01
Building the file
Missing pieces become a task for the customer. Payslips, identity documents, annual accounts and social contributions arrive structured instead of as scattered mail.
The platform
The engine under financing and leasing.
Adefin is not a form with a calculator behind it. It is the whole chain: query the sources, weigh the risk, collect the paperwork, compute the offers and keep the file honest all the way to delivery.
Solvability from real sources.
An enterprise number is enough to build the picture. What is public we fetch automatically; what is missing we request from the customer, precisely.
We connect to several business and credit information providers side by side, so you are not locked to one vendor and a source can be swapped without changing how you work. Every query is stored with its moment, so a decision stays traceable to what was known at the time.
Annual accounts and credit information
VIES
CBE and Official Gazette
Integrity check
Sourced research
These sources among others. What else comes in depends on what a file needs.
Annual accounts health score
Equity
Company age
Personal guarantee
Bankruptcy
A selection of the factors. Which ones count and how heavily is yours to decide.
A score model you tune yourself.
The factors live in a catalogue, not in the code. Per factor you decide three things: how a value becomes a strength, how much it weighs, and whether it is a knock-out.
Next to the standard Adefin model, every financing partner can have their own. The same file then produces different outcomes, exactly as it does in reality.
- A standard model plus one model per financing partner
- Versioned: every score points at the model that produced it
- Knock-outs such as an invalid VAT number or a hit in the integrity check
- An AI credit officer proposes, the weighted model stays the check
Documents read
Sources queried
Research
Reasoning written
Assessment
The AI credit officer does the groundwork.
Every file is worked through the way an analyst would: query the sources, read the documents that came in, pull the figures out and hold them against each other, and look up what there is to find about the company and its directors.
That does not produce a black box. Every finding carries its source, the time it was queried and the quote it rests on, so your analyst has nothing to believe and everything to check.
What comes out goes through the weighted model and through your knock-outs. The AI does not decide, and findings from the integrity check and from the research are never counted automatically.
- Reads annual accounts, payslips and identity documents and pulls the data out
- Searches the sources and cites where every figure comes from
- Writes the reasoning behind the file, in the customer's language
- Proposes what would strengthen the file, with what it would gain
- Never decides: the weighted model and your analyst stay in control
On our own AI infrastructure.
The AI credit officer is not a pass-through to a public chat service. It runs on our own setup, with the model that fits for each task, and the files stay inside our environment within the European Union.
Every step it takes is logged like any other step in the file: which source, which model, which version, which time. An assessment from last year is still explainable three years from now.
And what it proposes always goes through the weighted model and past a person. Your files are not used to train models, for anyone.
- Our own setup: no file leaves for a public AI service
- A model per task: reading, searching, summarising, proposing
- Every AI step logged with model, version and time
- Your files train no models, not even another partner's
- Managed under ISO/IEC 42001, with the EU AI Act as the frame
Where it runs
Models per task
Logging
Your files as training data
Governance
From score to signed contract.
The file stays one thing, from the first simulation to the delivery of the goods.
02
Matching on rate cards
Per financing partner you set terms, residual values, first increased instalment and price windows. The maths is a classic annuity with a residual, ranked by monthly payment including VAT.
03
Lifecycle
Request, offer, contract and delivery, with a signature tied to the version that was signed. Every transition is kept as an event.
Everything the console does, the API does too.
The platform was built as a machine interface first and given a screen afterwards. What your people do in the console, your own system can do over the API.
Authentication with an API key per source, or a user token for partners. Webhooks signed with HMAC, so you know a notification really came from us.
Create applicationWebhook
POST /api/v1/applications
X-API-Key: …
{
"applicant": { "vat": "BE0473773635" },
"asset": { "price_incl_vat": 38500, "term": 48 }
}
→ 201 { "ref": "APP-7F2K", "status": "received" }
POST /api/v1/applications/APP-7F2K/score
→ 200 { "score": 72, "tier": "B", "approved": true,
"actions": [ { "add": "personal_guarantee", "gain": 8 } ] }
What that looks like in the console.
Today's work at the top: what has to go to an analyst, which documents need a judgement and which files are still missing data, with the pipeline and the latest applications beside it. Screenshots from the product. Names, numbers and amounts have been replaced.
The model, and what it makes of a file.
On the left you tune the model: a grade per factor, a weight, and whether it blocks.
On the right the result on one file: every factor with its grade, and underneath what pulls the score down next to what dealing with it would be worth.
The questions that go with this.
- What does the AI actually do?
The groundwork: query the sources, read the documents that came in, pull the figures out, look up what is known about the company and its directors, and write the reasoning behind the file. Every finding carries its source, the time and the quote.
Weighing and deciding, it does not. That stays with the financing partner's score model and with their analyst.
- Does a model decide on a file?
No. Models propose, people decide. Findings from the integrity check and from the research are never scored automatically: they stay open for assessment, with their source attached.
- Can we still explain a decision years later?
That is what it is built for. Every source with the time it was queried, the version of the score model that produced the score, who changed which parameter and when, and every step as its own event with its time and its actor.
Want to see it on your own files?
We take a request you know, and show which sources answer, what the score does and which offers come out.