In a nutshell
The chat now occasionally speaks up on its own — with a concrete thought derived from what it already knows about your case, never on a fixed rhythm and never guessed, but only when it can genuinely be substantiated.
Why we deliberately keep this rare
An assistant that drops a “tip” into every other answer quickly starts to feel like a chattering sales assistant — and you soon learn to skim past such hints. We chose the opposite approach: the hint appears only when it genuinely follows from real, already-loaded data, is capped at two per document or conversation, and costs you no extra click — you simply read it or ignore it.
For consumers: a thought from your own case
If the chat has loaded a playbook that matches your topic — say, a termination notice, a rent increase, a purchase agreement — and recognizes at least one field that fits your specific conversation, it may remark on its own: “Based on X, Y could make sense — have you considered that?” This isn't the language model guessing: a server-side check verifies that the hint genuinely matches the loaded playbook fields before it ever appears. No match, no hint — the model never makes that call alone.
In the lawyer portal: two building blocks
In the lawyer portal's AI assistant, two things come together. First, the findings that the fact graph already gathers about your case file — contradictions, inferences, missing exhibit references — are now visible directly in the chat, instead of just waiting in the case overview. Second, the assistant can bring up a thought of its own during the conversation, derived solely from the case context already loaded — for example, something like “Based on the deadlines recorded so far, an objection (Widerspruch) rather than an appeal (Einspruch) would be the correct remedy.” This hint, too, is checked server-side against the case file's actual facts and relationships before it reaches the chat — if the assistant invents something that cannot be substantiated in the case file, no hint appears at all.
What “the AI thinks along” already means elsewhere in the system
The new chat hint isn't an isolated feature — it joins several existing building blocks of the same idea. After every analysis, Lexi already suggests the next sensible step — with visible provenance and the price shown before you click. The fact graph automatically detects contradictions between documents in the same case file. And the Copilot checklist in the lawyer portal keeps a case file's open next steps permanently on record instead of forgetting them once the analysis is done. The new hint marker now turns all of this into a conversation: instead of only answering a question, the chat actively brings such findings into the discussion whenever they fit the topic at hand.
Fail-closed, not creative
At its technical core, this feature is about restraint: the language model may propose a hint, but never alone decides whether it appears. A second, deterministic step checks every proposal against the data actually loaded — playbook fields for the consumer chat, facts and relationships for the fact graph. If a proposal fails that check, it disappears without comment. That's precisely what sets this apart from a generic chatbot: it wouldn't have a loaded case file to correct itself against.
Staying honest
A hint is food for thought, not a final legal assessment — whether it truly applies to your case is a judgment call that remains yours, or your lawyer's. More on how this works: AI transparency; why a tool with genuine knowledge of your case file differs from a generic chatbot: Why not ChatGPT.
