CASE STUDY · LEGALTECH AI SOLUTIONS
An AI contract review assistant that cut review time by 35%.
- 35%First-pass review down from about 3 hours to under 2 hours (35% less) on a sample of 200 contracts
01 · PROJECT OVERVIEW
The client and the brief.
An anonymised in-house legal team. Name withheld under confidentiality.
A clause playbook, retrieval-assisted review and a lawyer-controlled interface with citations. The playbook set out, for each clause type, what was acceptable, what needed a fallback and what had to go to a senior lawyer. The assistant read each contract, matched clauses to the playbook and proposed a note. Each note pointed to the exact text it was based on.
- 1Each contract needed about 3 hours of first-pass review.
- 2The team handled a steady flow of vendor and customer agreements, and the same issues came up again and again.
- 3Senior lawyers were reading boilerplate to find the few clauses that needed judgment.
- 4The team wanted to cut that time without handing decisions to software.
02 · WHAT WENT WRONG
What went wrong, and how we fixed it.
Every engagement hits problems. These were the ones that mattered here, and what we changed so they did not come back.
- PROBLEM 01In the first pilot week, the assistant flagged many indemnity and liability clauses as risky when they were within the team's accepted range
The playbook described limits in words that the model read too strictly. Lawyers lost patience quickly, which is a real risk with any review tool. We sat with them, rewrote the rules as clear thresholds with examples of acceptable and unacceptable wording, and added a test set for these clauses.
- PROBLEM 02The false flags dropped, and the lawyers started to trust the other suggestions again
- PROBLEM 03A second hurdle came at the end of the pilot
The assistant handled English contracts well, but a few agreements had been scanned with handwritten amendments in the margins. The model ignored the handwriting, so a suggestion could cite a clause that had been changed by hand. We added a check that flags pages with low text quality or marginal marks and tells the lawyer to read that page directly.
- PROBLEM 04It was a small change, and it removed a failure that would have been hard to forgive
03 · THE PROCESS
How we ran it.
- 01Weeks 1 to 2
Workshops with three lawyers to turn their habits into a written playbook. We collected a sample set of past contracts and the reviews that had been done on them.
- 02Weeks 3 to 6
Build in weekly cycles. Each Friday the lawyers reviewed a batch and marked every suggestion as right, partly right or wrong.
- 03Weeks 7 to 9
Pilot on live work with a small group, daily feedback in a shared channel, and a weekly review of misses.
- 04After launch
A fortnightly check of the evaluation set and a short change log for any playbook edits.
04 · RESULTS
What changed.
Measured against the way the work ran before the engagement began.
- NOW 01
Lawyers start from the assistant's marked-up first pass and spend their time on the clauses that need judgment.
- NOW 02
The playbook is now a living document that the team updates when a negotiation teaches them something, and each change is tested against the evaluation set before it goes live.
- NOW 03
New joiners learn the team's standards from the same playbook.
- NOW 04
The team also tracks where the assistant is unsure, because those clauses usually show where the standards are unclear.
- 35%
First-pass review down from about 3 hours to under 2 hours (35% less) on a sample of 200 contracts
Lawyers accepted or edited every suggestion
05 · WHAT WE BUILT
How it works.
- Contracts were split into clauses and paragraphs, with page references kept.
- The review step retrieved the relevant playbook rule for each clause, then asked the model to compare the clause against the rule and explain the gap in a sentence.
- The output was structured, so the interface could show the clause on one side and the suggestion on the other.
- Controls were part of the design.
- Every suggestion stored its source text, the playbook rule used and the lawyer's action.
- Documents were processed in a private environment.
- Controls after launch
- The team kept three habits. A named lawyer owns the playbook and approves each change. Each month a sample of reviewed contracts is scored again against the evaluation set, so a model or prompt change cannot quietly lower quality. A short report lists the most frequently rejected suggestions, which shows where the playbook needs work.
- Handover
- We ran sessions for the legal team and their IT contact, covering how suggestions are produced, how to edit and test playbook rules, and hosting, logs and access. The team received a written guide and the evaluation set, so they could carry on without us.
06 · QUICK ANSWERS
The short version.
- What problem did the client have?
- Each contract needed about 3 hours of first-pass review. The team handled a steady flow of vendor and customer agreements, and the same issues came up again and again. Senior lawyers were reading boilerplate to find the few clauses that needed judgment. The team wanted to cut that time without handing decisions to software.
- What did Brandsmashers build?
- A clause playbook, retrieval-assisted review and a lawyer-controlled interface with citations. The playbook set out, for each clause type, what was acceptable, what needed a fallback and what had to go to a senior lawyer. The assistant read each contract, matched clauses to the playbook and proposed a note. Each note pointed to the exact text it was based on.
- What were the results?
- First-pass review down from about 3 hours to under 2 hours (35% less) on a sample of 200 contracts; lawyers accepted or edited every suggestion.
YOUR TURN
Have a similar requirement?
Tell us what is stuck. We will come back with a plan and matched, pre-vetted engineers.
