How we built and evaluated the law firm's drafting check

We measured the review burden without exposing deal content to us.

We studied one working week in their deal-drafting group before choosing the intervention. The firm needed to understand who could see client work, so we explained how the study would report categories of work without exposing deal content to us.

The study ran on their own systems and classified work as drafting, reviewing, redoing, research, and client communication. It included lawyers who used AI and lawyers who didn't. That gave us a view of the review burden across the group, which we documented in a memo and workflow maps.

Their partners approved the standards before the check used them.

Past deal files recorded prior decisions, but they didn't establish which negotiating positions the firm wanted lawyers to use in a new draft. We captured those positions from partner mark-ups and closed deals, wrote them down, and had their partners review and approve them.

We configured existing tools first, then built the missing drafting and review check around those standards. The check compared draft positions with the approved ones and pointed to the prior deal supporting each position. That gave partners a source to inspect when accepting, correcting, or rejecting a suggestion for the current deal.

The check followed their existing document permissions.

Source documents stayed in their document-management system. Their security team applied existing access rules before retrieval, so searches followed the permissions the firm already used.

The work study and the drafting check used information differently. The study reported categories of work without exposing deal content to us. The drafting check retrieved source material under their document-access rules so reviewers could inspect the basis for a suggested position.

We kept the completed test deal out of its own sources.

We evaluated the drafting check against closed deals, but restricted each test to earlier deals as sources. The system therefore had to make suggestions without drawing from the completed test deal's own file.

Their partners reviewed the suggested positions against work they already knew and accepted, corrected, or rejected them. We used the agreed first-draft and partner-review criteria to evaluate the work. Partners continued reviewing drafts after the system entered daily use.

We measured standard-position acceptance and active partner review.

The results come from our founder's account of two one-week studies of the same deal-drafting group. Both used the same instrument and work categories. Staffing and deal mix stayed stable, while the follow-up week had more concurrent matters.

• First-draft acceptance: 78% of first drafts passed partner review without overrides on standard firm negotiating positions, up from a baseline of ~15% to 20%. Partners could still make other changes required by a deal.
• Partner review: median active review time per primary deal document fell from 90 to 20 minutes. This measures the partner's review effort, not elapsed time to client delivery.

Both results describe the studied deal-drafting group.

The firm also reported removing $200K to $300K a year in recurring AI tool expense. The $200K+ headline uses the lower end of that reported range. This is the expense removed, rather than net profit after the costs of the new system.

We trained their team to extend the method, and they applied it in another practice.

The firm put the check into daily drafting and review, directed lawyers to the approved route, and retired the former route. We tracked rework, repeated questions, and the types of deal work brought to the system.

Because the firm's standards would need to cover new deal types, we trained their knowledge team and practice champions to run studies and add approved positions. Their own team subsequently applied the method in a second practice without us.