In brief
Every tender submission comes with forms: self-declarations, reference forms, the ESPD, German Formblätter. Always the same company details, always in a slightly different place. TenderRender recognises the answer slots on the form itself and lays them over the document as coloured boxes. Green means: something can go here, click and type. Purple means: an answer is already in place, and a click on the cross removes it. The AI uses those same boxes to fill forms automatically with verified data from your Library (your previously completed forms and evidence) and your Brain (your company profile with fixed company facts).
We tested this on over fifty real tender forms from the Netherlands, Belgium and Germany. On our own benchmark the system filled in 94% of everything the sources allowed. On forms it had never seen before, that was 85 to 97%. The Dutch ESPD was filled in completely, including the exclusion-ground declarations.
Two kinds of forms, one problem
Tender forms come in two flavours. Sometimes you get a tidy PDF with real fillable fields. Click the blue box and type:
But most forms look like this. The same kind of form, only as a flat print. No field to click, just a table with white space where the answer belongs:
You see at a glance where the answer should go. Software could not, for a long time. Every bid manager knows the consequence: retyping the same twenty details for every tender. Registration number, IBAN, insurance, turnover figures, references. Error-prone work that says nothing about the quality of your bid. Word forms with a dotted line after "Name and position:" have the same problem: there is nothing to click.
Where the answers come from
The AI draws on two fixed places in TenderRender. The Brain is your organisation's memory: verified company facts such as your registration number, IBAN, insurances, turnover figures and certifications. The Library is your document cabinet: previously completed forms, declarations and evidence from earlier submissions. That is where the AI looks first. The project documents of the tender itself supply the tender-specific details, such as reference numbers and the name of the contracting authority.
How it works, step by step
The steps below show a flat PDF, but this works on all three shapes you meet: PDFs with form fields, flat PDFs without, and Word forms too. Word forms we fill in the same way, in the form's own layout. The difference: with Word, the purple boxes show what the AI has filled, and you simply edit the text in the document.
Step 1: the form arrives as it is. No pre-processing, no template we set up in advance. This is a real German form (Wirt-124, "Erklärungen zu Ausschlussgründen"), blank:
Step 2: the software draws the boxes. A measuring layer reads the form the way you do: where is the label, and where is the empty space the answer belongs in? Every place it finds becomes a green box on the document, an answer slot. Nothing has been filled yet. The form simply knows where something can go:
Step 3: the AI fills what it can prove. Through those same boxes the AI fills the form with data from your Brain and Library. Filled slots turn purple. Anything the AI cannot back with a source stays green and open:
Step 4: the chat asks what is still open, and you finish it. After the fill, the chat lists which slots are still open and why. You can complete them in two ways. Type the answers into the chat and they land in the right place on the form. Or click a green box and type the value directly in the document: press enter, and the text is there, in the form's own typeface. Made a mistake? Click the purple box and change the value, or remove it with the cross. That works even if the AI wrote it. Every action becomes a new version, so you can always go back:
The AI fills, you check
Automatic filling follows one hard rule: only data that demonstrably comes from your sources. No source? The slot stays empty and goes on the review list, with the reason attached. You walk through the list, fill the last boxes, and the document is ready for submission.
Signature, place and date always stay empty. That is not a promise but a check: every filled version is compared with the original and only saved if the pages, the printed text and the signature block are unchanged.
Yes/no declarations, such as the exclusion grounds in the ESPD, are filled by the AI only if it finds an earlier declaration from exactly the same legal person in a dated source document. Such a proposal always appears on the review list with the source document and date. You confirm whether it still holds today.
Refilling is always possible, for instance after you have expanded your Library. The AI then takes the filled form as its starting point, improves what can be better substantiated and removes what no longer holds. Each round becomes a new version, so you can always go back.
Tested on over fifty different forms
Does something like this only work on the form it was built for? That was exactly our worry. So we tested on over fifty forms in total: first a benchmark of ten to learn from, then in four rounds forms the system had never seen. The benchmark held ten types. German self-declarations with and without form fields, Dutch declarations in Word, a reference form, the ESPD. The rules were strict: no instructions beforehand, only a company profile and a library of previously completed forms. Exactly what you supply as a customer.
The first rounds scored poorly, and every miss became an improvement. Turnover figures that need labelling per financial year. Older forms showing earlier financial years, which are therefore not contradictions. Forms your company cannot fill at all, such as a healthcare reference form for a construction firm, which should stay empty. Final score: 94% of everything that could be filled from the available data was filled correctly.
Then the real test: forms the system had never seen, from six German states, the Netherlands and Belgium. Among them multiple ESPDs spanning 29 pages, a French-language Belgian offer form and German Formblätter filled with declarations. Same library, zero instructions. The first blind round reached 85 to 90 per cent; the Belgian form even received the correct IBAN and BIC. The exclusion ground declarations remained too often empty in that round, and that became the next improvement. In the final two rounds, coverage sat at 97 per cent, and the Dutch national model of the ESPD was filled completely. What stayed empty was mostly data that simply was not in the sources. Which is how it should be.
What this means for your working week
The retyping fades into the background. The AI fills what is verifiable, you walk through the review list and fill the last tender-specific boxes. What remains is the work that actually matters: the content of your bid.
The best start? Upload your previously completed forms to your Library and complete your Brain. The more complete those two are, the more the system can fill for you. Across over fifty forms, it became clear: that is the difference between an empty form and one ready for your final review.
