Filling in the SUGB form with AI, with a source for every answer
How we built an AI assistant that pre-fills the SUGB form from the CAO and company policies. Every answer shows its source. No source, no answer.

For an organisation in the staffing sector, we built an AI assistant that pre-fills the SUGB form from the CAO (collective labour agreement) and the company's own policies. Every answer shows where it comes from. And if the answer isn't written anywhere, the field simply stays empty.
Since 1 January 2026, agency workers under their new CAO have been entitled to equivalent pay. That pay is equal in value to what a comparable employee receives at the company where they work. For that, the staffing agency needs to know what the company offers. This is what the SUGB is for: the Standaard uitvraag Gelijkwaardige Beloning, the standard equivalent pay questionnaire from the ABU and the NBBU.
The project in brief
- Client
- an organisation in the staffing sector (anonymous).
- Sector
- staffing and recruitment.
- What we did
- clickable prototype, UX/UI design, web application, AI implementation, document processing with OCR, PDF export and archive, DocuSign integration.
- Technology
- React and TypeScript, Python and FastAPI, Azure OpenAI, Azure AI Document Intelligence, Unstructured and Tesseract, FAISS, Azure Blob Storage and DocuSign.
Why does the SUGB form take so much time?
Because it is a long questionnaire and every answer lives somewhere else. The SUGB asks about pay and pay scales, allowances, travel expenses, leave, pension, the individual choice budget and a series of additional arrangements. Whoever fills it in looks everything up, item by item, in the CAO and the company's own policies. For every company, all over again.
That search work has to be precise. A missed allowance or a wrong percentage carries through into what an agency worker is paid. And the CAO is not always the whole story. A company can have its own policies that differ from it, for example in a staff handbook.
Why the rule here is: no source, no answer
Because an invented answer is worse here than an empty field. A language model will fill in a form like this without hesitation, but the question is whether it is right. So the assistant only answers a question if it finds the answer in the documents provided. It then shows the exact passage, with the file name and page number.
If it finds nothing, the field stays empty. The employee then sees a message that the documents contain no information on it.
For me, this is the core of using AI responsibly. An assistant that honestly says "I don't know" is worth more than one that fills in everything. You can read more about that approach in implementing AI in SMEs.
Do your people also fill in forms from piles of documents? Then let's look together at whether an AI assistant can prepare that work, with sources.
See how we implement AI
How did we build the AI assistant?
We started with a clickable prototype of the wizard, filled with sample data. That fixed the route through the form before any AI was involved. After trial versions of the AI analysis came the final set-up. It consists of separate document processing in Python and a web application for the employee.
A search instruction for every question
A lot of the work went into the questions themselves. For every question, we defined which section it belongs to and which keywords the assistant searches with. The type of answer and the answer options are fixed as well.
On top of that, every question got a short working instruction. Take standard weekly working hours. The instruction says: search the CAO for the normal working hours. If it says 40, 38 or 36 hours, choose that option. If it gives a different number, choose 'Other, namely' and copy the number exactly. If the assistant finds nothing, it gives no answer.
The instructions also allow for words that differ between documents. Holiday allowance is called vakantiebijslag in one CAO and vakantiegeld in another.
One company per session
When uploading, the employee states which company the form is for. Every session gets its own search index, so documents from different companies stay separate. Passages that mention the company's name get priority.
Awkward documents
The application works with PDFs, digital or scanned. It extracts the text and tables from each PDF, such as pay tables. Scanned pages go through OCR in Dutch and English. If the first reading method returns nothing, the application tries a second. If that fails too, the other documents carry on as normal.
How does the assistant fill in the SUGB form?
For each question, the assistant finds the passages that say the most about it. A language model picks the answer from those. For a multiple-choice question, it only picks options the text supports. Next to every completed field is the passage, up to 300 characters, with the file name and page number. Where there are several sources, it shows up to three.
Many options ask for a value, such as a percentage of periodic pay or an amount per kilometre. The assistant then copies that value exactly. If the model still returns 'yes' or 'no' as a value, the assistant rejects it. It then searches the passage for an amount, a percentage or a number of hours. If it can't find one, the value stays empty.
A session runs in five steps:
- The employee enters the company name and uploads the CAO and the company documents.
- The application then reads the documents and makes them searchable.
- The assistant pre-fills the steps, with the source for every answer.
- Next, the employee checks everything, makes changes and fills in the gaps.
- In the review step, the application creates a PDF for the archive.

What does the design look like?
The design is a wizard that guides the employee through the form step by step. They choose between autopilot, where the assistant fills in everything in one go, and a manual mode. In manual mode, they decide per step whether the assistant takes part. A review step at the end lists everything.
Each section shows 'Completed' or 'Attention required', with any errors and warnings. The Edit button takes you back. You can only submit once everything is in order.
The archive shows the company name, date and status for each form. From there you download the PDF or the answers as a data file. If you reopen a form, the assistant does not fill anything in again. So the employee's changes stay in place. A form can also be sent out for signature via DocuSign.

Why the assistant only answers with a source
The choice for 'no source, no answer' came from the first trial versions. In them, the model sometimes gave values that were not in the documents. They sounded plausible, but they weren't written anywhere. So we turned the approach around: the assistant only answers when it can point to the source. It became the most important lesson of the whole project, and I take it with me to every assistant we build next.
What did it deliver?
The employee no longer starts with an empty form and a pile of PDFs. They start with a proposal they can check. Every answer points to the place in the CAO or policy it comes from. Where the documents say nothing, there is no guess, just an empty field.
The judgement stays with the people who are responsible for it. The assistant reads, searches and makes a proposal. The employee decides what goes into the form.
Curious what this approach looks like for integrations rather than forms? Read the case on the quoting tool and CRM integration for De Duurzame Jongens. Want to know what an assistant like this costs? See the cost of an AI agent.
Want to know where AI really saves time in your organisation? We work it out in five working days, before you have anything built.
Book your AI auditUpdated
Get in touch with us directly. We'll get back to you as soon as we can


Your message goes straight to our team in Amsterdam.What is the SUGB?
The Standaard uitvraag Gelijkwaardige Beloning is a questionnaire from the ABU and the NBBU, the Dutch staffing industry associations. The staffing agency uses it to ask the company where an agency worker is placed for the terms of employment needed to pay equivalently, from pay and allowances to pension and travel expenses.
How do you stop the AI from making up answers?
By only allowing answers that are in the documents provided. Every answer comes with an exact passage, with the file and page number. Without a source, the field stays empty. The employee then checks the answers in the review step, before the PDF is created.
Which documents can the assistant read?
PDFs of the CAO and of company policies, such as a staff handbook. Scanned documents too: they go through OCR in Dutch and English. The assistant reads tables as well. Other file types are not processed.
Does the assistant work per company?
Yes. When uploading, the employee states which company the form is for. Every session gets its own search index, so documents from different companies don't get mixed up.
What technology does the assistant run on?
The web application is built in React and TypeScript. The document processing runs on Python with FastAPI. Unstructured and Tesseract extract text and tables from the PDFs, with Azure AI Document Intelligence as a fallback. FAISS searches the documents, a language model in Azure OpenAI picks the answer, and the files are stored in Azure Blob Storage.

