LegalTrack AI: Contract Lifecycle Management at Scale

A corporate legal function does not usually fail at drafting. It fails at scale — at the point where the number of live agreements exceeds what anyone can hold in their head, and the things that go wrong are not bad clauses but missed dates.

LegalTrack AI came to us running more than a thousand enterprise contracts: vendor master services agreements, NDAs, and the amendments that had accumulated on top of them over several years. The drafting was fine. The problem was everything that happens after signature.

What was actually going wrong

Three failures, all administrative rather than legal.

Renewals were tracked outside the contracts. Key dates lived in a spreadsheet maintained by whoever had last been asked to maintain it. A contract auto-renewed on terms the business no longer wanted because the notice window closed while nobody was looking at that row.

Nobody could answer portfolio questions. "How many of our vendor agreements have uncapped indemnity" is a reasonable question from a risk committee. Answering it meant a person opening contracts one at a time.

The operative terms were not in the operative document. An MSA signed in one year, amended in another, with a side letter carving out a business unit — the notice period actually in force was not written in any single document, and reconstructing it took a lawyer an afternoon.

That last one is the difficult problem, and it is what makes naive contract search dangerous rather than merely unhelpful.

What we built

The deployment runs on iLeaf's CLM platform, with LEXI as the AI layer, covering the full lifecycle: draft, review, negotiate, execute, monitor, renew.

Extraction with the chain intact. Every agreement is ingested and its key terms pulled out — execution, effective and termination dates, payment terms, renewal notice windows, auto-renewal provisions, indemnity limits and liability caps. Critically, amendments are modelled as related documents rather than as separate records, so the platform resolves what is currently in force by following the chain rather than by matching text.

This is the part that makes the difference. Passage-matching retrieval finds the master agreement, because it is the longest and most contract-shaped document and uses the relevant phrases in a clause explicitly about them. The amendment that replaced the clause says something terse like "Clause 14.2 is deleted and replaced with the following", and scores lower on almost any similarity measure. A system built that way returns the superseded term, cites a real clause from a real executed contract, and gives no indication that anything is missing.

Playbook-driven vetting. When a counterparty returns redlines, every deviation is compared against the organisation's approved playbook and flagged with a suggested compliant fallback. Reviewers see what changed against what was acceptable, rather than reading the whole document again.

Proactive alerts on real dates. Renewal windows, payment deadlines and milestone triggers are pushed before they matter, derived from the extracted terms rather than from a parallel spreadsheet.

A full audit trail. Every edit, approval, signature and alert is logged with timestamp and identity. For a team that periodically has to demonstrate what happened and when, this turned out to matter more than any single feature.

The design decisions that mattered

Human-in-the-loop is mandatory, not optional. Nothing is finalised without a checkpoint. The platform does not advise; it extracts, cites and flags, and every output points at the exact clause and statute it came from. What a clause means for a negotiating position remains a lawyer's judgement, and the system is built not to offer one — a model asked what a clause means will answer confidently, because that is what models do, so being built not to is a decision that has to be made and enforced.

Recall is preferred over fluency. Where the amendment chain is ambiguous, the platform says so and shows what it found rather than picking the most plausible branch. A missed uncapped indemnity costs more than a hundred over-cautious flags.

Every answer carries its clause. Not a paraphrase — the text, from the document, with the reference. An assertion without a citation cannot be used in legal work, so producing one is wasted effort however accurate it happens to be.

Where it sits now

The platform runs end-to-end CLM across the portfolio: authoring from a template and clause library, approval routing, digital signature, central repository, obligation tracking, and analytics over the whole estate. Portfolio questions that previously required opening contracts individually are now queries.

The measure the team uses is not drafting speed, although that improved. It is that no renewal has passed unnoticed since the platform took over the dates — which was the failure that started the conversation.