Legal teams at PE-backed companies drown in the routine — NDAs, vendor paper, diligence requests — while the judgment work waits. Our skills clear the routine; the judgment stays with your counsel.
The Principle
Every output traces back to a source document — no citation, no output.
Maker-checker
The same control structure banks use for wire approvals.
Confidence gating
Calibrated to within ±5%
When the system says 90%, it’s right about nine times in ten — verified, continuously.
The Skill Catalog
Named, callable workflows — a button your legal team presses, not a blank prompt box.
Compares inbound paper against your template, flags deviations, and proposes redlines matching the language your team used last time.
Reviewed, not drafted
Accept, push back, or escalate — decided per clause, with a response email in the deal owner’s voice.
Per-clause decisions
Liability caps, IP assignment, termination triggers, audit rights — pulled from every agreement into one comparison table.
One table, not forty PDFs
A data-room dump routed into workstreams, with per-document summaries and a gap list against the checklist.
Data room → workstreams
Each new patent becomes a marketing one-pager and an engineering claim summary; inbound claims get a defensive read.
Two audiences, one source
Security attestations, privacy posture, certifications — drafted from source-of-truth documents with a citation on every claim.
Every claim cited
Worked Example
Compare, flag, propose, escalate — the head of legal opens a finished review, not a blank document.
Inbound NDA received
Counterparty paper, 24 clauses
Clause-by-clause comparison against standard template
21 of 24 clauses match corporate position
Non-compete: 3 years vs. our standard 18 months
Redline proposed — matches the language your team used the last six times
Governing law: matches corporate position
No action required
IP assignment: broader than standard
Escalated — flagged for the head of legal with the deviation cited
Proposed redline + response email drafted
In the deal owner’s voice, ready for review
The flywheel
Accepted, corrected, or escalated — every outcome feeds the eval set.
The 90-Day Legal Playbook
Codify what your team actually does, prove it on live paper, widen week by week — no pilot purgatory.
Weeks 1–2
Your positions extracted from the last 50 executed agreements — what you actually accepted, clause by clause. The agents learn your playbook, not a generic one.
Weeks 3–4
NDA and DPA triage in production, with the head of legal reviewing every run.
Weeks 5–8
Playbook redlining and vendor comparison go live. Deviation patterns show which clauses cost the most negotiation time.
Weeks 9–12
The M&A classifier and compliance memo drafter go live. Privacy assessments complete; telemetry reviewed with the GC.
Week 12
The next skills are scoped, and the eval set becomes a permanent asset every future skill inherits.
The Compliance Workstream
The workstream that keeps every deployment defensible — in any function.
United States
State privacy laws — CCPA/CPRA and their successors.
Canada
PIPEDA federally, plus Quebec’s Law 25 — stricter than both.
Documented before launch — not after the first incident.
People are told when AI affects a decision about them.
Access and deletion reach AI logs; retention set on day one.
A named owner tracks new AI law as it lands.
Mapped before a single skill goes live — nothing built twice, nothing built wrong.
Privacy Incidents, Automated
The same pattern that runs your contract queue applies to your incident queue.
A breach lands and the questions cascade: which regimes apply, what crosses the notification threshold, what the clock is. For a privacy-technology company we built the system that answers them — regulation in, structured decision out: a multi-factor risk assessment producing a defensible determination instead of a scramble.
IP & Liability
Six questions with written answers before the first AI-touched contract ships.
IP ownership — Settled under vendor terms before the first contract ships.
Copyright exposure — Evaluated for anything externally published.
Vendor indemnification — Read in full — what’s covered, what’s quietly excluded.
Open-model licenses — Confirmed for commercial use. “Free to download” isn’t a position.
Customer contracts — Updated wherever AI is part of service delivery.
Marketing claims — “AI-powered” vetted for truth-in-advertising exposure first.
One coordinated workstream with privacy — not two silos.
Guardrails by Design
Two rules, enforced in the design — not in a policy PDF.
Wherever the decision is consequential, a person signs. Hiring, credit, legal positions, safety — no confidence score overrides that, ever.
Consequence decides, not convenience
Legal skills read the contract repository and nothing else. No payroll, no deal pipeline, no board folder.
Least privilege, by default
Specialist Knowledge at Scale
Jurisdictional expertise doesn’t scale by hiring — agents carry the whole map at once.
For a collections platform serving banks and credit unions, we shipped specialist agents carrying bankruptcy handling and repossession law across all 55 US jurisdictions — tuned to each institution’s compliance requirements and tone with zero code changes, and with instant human hand-off the moment a situation changes. The specialist knowledge lives in the system, not in one person’s calendar.
Which regime applies and what changed last quarter — resolved per matter.
Notice periods, non-competes, termination rules — per state, per offer letter.
Permissible claims checked before the campaign ships, not after the complaint.
Every jurisdiction on every matter, with a human the instant judgment is needed.
Case Proof
A compliance software company needed AI to qualify community-development activities under federal banking regulation — where a wrong answer is an exam finding, not an inconvenience. We shipped a citation-backed decision engine: every determination linked to its source, every confidence score calibrated, five validation layers aligned to bank model-risk guidance (OCC SR 11-7, NIST AI RMF).
The GC’s Dashboard
Every legal skill reports into one telemetry layer — your numbers, not a vendor scorecard.
By paper type — NDA, DPA, vendor, customer. The number the business feels first.
Which clauses actually cost negotiation hours — where your template should move.
Falling as the playbook codifies; a rising line means it’s out of date.
On every prompt or model change. Nothing ships on a hunch.
Every review, every override, every version. The answer is a record, not a recollection.
Wave Two
Each second-wave skill inherits the positions, the guardrails, and the telemetry.
Renewals, exclusivity windows, price-rise clauses — tracked so nothing lapses quietly.
Answers with handbook citations; regulated topics route to a human.
Issued, acknowledged, and tracked — with the paper trail already assembled.
Drafted from the meeting record in your house style.
COIs collected, checked, and chased before they expire.
Screening results condensed into a decision-ready summary with sources.
The Bottom Line
The Foundation
Built into every band from day one — not bolted on after.
The Rest of the Spectrum
Ready to move
The fastest proof in the legal function: two weeks from your standard template to triaged inbound paper. We’ll show you the deviation report on your own contracts.
Talk to LightCI