What has changed since 10 May
On 10 May we explained why Smart Legal Pro replaces personal-data patterns with placeholders before the LLM call. Since the end of May, the masking has become considerably more robust — without you having to change anything in the interface.
Stage 2 at a glance
Identity seeds from your account. If you have stored letterhead details (account_identities), your name, company, address and billing email are pre-masked in addition to regex patterns. The AI sees, for example, __PII_PERSON_1__ instead of "Müller GmbH" — even when the name appears in the text without a typical address format.
Session consistency across multiple AI calls. Analysis → clarification → letter generation: within a single process, placeholders remain stable (__PII_PERSON_1__ is always the same person). This prevents mix-ups in long dialogues — especially in the Schreib-Assistent and during document analysis.
Persistence in the writing intake. writing_intake_sessions.pii_session stores the reversible map in the EU database, so that follow-up rounds and letter generation use the same masking.
Where it is live
The extended pipeline runs in 24 edge functions — including document analysis, contract comparison, legal questions, lawyer pre-chat, writing intake, letter fields and platform help. Exceptions (e.g. when the sender has to appear in the output) are documented in Privacy Policy §4c.
To complement the technical side: Security overview · AI transparency
What you will notice in practice
- Fewer "leaks" with repeated names in a long chat.
- Letters with your real letterhead — even though the AI only ever saw placeholders.
- The same data-protection logic in the public quick check and in the case file.
Limits (still important)
PII masking does not replace third-party confidentiality obligations or a legal data-protection review before uploading. Details remain in the legal-services notice and on the trial page.
Addendum of 8 September 2026 — two statements above are out of date
We do not rewrite a dated article retroactively. Two statements from back then are no longer accurate today, and both concern data protection — hence this addendum rather than a silent correction.
Session persistence in the writing intake no longer exists — and that is an improvement. The column that permanently stored the mapping table between placeholder and clear value was removed without replacement in September 2026. It violated our own first invariant: such a table is a re-identification key and must never sit in a database. Since then the writing intake has worked with a fresh mapping for each round that lives exclusively in memory — with no loss of functionality.
What was 24 edge functions has become considerably more. That figure described the state of affairs in May 2026. Rather than keep updating it, we no longer state a number of functions at all, but the thing that matters: which areas work masked and which expressly do not — set out in Privacy Policy §4c and, for the lawyer portal, in §4g.
And one figure that did not exist back then: since September 2026 the detection quality has been measured rather than estimated — against our own test corpus of 246 documents, around a third of which consists deliberately of hard cases. What came out of it, including the classes our rules do not detect, is set out in Privacy Policy §4c.
