- Client
- Commercial real estate law firm based in Illinois, about 30 attorneys and paralegals
- Function
- Legal operations and transactional practice
- Headline result
- First-pass redlines fell from 6 hours to 2, and the firm recovered about 1,400 hours a year.
The challenge
The firm's margin depended on how much senior time each deal consumed. Too much of that time went to work that was skilled but repetitive.
Three bottlenecks stood out:
- Redlining. Every counterparty draft needed a first pass against the firm's positions. Quality varied with who held the pen and how busy they were.
- Deal summaries. Paralegals built critical date sheets and lease abstracts by hand. A single missed date in a purchase agreement carries real liability.
- Version comparison. When one version existed only as a PDF, converting it and running a Word compare flagged about 2,000 formatting differences. The few substantive changes were buried among them.
Generic AI tools were not a realistic answer. Client documents could not leave a controlled environment, certain matter folders had to stay off limits, and partners would not accept output they could not verify.
Our approach
We started from the firm's own standards rather than from the technology. We codified its redlining conventions, playbook positions and date-calculation rules into written guidance. We then built one private workspace around that guidance, connected to the firm's practice-management system.
The rollout was scoped at six weeks and rested on four design choices:
- Encode the playbook, not just the prompt. Contract review extracts clauses, scores risk and proposes redlines against the firm's positions. It reads the deal from the client's side, because the same clause means opposite things to a buyer and a seller.
- Make human review structural. Every AI result is a draft. It moves from draft to in review to approved, and nothing can be exported or sent until an attorney approves it.
- Fit the output to the existing work product. Critical date sheets arrive as Word documents in the firm's template. Lease abstracts arrive as spreadsheets, one tab per lease, with issue flags for the reviewer.
- Build confidentiality into the data layer. Each firm's data is isolated. Restricted folders are excluded before ingestion, so they are never parsed, indexed or retrievable.
Every AI result stops at an attorney before it leaves the firm
The workspace also added cited chat over the firm's knowledge base. A PDF comparison tool shows only substantive changes, and a table-of-contents generator works even when Word styles are wrong. Attorneys file approved documents back to the matter with one click.
Impact
After launch, attorneys spent less time producing first drafts and more time on judgment calls.
| Measure | Before | After | Change |
|---|---|---|---|
| First-pass redline, per document | 6 hours | 2 hours | -67% |
| Lease abstract, per lease | 4 hours | 45 minutes | -81% |
| Critical date sheet, per deal | 2.5 hours | 30 minutes | -80% |
| Items to review in a PDF compare | About 2,000 | About 40 | -98% |
| AI drafts approved with minor edits only | Not applicable | 85% | New measure |
At current volumes, the firm recovers about 1,400 attorney and paralegal hours a year. That is worth about $490,000 of capacity at a blended rate of $350 an hour. Adoption reached 90% of fee earners within 60 days, and the review gate means no AI output can leave the firm unapproved.
What made it work
- Trust came from the gate. Attorneys adopted the tools because the system could not bypass them.
- Written standards were the real asset. Codifying the playbook improved consistency among junior staff even before AI was added.
- Configuration beat custom builds. Each workflow is a feature switched on per firm, so the next firm starts from configuration rather than code.