In a nutshell
The AI assistant now accepts whatever occurs to you while an answer is running — and at the same time shows you what it is actually working on right now.
The problem was never the AI, it was a locked input field
Nobody doing legal work with an AI assistant thinks in neat question-and-answer pairs. While the answer on limitation periods is being written, you remember that you also wanted to draft the reply to the statement of defence — and that there is a deadline in the file you should check. Until now that thought had literally nowhere to go: the field was locked while an answer was running. Two poor options remained — wait and risk losing the thought, or abort the running answer you had already paid for.
That is fixed. The field stays usable, and what you write is queued. As soon as the running answer is finished, your next question is asked automatically. You see the queue in a slim bar above the input field and in full in the work-status panel: every entry can be moved to the front, edited, saved for later or discarded.
What happens when something occurs to you mid-answer
Not every interjection is the same, and that is precisely the point. “Ask that next” is different from “no, do this instead”, and both are different from “by the way, include this too”. So as soon as you start typing during a running answer, several routes appear — and each one says what it actually does:
- Queue it is the default. Nothing is interrupted, nothing is lost.
- Addendum after the answer: the running answer finishes and is then extended with your addition.
- Recreate the answer with the addendum: the run stops and the same answer is produced again — this time with your addition in view from the start.
- Send now overtakes; discard & take back removes your last question together with the answer already begun from the conversation.
On the keyboard: Enter queues, Ctrl+Enter overtakes, Alt+Enter adds an addendum.
Why we do not claim that a running generation still picks up your addendum
That would be the more convenient story, but it would be false. A generation already under way cannot be fed after the fact by any provider — the prompt is irrevocably fixed with the first token. The large assistants you might have in mind do not do it either; they queue.
That is why the option closest to an addendum is called “Recreate the answer with the addendum” and not, say, “Add addendum”. It also states what it costs: the part already written is discarded, and the tokens started on it have been used. This is precisely why the default for addenda is the lossless variant — letting the running answer finish and extending it afterwards. Papering over a technical limit with friendlier wording is the wrong path: once someone notices that one promise does not hold, they stop believing the next one too.
One question at a time — and why that is deliberate
The queue works strictly serially. That is not technical convenience but a decision about your case record: two answers running at once would interleave their messages, so the log would end up holding two questions followed by two answers with no discernible pairing. For a tool whose conversation belongs to the file, that is unacceptable.
Just as important: the queue never carries on blindly. It visibly pauses when you stop the running answer — anything else would be a broken stop button. It pauses when a run fails, instead of sending the next questions into the same error. And it pauses when your quota is used up, instead of quietly consuming more. In all three cases you can see why it is standing still and decide for yourself whether to resume.
Anything you would rather save for later stays noted in this conversation first — and can be moved permanently into your personal list under “My work”, where it outlives both the conversation and the browser.
You can now see what the AI is working on
The second part of this change is less conspicuous and, to our mind, the more important one. Until now the assistant showed three bouncing dots while you waited — for up to 45 seconds. The reason was architectural: building the case context, reading the attachments and the entire tool round all happened before there was any connection to the display. The steps were only reported afterwards as already completed — a retrospective, not a display.
That order has been reversed. You now see live, with a duration per step, what is happening: assembling context, reading attached documents, searching the file, looking up a statute, obtaining a second opinion, writing the answer — and finally the citation check, which used to be a completely silent gap after the last word. Four tools that were never reported at all have become visible too: the legal library, case-law search, neighbourhood search in the fact graph, and re-reading an earlier message.
The step trail does not vanish with the answer. It stays attached to it and can be expanded again at any time.
What this has to do with trust
A wait without an explanation always feels like a risk. You do not know whether work is happening, whether something is stuck, whether it is getting expensive. That gap is exactly where the mistrust many colleagues feel towards AI tools grows — and it is justified mistrust as long as nothing is visible.
A visible work trail is therefore more than cosmetics. It answers, after the fact, the question of what an answer actually rests on: which documents were read, which provisions looked up, whether citations were checked. Together with the source references the platform attaches to every statement anyway, a text you have to believe becomes a result you can follow. That is the difference that matters in this profession.
