Legal AI and the Changing Role and Design of Work Product Review
Table of Contents
The AI Final-Version Gap
Review Methodology as a New Problem to be Solved
How We Used to Work: Drafting is Reviewing is Drafting
How We Work Now: Reviewing as a Separate, Subsequent Step to Drafting
The Need for Review-Process Design to Narrow AIFVG and Provide the Efficiencies that AI Has Promised to Lawyers
Conclusion: What We Know Lawyers Must Solve For in Legal AI
The AI Final-Version Gap
AI is compressing the time it takes to generate legal documents. If you know nothing else about AI’s impact on legal, you of course still know this. However, it doesn’t necessarily follow that AI - without more - inevitably compresses the time to generate final versions of legal documents (let’s call this disconnect the “AI Final-Version Gap” or “AIFVG”). But of course, doing so is the goal, because internal and external clients do not benefit from speed-to-first-draft. They benefit from speed-to-client-ready-version.
Review Methodology as a New Problem to be Solved
As AI improves, AIFVG will certainly compress, because the quality of outputs will become better and more capable of capturing nuance. But right now, and even as AIFVG gets narrower, human review of AI outputs and drafts remains key in high-stakes matters and practices. This is the “last 10 yards” problem in contexts where precision is key and judgement calls live in shades of gray, and this is where bottlenecks can emerge to try to swallow AI throughput gains.
Consequently, clever, human-centered design to optimize efficacy and speed of output review will be a differentiator between lawyers and legal teams. Put differently, while getting an AI output is important, knowing what to do with it - and having set up an arena that allows you to do so most effectively - is even more critical.
How We Used to Work: Drafting is Reviewing is Drafting
Think about it this way. Five years ago, you were an associate at a law firm, and you were just assigned the initial draft of a contract. You were given some key terms from your manager or partner, emails or notes from a client call, some background context and a precedent or template to begin with. And while time pressure was very real, it was at least theoretically understood that you could not literally be typing two documents at once. In other words, the expectation was for you to take those inputs and draft a document as quickly and accurately as possible, from start to finish, circling back for feedback if needed.
While you performed the task (at least after you’d achieved a certain level of experience), you were making judgement calls throughout - what standard of care to use here, whether to port the 10b5 rep verbatim from the negotiated precedent or dial it in to be more favorable to your side - and probably also reading through even the sections you didn’t anticipate needing to change, in order to ensure no changes you were making had unanticipated domino effects in those sections (let’s call this whole process “Manual Drafting”).
All that is to say: by the time you’d completed your draft, you knew what was in the document. If the partner, manager or client had questions about how you decided to draft something (or leave something untouched), you most likely had an answer. And that partner, manager or client also had at least some understanding that undertaking that Manual Drafting took a non-zero period of time.
How We Work Now: Reviewing as a Separate, Subsequent Step to Drafting
What does that same scenario look like now? The partner or manager knows that, with the AI tooling they’ve paid for you to have, you essentially can draft multiple documents at once. So the assignment is no longer a set of inputs for one document, which you’re expected to take back to your desk and work through for several hours in a relatively straight line.
Instead, it’s a series of inputs for, say, three documents that you’ll be responsible for landing, in roughly the same time period as the one doc took you five years ago, because the understanding is that you won’t be doing the Manual Drafting. You’ll be feeding the inputs, mixed with a bit of your commentary, into AI tools, and then you’ll let them spin until you have AI-generated drafts for each of the three docs. Then what?
“Then what?” is the $64 million dollar question. How do you engage with a newly drafted document, without the benefit of having really been the one that had to navigate and revise it, to ensure it accomplishes the business and legal goals you were aimed at? How do you do that for three such documents in the same time you used to be allotted to handle just the one? How do you manage the cognitive load when the job shifts toward solely making decisions around the thorniest or most technical issues and away from less-mentally-taxing keystroking?
The Need for Review-Process Design to Narrow AIFVG and Provide the Efficiencies that AI Has Promised to Lawyers
You accomplish these things with the help of rethought review processes and patterns, aimed toward at least a few things (let’s call them, collectively, the “Review Model”):
Prioritizing provisions/concepts of highest consequence
Accelerating understanding of potential cross-provision domino effects
Using visuals to draw the lawyer from key area to key area within a document
Easing the cognitive fatigue of context switching
Configuration of review materials in the lawyer’s field of view (i.e., what documents and source materials are visible, how are they navigated; in our example, how are they organized by/under each of the three new documents? In other words, how do you avoid being the lawyer with ten Word docs and emails scattered open across their desktop, constantly/maddeningly toggling and scrolling among them?)
Can the review be guided/semi-automated?
Creating a Review Model for specific use-cases/documents is how lawyers can begin to more quickly narrow AIFVG and actually get documents out the door (not just generate first drafts). Some of the issues within a given workflow’s Review Model may be addressed via products, others via tool configuration and still others via human coaching/teaching.
And more likely than not, the Review Model will vary from document to document and project to project (e.g., the set-up of the Review Model probably looks much different for drafting a placement agent agreement, compared with producing an underwriter’s diligence/factual-backup questionnaire for a draft Form S-1). But creating the Review Model, and then tooling/educating against each of its elements appropriately, will be a key unlock for lawyers using AI.
Conclusion: What We Know Lawyers Must Solve For in Legal AI
We know a few things:
AI makes parts of certain legal tasks faster
Clients, internal and external, expect those compressed tasks to amount to faster delivery
They also expect more things to be done simultaneously with the human lawyer as orchestrator and not drafter
This pushes more of the lawyer’s work to review of newly-formed drafts with less context than lawyers (who traditionally created the draft) historically had, in order to finalize a document
This means a much greater proportion of the lawyer’s time-spent is devoted to cognitively draining tasks (i.e., review of key items without being truly grounded in a document) done at speed and across multiple contexts
This makes lawyers’ roles potentially less satisfying and more burnout-inducing, and errors (or, at a minimum, homogenization of work product) more likely
To solve for all of this, we must design workflows and processes anew. This will mean changing some old ways, emphasizing others and dealing with periods of uncertainty and extreme discomfort during adoption. As the trope goes, you can’t buy your way into AI fluency and substantive adoption. Instead, you must design your way to it. Figuring out what to do with the artifacts that AI gives you, and not just figuring out how to get it to give you something, is a massive part of that design.
Ten Manor Advisory LLC is not affiliated with, controlled by or in control of any other business or entity.