Know-Your-Client: from a name to a signed dossier

Lawyers, fiduciaries and compliance teams have to know who they take on. Our own product turns a name into a due-diligence dossier where every claim links to its source, and the decision stays with a person.

Know-Your-Client is our own product. It is for lawyers, fiduciaries and compliance teams, who have to check a person before they take them on: sanctions, politically exposed roles, registers, leaks, press.

The analyst types a name. The product comes back with a dossier in which every statement links to the page it came from, every source says whether it answered, and the risk level is a suggestion until a person signs it.

Product
Know-Your-ClientMont3's own, for AML and KYC due diligence.
Live
Since June 2026At knowyourclient.ch, in English, German, French and Italian.
Editions
Cloud or localHosted for you, or installed on the firm's own machines.
A Know-Your-Client report on Elon Musk, Overview tab: his photo, name and roles with citations, tabs for Profile, Career, Screening, Review and Sources, a suggested indicator of High, not signed, with its reason, five questions with their answers, and a summary and key facts with numbered citations
A real check, run on Elon Musk for this case study. Every statement carries its sources, and the High indicator stays a suggestion until an analyst signs.

What made it hard

  • Every statement needs a source. A compliance officer cannot act on what a model believes. A statement without a source had to be dropped.
  • Clear has to mean checked. A source that did not answer is not a clean result. The dossier had to say which sources were never screened.
  • Same name, different people. The first job is to work out who the subject is, before reading anything about them.
  • The decision belongs to a person. The analyst confirms or dismisses each finding, signs, and the record of that has to hold.
  • Some firms cannot send a name anywhere. For them the product had to run on their own machines, with their own models.

What shipped

  • Discovery. A name becomes a short list of candidate identities, each with how sure the match is.
  • A deep-dive. The person's profile, digital footprint and a search of adverse media that cannot be skipped.
  • Screening. Up to 19 sources at once: sanctions and PEP lists, INTERPOL, FINMA, Swiss and UK registers, the SEC, leaked-data archives, world news.
  • Twelve compliance rules. Plain code that suggests a risk level from what was found.
  • Sign-off. A report in seven tabs, a PDF, and an audit log in which every entry seals the one before it.
  • Monitoring. Signed-off subjects are checked again, weekly by default, with an alert only when something changed.
The How it works section: a line with three stations, enter the subject, the agent screens every source, review and sign off, and below it the list of up to nineteen authoritative sources
Three steps: enter the subject, screen every source, review and sign off.

How a check is assembled

  • Who. The analyst types a name, with a company and a city if they know them. An AI agent proposes the people it could be. If nobody matches anywhere, the check stops there.
  • What. One sub-agent per candidate reads the web, and every sentence it writes must cite a link.
  • Where. At the same time, the official sources are queried directly, all together. A source that does not answer reads "not screened", never "clear".
  • Checked by code. Sentences without a source are removed. Each fact is labelled as self-reported, corroborated or official.
  • Decided by a person. The rules suggest a level: a sanctions hit means elevated, a leak or adverse press means review, an unclear identity means indeterminate. The analyst decides each finding, signs, and the case locks.

How we built it

  • June 2026. First version, and in production on knowyourclient.ch within the month.
  • July. Version 1.0: a local edition a firm can install and license.
  • September. The hosted edition, open to sign up on your own.
  • October. Production moved to a Swiss host.
  • Every release. A labelled set of test subjects runs before shipping, and a single wrong answer stops the release.

Result

  • A dossier in which nothing is unsourced: every claim links to where it was found.
  • Up to 19 sources and more than 250 sanctions lists, screened in one check.
  • The level is suggested by rules a person can read, and decided by a person who signs.
  • Two ways to run it: hosted on a Swiss server, where the AI model runs at its provider, or installed on the firm's own machines with local models, where nothing leaves.

Stack

  • Server. Node.js 22, a single service with no third-party packages.
  • Storage. Encrypted files, one key per analyst. No database.
  • AI. Claude with web search, or any compatible model, local ones included.
  • Interface. Plain HTML and JavaScript, in four languages.
  • Hosting. A Swiss server for the cloud edition; the firm's own machines for the local one.

Article 035 · 13 November 2026 · 4 min read.
Reply to paolo@mont3.ch.