Legal AI
Built with Prodular UIOne clause in 52 pages. Found in four minutes.
Brace, an AI agent for in-house legal teams. It reads the customer's paper against the playbook, redlines it in language legal already approved, and cites a clause for every change. A person makes every call.
- Scope
- Flow, product, how the AI behaves
- Surfaces
- Legal web, CFO phone, sales chat
- Hard part
- Every claim traced to a clause
- Deliverable
- Coded product on Prodular UI, recorded flows
00Overview
A quarter-end deal, on the customer's paper.
Fernway sells scheduling software to retail chains. Its largest deal of the year, $1.44M over three years, arrived on a Friday night as Brightwater's own contract. The quarter closed on Wednesday, and legal is one lawyer.
Pages of MSA and exhibits, read against 46 playbook positions
Clauses Brace sorted into match, redline, decide and blocker
Minutes from opening the file to the first finding
Blockers outside walk-away, each sent to the person who owns it
01The finding
The MSA caps liability. Exhibit B takes the cap away.
Section 12.2 limits a claim to a year of fees. Twenty-four pages later, Exhibit B says it prevails and that a data breach has no limit at all. Brace ties the two together in the margin and shows what the playbook allows instead.
One bracket across 24 pages: the clause that set the cap aside, found before anyone negotiated
Highlighted where the contract says it, not summarised
Every sentence cites a clause; the redline cites the playbook
02The week
Friday's paper, signed on Wednesday.
Every number matches across screens: 214 clauses, 3 blockers, 42 changes in the redline, 39 accepted, a 4× cap of $1,920,000.
03Three people
Legal decides. Finance signs off. Sales gets an answer.
Clara works in the contract. Rachel approves from her phone with the clause in front of her. Ben asks in plain words and gets the clause, the playbook and who decides.
Rachel, CFO: two approvals, 1 h 32 min and 40 min after they were asked
Ben, sales: “What does 4× actually mean if we get breached?”
04The impact
One quarter with Brace.
Fernway's legal team in the quarter before Brace and the first quarter with it. Same people, more contracts, the deals in the quarter they were due.
Customer paper to first redline, median
Quarter-end deals signed in the quarter they were due
CFO approval on an exception, median
Outside counsel spend in the quarter
Contracts the one lawyer carries in a week, up from 9
Exhibits that overrode an MSA cap, found before signature in 38 contracts
Lines of new clause text in 412 redlines. Every one came from the playbook
From the product’s scenario: Fernway, Q2 2026 before Brace and Q3 2026 with it.
05Decisions
Three calls made before a screen was drawn.
Each one has a price, and the price is written down.
- 01
Brace proposes, a person commits
Nothing reaches the customer until legal sends it, and every exception goes to the person who owns it. The cost: Brace can be four minutes from a redline that still waits an hour for a CFO.
- 02
Only the playbook's language, never new clause text
Every redline is a fallback someone in legal already approved. The cost: when the playbook has no answer, Brace says so and stops, instead of drafting one.
- 03
Every claim cites a clause, and says when it cannot
An answer without a source is not shown, and a missing document is a question, not a guess. The cost: more questions to finance on the first contracts.
06What we test next
Three assumptions that need evidence.
With in-house lawyers, finance leads and account executives, on real contracts with names removed, with tasks rather than opinions.
- Assumption 1
The bracket
Does a lawyer trust a cross-reference she did not find herself, or reread all 24 pages anyway?
- Assumption 2
The phone
Will a CFO approve a $1.9M cap from a phone, with the clause on screen, or wait for her desk?
- Assumption 3
The chat
Does sales stop asking legal once the answer comes with the clause and the person who decides?
Where does your AI feature need to show its sources?
We design the moment a person has to trust what the model found, and hand it back designed in five days.