The first draft is not the valuable part of legal work, but it consumes a disproportionate amount of the time. A paralegal pulls a precedent, adapts it to the matter, checks the jurisdiction-specific provisions, and produces something an attorney then reads properly. The reading is the skilled part. The assembly is not, and it is most of the hours.
That is the gap iLeaf's LEXI drafting layer was built for, and the constraint that shaped it is simple: a draft an attorney cannot check quickly is worse than no draft at all.
The problem with generic generation
Asking a general-purpose model to draft a contract produces something that looks correct. It uses the right register, includes the expected sections, and reads fluently.
What it does not reliably do is apply the right law. Jurisdiction matters enormously — a provision that is standard in one state is unenforceable in another, mandatory disclosures differ, and the federal overlay depends on what kind of agreement it is. A fluent draft with the wrong governing framework is a trap, because the errors are invisible at the level of style and only surface on careful reading, which is exactly what the draft was supposed to save.
The second problem is unattributed text. A clause that appears in a draft with no indication of where it came from has to be verified from scratch. If verification takes as long as drafting, the tool has achieved nothing.
How the drafting layer works
Guided questions rather than a prompt. The attorney or paralegal selects a contract type from the template library and answers structured, jurisdiction-specific questions — parties, term, consideration, governing state, the provisions this matter actually needs. The system asks what it needs to know rather than inferring it from a paragraph of free text, which removes the largest source of silent error.
Jurisdiction applied from the first line. The draft is assembled against the law of the selected state, with the applicable federal framework layered in: UCC articles for goods, the Federal Arbitration Act for dispute provisions, FLSA and ADA for employment, CCPA and GDPR where data is involved, the Lanham Act for trade mark terms, the DMCA for digital content. Fifty states and twelve federal statutes are covered, so a firm working across state lines is not maintaining fifty precedent sets by hand.
Every clause cites its authority. This is the design decision that makes the output usable. Clauses carry the statute or precedent they rest on, inline, so review is a matter of checking the citation rather than reconstructing the reasoning. An assertion without a citation cannot be relied on in legal work, so producing one is wasted effort regardless of how accurate it happens to be.
Amendments from the original. Paste an existing contract, specify what changes, and the system maps every clause and generates a formal Amendment No. X with version tracking — rather than a fresh document that has to be diffed by hand against what it replaces.
Vetting on the way back in. When a base contract and its amendments are uploaded, clauses are compared side by side with the differences shown against the original, with redaction where the document contains material that should not be shared.
The boundaries, stated in the product
Three limits are enforced rather than advised.
It drafts; it does not advise. The output is a draft for attorney review, and the product says so on every document it produces. Whether a term is right for this client, in this negotiation, is a judgement that requires a licensed professional and knowledge the system does not have.
Human-in-the-loop is mandatory. Nothing reaches execution without a checkpoint. There is no configuration in which a generated document proceeds unreviewed.
No black-box output. Every clause is traceable to its source, and the full process is visible. A firm cannot stand behind a document whose provenance it cannot establish, and a system that cannot establish provenance is one the firm will eventually stop using.
What changed in practice
The platform's own measure is an eighty per cent reduction in drafting time against a manual paralegal workflow, with a complete first draft produced in about two minutes and three free drafts available before any commitment — which is deliberate, because the only convincing demonstration in this field is a document a lawyer can read.
The more interesting effect is on what the time gets spent on instead. Assembly was never where the expertise was. Moving those hours to review, negotiation and the questions that actually require judgement is the point of the exercise, and it is a better argument for the tool than any speed figure.
Have a system that needs to do this?
500+ systems shipped since 2011, and we still maintain most of them. Tell us what you are trying to move.
