Filling a client wealth form from the client's own documents: a test on a real insurer's form
ID card, life-insurance statement, property-company and holding accounts in; the insurer's official form out. What fills itself, what doesn't, and what it costs.
Oct 16, 2026
A French mortgage broker or wealth manager spends part of every client file copying. The client's name and date and place of birth are on their ID, their address on a statement, the value of their life insurance on another. The fiche patrimoniale, the client-wealth form, gathers all of it. Then the bank wants the same information on its own template, and the insurer on a third.
We wanted to know how much of that work an extraction system can do on its own, and at what cost. Not on a prepared demo: on a real form, filled by hand in 2024 by a real client, which gave us the answer key.
Contents
How we ran it
The form. A large French life insurer's (Swiss Life) supplementary client-information form, three pages, and the policy-loan request that goes with it, three more. Both are fillable PDFs: 137 fields on the first, 150 on the second, so answers can be written straight into the official form.
The answer key. The file is mine. I filled both forms by hand in May 2024 for an operation on my own contract. Read field by field, they give 80 answers to score against.
The input documents. Six, the ones a client actually sends their adviser:
| Document | Pages sent |
|---|---|
| National ID card, both sides | 2 |
| Bank details slip | 1 |
| Annual life-insurance statement | 4 |
| Property company's annual accounts (the balance-sheet pages, out of 19) | 5 |
| Holding company's annual accounts (balance sheet and tax return, out of 24) | 6 |
| A furnished-rental investment study | 1 |
We deliberately left out the tax notice. In France it covers the whole household, so it carries another person's income, and that person had not asked to be in a test. It is a limit of this test, and we come back to it below.
The method. Each document goes through our extraction service with a template for its type: an ID card's fields are not a set of accounts' fields. Two models from different families (Gemini and Mistral) read every field without seeing the other's reading. If they agree, the field is verified. If they disagree, a third (Claude) re-reads it blind, and it can only settle the field by confirming a reading already on the table. A field everyone reads as empty needs corroborating too. Then deterministic code, with no model in it, assembles the values and writes them into the form.
Results
Of the 80 answers in the key, 41 could in principle come from the documents. Here is what came out:
| Outcome | Fields | Which |
|---|---|---|
| Filled and right | 34 | Identity, date and place of birth, nationality, ID number and dates, contract name and number, client ID, full address down to the flat number, occupation and social-security regime |
| Filled but different | 2 | The split of wealth between financial investments, property and business assets |
| Left empty | 5 | Postcode of birth, the ID card's issuing authority and place of issue, business sector |
| For the client to answer | 39 | Everything declarative (next section) |
Three things stand out.
The right fields are exactly right. Not roughly: the 13-digit contract number, the ID number, dates in the form's own format, a flat number pulled out of an address line that mixes it with the building's name. This is where a one-character error gets a file rejected.
The system can say it does not know. The two readers disagreed on the property price in the rental study. The field was not filled with a guess: it went to review, with both readings. That is the behaviour you want on a file a bank or an auditor may challenge.
The empty fields are legitimately empty. The 2021-format French ID card no longer prints its issuing authority: all three models confirmed the field was blank, which is true. The postcode of birth is not on the card either, but a lookup table derives it from the town. As for the sector, the holding company's accounts say "fund management", while I had declared my actual line of work, software. The document is right about the company; the client is right about himself.
The two differences are worth a closer look. Computed from the documents, the wealth split comes out at 12% financial investments, 15% property and 73% business assets. I had declared 20, 40 and 40. Neither is wrong: I estimated net values off the top of my head, the calculation starts from gross book values, and the property company's accounts show, for instance, partner loans larger than the property's value. Reconciling the two is exactly the adviser's job, and it is easier with the figures in front of them than with an estimate.
What cannot be automated, and why
Almost half the form, 39 of 80 fields, cannot come from any document, however good the extraction. They are declarations:
- family situation, marriage or civil-partnership regime, number of dependants;
- politically-exposed-person status, for the client and their close relations;
- tax residence and US tax status (FATCA);
- the purpose, amount and object of the operation;
- the origin of the wealth and the split of income by source;
- email address, place and date of signature.
Some of them move to the document side with the tax notice we left out: tax number, household income, income bracket, family situation and number of dependants. With it, more than half the form comes from documents. The rest takes a short client questionnaire, a few minutes long, whose answers go into the same form.
So the realistic product is not "AI fills in your form". It is extraction of whatever is written down somewhere, plus a questionnaire for whatever is not. The gain is real without being magic: the adviser stops copying, and spends the time checking and asking the right questions.
What it costs
| Document | Pages | Cost |
|---|---|---|
| ID card | 2 | $0.021 |
| Bank details | 1 | $0.009 |
| Life-insurance statement | 4 | $0.027 |
| Property company's accounts | 5 | $0.063 |
| Holding company's accounts | 6 | $0.228 |
| Rental study | 1 | $0.021 |
| Total | 19 | $0.37 |
The documents run in parallel. The slowest, the holding company's accounts, took 59 seconds.
Cost follows how dense the pages are, not how many there are: a company tax return is a page packed with figures. In this test, the private client's own documents (ID, bank details, life-insurance statement) came to under ten cents together.
One practical note: one of the models refused the full 19- and 24-page sets of accounts. A first step that picks out the useful pages (balance sheet, income statement) is needed before extraction, and it cuts the bill too.
What we take from it
On this file, the half of the form that can come from documents fills itself, correctly, for a few tens of cents. The other half belongs to the client, and that is as it should be: those are the questions the adviser has to ask anyway.
What is left is mechanical work: ticking the form's checkboxes (text fields are written already, checkboxes not yet), adding the postcode table, and mapping each bank's and insurer's template. None of it is research.
One file is not a statistic. The next step is to repeat the test on varied files: employees, self-employed clients, couples, clients with several loans.
Are you a mortgage broker or a wealth manager? We are looking for a few firms to test this with us on their own files, anonymised, and on the forms they really fill. Get in touch.
Further reading: turning chat messages into business objects, and reviving a legacy ERP.