It reads the clause before it answers.
A chat scoped to one contract, which fetches the actual clause text and your actual playbook position with tools before it says anything. Not a summary of your contract stuffed into a prompt and hoped over, and you can see every fetch it made.
“Can we live with 8.1 as drafted?”
Choose one of the three quick actions and the thread carries on. Everything here is the real interface.
Short answer: not comfortably. 8.1 leaves indirect losses uncapped, which breaches L-04 (escalate condition: "liability uncapped in any form") and sits outside your £2m insurance aggregate.
If they genuinely won't move, two live routes:
1. Offer W3 (3× super-cap), a concession that still caps exposure. Meridian accepted a 3× cap with Halcyon-style carve-outs in your last two negotiations.
2. Trade against 5.2, accept their 60-day payment terms in exchange for the cap. Their commercial priority in round 1 was payment terms, not liability.
Signing as-is means unlimited exposure on a £48k engagement, that's an escalate-to-counsel decision, not an ops sign-off.
“Thanks for the revised draft. We can work with your payment terms at 60 days as they stand.
In return we do need clause 8.1 to carry a cap. Our position is twice the fees paid or payable in the twelve months preceding a claim, mutual on both sides, we’re not asking you to accept anything we aren’t accepting ourselves.
On that basis we’re ready to sign this week.”
This is your tier 2 fallback, so using it is within policy, but note it is your least-accepted wording. W1 at 2× lands 78% of the time and is worth one more attempt before you concede the multiple.
Three of your last four signed at 2× or better, so Meridian is the outlier rather than the market. Calloway is worth a look separately, you signed uncapped there, which is the precedent that would be quoted back at you.
It fetches. It doesn't recall.
The usual approach is to paste the whole contract into the prompt and let the model find its way. That works until the contract is long, at which point the model starts remembering things approximately. Ask Vern is given tools instead, and has to go and look.
Returns a clause verbatim with its character offsets, so a quotation in an answer is the contract’s words and not a paraphrase.
Finds every mention of a concept across the document, which is how it notices that liability is dealt with in three places, not one.
Pulls the position from the published version, including its fallback tiers, so the advice argues your case, not a general one.
Reads the run’s own findings, so the chat and the report never contradict each other in front of a client.
Four places.
One conversation.
Threads are stored against the matter, so a question asked in Word is still there in the dashboard tomorrow.
The default way it opens, a panel docked over whatever you were looking at, with the clause you had selected already attached as context. Expand it to the full tab when the answer turns into a piece of work.
The version above, room for a thread list, longer answers, drafted emails and comparison tables. Sits alongside Findings and Versions on the contract.
Highlight a paragraph in the document and ask about it. The selection becomes the context chip, so “is this normal?” has an unambiguous subject.
Connect Vern to the assistant your team already uses and ask it there. It calls the same tools against the same contract, so the answer is the one Vern would have given.
Mostly not “what does this mean”.
The review already explains the clauses. What people want next is help doing something about them, which is why the quick actions are drafting actions rather than explaining ones.
And because threads live on the matter, the answer you got in round one is still there when the same clause comes back in round three.
Ask it something you'd normally ask a lawyer.
Put your own question to an anonymised contract from your sector on the call. You'll see every clause it reads before it answers.