One switch limits every model level to EU-hosted models. A model on your own hardware keeps the text on your own network. The switch covers AI models. Connected apps keep their own settings.
On every licence. Self-hosted, the licence is checked offline and by default nothing is sent to Bee Flow. New organisations start with managed-key encryption on every licence. Zero-knowledge encryption for private chats is the stronger option, in the Enterprise licence.
Live demo
What the filter looks for, what it does when it finds something and what leaves your network. Take a question about a payslip with a BSN and an IBAN in it. With replacement on, the model receives the same question with a placeholder where each value stood, and your colleague reads the answer with the real values back in place.
This is the real interface running on sample data. Nothing you do here is saved, and nothing leaves your browser.
The guide on responsibility
The guide calls AI a tool with which the professional keeps “de volledige en ongedeelde verantwoordelijkheid”: full and undivided responsibility (section 2.1). These settings keep a person in charge.
Where an approval step is built into an automation, the run waits for a person: one approver, a group or a chain. Bee Flow records who decided, when, why and what they saw. Approval steps need the Pro licence.
For each connected app and each action an administrator sets what an assistant may do: reading runs directly, writing can ask first, sending always asks. In the Free licence. The product calls an assistant an agent.
AFAS Profit and NMBRS can be read and cannot be changed, in the Free licence. This holds for those two systems. Mail and files follow the rights you set per action.
An assistant can start when a new mail arrives in Gmail or Microsoft 365. Per assistant you choose: read only, prepare a draft, or reply in the same thread. A reply falls back to a draft when the run read anything outside that thread.
Fictional data. Nothing here is sent.
Hello Anja, for the year-end accounts of Bakkerij Voorbeeld we still need two documents: the bank statements for Q4 and the lease contract for the new van. You can reply to this mail with the files. Kind regards, the team.
An approval step in an automation: the run waits until a colleague decides. The approver sees the client, the step that is waiting and the exact message that would go out. The decision is recorded with who decided, when, why and what they saw. Approval steps need the Pro licence.
Before go-live and after every change
What if it sends something wrong to a client? This is how you find out before a client does.
A dry run reads your real data and simulates every send, create, update and delete. Anything that is not known to be read-only is treated as a write. In the Free licence.
Runs you trust are kept and replayed after every change, with a comparison step by step. In the Free licence.
An AI step is held to a schema and to allowed values. It is retried once and then fails visibly. Reading is never 100% certain, so doubtful cases go to a person.
An assistant can be checked against saved test questions, with a verdict per question. In the Free licence.
The guide on validation
The guide asks you to record in the file “gehanteerde prompts, zoektermen, geraadpleegde bronnen, de signalen, uitgevoerde verificatiestappen en de uiteindelijke conclusies”: the prompts and search terms used, the sources consulted, the signals, the verification steps and the conclusions (section 2.2). It also says that AI output alone can never count as sufficient and appropriate audit evidence. We say the same.
An answer from a knowledge base cites the document, page or row it came from. In the Free licence. This holds for knowledge-base answers, and a plain chat answer has no source list.
Under an answer you can open which tools and documents were used and what the personal-data filter did. In the Free licence.
Every run of an automation is kept step by step, and a failed run can be retried from the step that failed. In the Free licence.
Outgoing calls from tools and automations are logged with host, company and country, without content, on every licence. The report is part of Advanced monitoring in the Pro licence. Calls to AI models are not on it.
Live demo
The documents a knowledge base keeps up to date, with what came out of each file and what was skipped. An answer from this base cites the document it used, and your colleague can open that source to check it.
This is the real interface running on sample data. Nothing you do here is saved, and nothing leaves your browser.
The mapping
Quotes from NBA and NOREA, AI-leidraad 2: AI Toegepast, June 2026, sections 2.1 and 2.2. The right column says what Bee Flow offers for that condition. Meeting the guide stays the work of your firm.
The quotes were checked against the published PDF on 16 November 2026. This table is no statement that a firm using Bee Flow meets the guide.
