In one sentence
When our assistant or an agent researches the web, it really reads the page it found — and before a query names a specific person, it asks you first.
Two questions, two answers
Not every question to the AI needs the same caution. "How is the limitation period for a contract for work calculated?" is an anonymous technical question — it names no person, no company, no identifiable target. Such questions are researched immediately, without any follow-up.
"What can be found online about Müller GmbH?" is something else: a query that names a specific target. This is exactly where the real change starts — not in the technology of the search itself, but in who decides whether it happens.
Read first, judge second
A search engine returns hits. Whether a hit is any good only shows once someone actually opens the page and reads it. That is now exactly what the portal does, consistently, staged by cost: first it checks its own legal source index — free, fast, but limited to what is already recorded there. If that is not enough, it queries a web search index focused on official legal sources; it returns verified references but does not write an answer itself. Only if that still is not enough does the more comprehensive web research with its own summary follow. Every source found passes through the same citation gate as any other research in the portal: only a reference that shows the claimed quote word for word counts as confirmed.
The result shows you how many web sources were used — with the note "Powered by Brave Search" whenever the middle stage contributed.
The approval nobody should automate
Back to the named query: instead of simply searching, the portal pauses right here and offers three ways forward, directly in the chat:
- Once — for exactly this one query.
- Remember for this case file — every future query about this name in this case file runs without asking again.
- Never ask again — the approval applies across all of that lawyer's case files.
Every approval granted is logged, even the one-off kind — and can be withdrawn at any time.
The reason we do not simply mask here, as the portal does in many other places: a search engine searches for whatever you give it. Replace the name you are searching for with a placeholder, and the search no longer finds anything useful — masking would destroy the function itself instead of protecting it. So a deliberate decision takes its place: not concealing who is being searched for, but deciding whether a search may happen at all.
A firm-wide switch, in addition to the individual approval
For the firm as a whole there is a separate, independent switch — off, anonymous questions only, or named searches allowed (with individual approval in the specific case). It sits in the same place as the switch for masking personal data, but it is its own, independent state with its own meaning — more on that and the full approval logic in the lawyer portal.
What already existed and now belongs with it
The assistant could already fetch a specific web address and read its actual content — including protection against unsafe targets. That capability was the first building block of the same idea: an AI that claims something should have checked it at the source instead of guessing. The new web search closes the gap before that — the step of first finding the right source. Together, both carry the same stance as our co-pilot for law firms: control is not a settings menu for us, it is a product promise.
Honest limits
As with any web search, query data stays with the search provider for a limited time before it is deleted — details are in our legal documents, not in this announcement. Your queries do not feed into training any model. And an approval permits the search — it says nothing about whether the results found are correct. The professional judgment stays, as always, with you.
Not on board yet? For lawyers shows how to get started; what else the AI sees of a case file is explained by AI case file coverage.