| What the guide asks | The setting, and its licence | |
|---|---|---|
| A closed environment | “Deze data moet in een beveiligde, afgeschermde omgeving blijven.” (2.1) | Your own server or Bee Flow Cloud in the EU, on every licence. One switch limits every model level to EU-hosted models, and a model on your own hardware keeps text on your network. |
| Client data and public models | “Cliënt- en persoonsdata mag niet zomaar in publieke AI-modellen worden ingevoerd” (2.1) | The personal-data filter detects and blocks 21 kinds of personal data (Free, guard service installed) and replaces them before the model sees them (Pro). Detection can miss. |
| Responsibility | The professional keeps “de volledige en ongedeelde verantwoordelijkheid” (2.1) | An approval step with a record of who decided, when, why and what they saw (Pro, where a step is built in). Rights per app and per action: read, ask first, always ask (Free). |
| No black box | “Daarbij is black box-AI niet te verantwoorden en een onacceptabel kwaliteitsrisico.” (2.1) | Under an answer: the tools and documents used and what the filter did (Free). Every automation run kept step by step (Free). |
| Sources in the outcome | “In de uitkomsten van de AI-toepassing verwijzingen opnemen naar de gebruikte bronnen.” (2.2, point 1) | Answers from a knowledge base cite the document, page or row (Free). A plain chat answer has no source list. |
| A record in the file | “Het vastleggen in het dossier van gehanteerde prompts, zoektermen, geraadpleegde bronnen, de signalen, uitgevoerde verificatiestappen en de uiteindelijke conclusies” (2.2, point 7) | The trace under an answer and every run step by step stay available in Bee Flow (Free). Outgoing tool and automation calls are logged without content, with the report in Advanced monitoring (Pro). Putting it in your file is your own step. |
| A second AI as a check | “Dit kan bijvoorbeeld op basis van een tool of LLM van een andere leverancier.” (2.2, point 8) | Nine hosted model providers and models on your own hardware. You can put the same question to a model from another provider. |
| AI output is not evidence | “Alleen AI-output kan nooit als voldoende en geschikte controle-informatie worden beschouwd.” (2.1) | No setting replaces this. Bee Flow prepares, and your professional judges. |
Honest about limits
AFAS Profit and NMBRS, both read-only, and Outlook, Gmail, OneDrive, Google Drive and Nextcloud. The built-in integrations are in the Free licence.
There is no connector for Exact Online, Twinfield, SnelStart or Moneybird, and none for practice software. Collecting documents works through mail and folders.
Another system goes through that system's own API and is agreed per system. We say in the first call whether that pays.
The first process
We set up the first process with you. You weigh it against the hours your team spends on that work now.
Bee Flow asks clients for the documents that are missing, sends reminders from your own address and files what comes in. A good one to have running before the busy season.
A recording of a meeting becomes notes per speaker, with a summary, decisions and action items, processed where you choose. From the Pro licence.
How we work
20 minutes. We ask how the work arrives, what happens next and which systems you use. You get an honest go or no-go.
With your own examples we map the steps, the exceptions and who decides. You get a planning and a fixed price.
We set it up in Bee Flow, connect your mailbox and systems and test it on your own examples in a dry run that sends nothing.
It goes live step by step, with approval steps where a person decides. After that we manage it per month: we keep an eye on it and adjust it.
Forward this page. The facts they will ask for are written down, including what is missing.
Prompts, files and answers: where they go, who can reach them, which model providers are involved and what is logged.
Sign-in, roles, retention and encryption, and what is not covered.
One command with Docker on amd64, or Bee Flow Cloud in the EU.
FAQ
Keep it for Word, Excel and Teams: it goes deeper there. Bee Flow is for work with client data. Personal data can be replaced before a model sees it (Pro, with the guard service), you choose the model and where it runs, AFAS Profit and NMBRS can be read and not changed, and an approval step can sit before anything is sent (Pro). Two dated facts to weigh yourself: on 13 July 2026 the Dutch Data Protection Authority published a tool for organisations that want to use generative AI, to check beforehand whether that use is lawful, and on 27 May 2026 SURF kept its risk rating for Microsoft 365 Copilot at orange and left it to each institution to accept the remaining risks. Last checked: 16 November 2026.
No, there is no built-in connector for those. Built in are AFAS Profit and NMBRS (read-only), Outlook, Gmail, OneDrive, Google Drive and Nextcloud. Collecting documents works through mail and folders. Another system goes through that system's own API and is agreed per system. We say in the first call whether that pays.
Check them, as the guide asks: AI output alone can never count as sufficient and appropriate audit evidence. Bee Flow gives you the means to check. Answers from a knowledge base cite their source, an assistant can be checked against saved test questions with a verdict per question (Free), and an approval step can sit before anything goes out (Pro). We make no claim that an answer is correct.
No. You can run Bee Flow on your own server, so client data never reaches us, and the source can be read (source-available under the Sustainable Use License). The page for IT and privacy officers lists the technical details, including what is missing.
Whoever runs a server can read the data on it that is not zero-knowledge encrypted. On Bee Flow Cloud that access is bound by the processing agreement, which you can request from us. Self-hosted, your firm or your IT partner is the operator.
A self-hosted install keeps running. Without a valid key it falls back to the Free licence and deletes nothing, and approvals that are waiting can still be decided. New steps that need a paid licence stop.
The licence is per user, and the prices are on the plans page. Use of an AI model is paid to the provider at cost. Setting up and the first process have a fixed price, agreed before building starts. The cost page explains what drives that price.
No. The web app works in the browser on a phone.
Read more
What happens to what your staff type into ChatGPT, what differs per subscription and which questions to put to any AI supplier.
Read the articleRead more
The weeks before a deadline, when the work could start but the documents are not there: what can be automated, and why judging the figures stays with you.
Read the article20 minutes. An honest go or no-go, and what a fixed price would look like.
Plan a call Try the live demo