In July we published a walkthrough of one of our papers: on routine tasks AI acts as an equalizer and flattens the gap between novices and experts; on complex tasks it acts as an amplifier, magnifying the judgment its user already has, including the lack of it. That piece ended by saying the framework shapes how we build enterprise AI. This one shows what it looks like in contract review.
Contract review is really two kinds of work
When a purchase agreement lands on an in-house lawyer's desk, the work splits roughly in two.
The first kind is checking. Is there a payment clause? Is liability for breach defined? Which court handles disputes? Is the confidentiality period stated? These questions have clear right answers. Go through a checklist and a junior lawyer and a senior one should reach the same result; the only differences are speed and whether something gets missed.
The second kind is judgment. Is the liquidated damages rate too high for us? Is the liability cap set at a sensible level? Will the half-sentence the other side changed leave us exposed at acceptance? There is no standard answer. They have to be weighed against the deal, the counterparty's bargaining position and how much risk the company can carry. Here the gap between a senior lawyer and a junior one is real, and it should be.
In the paper's terms, the first kind is a routine task and the second is a complex one. Covering both with the same AI workflow usually serves neither well.
Checking: let AI be the equalizer
For checking, the goal of AI is to let everyone do the work equally well and equally fast.
In Zico AI this is the quick check: it goes through a fixed checklist, item by item, to find which key provisions a contract is missing, returns a result in seconds, and marks anything it is unsure about as "to be checked" instead of forcing a conclusion. Because the checklist is fixed, results are comparable: the same contract gives the same table, whoever runs it.
The part worth stressing is "mark it when unsure". An equalizer is valuable because it levels reliably, not because it appears to know everything. A checklist that honestly shows its gaps is better suited to a less experienced reviewer than a fluent opinion that is occasionally wrong.
Judgment: let AI be the expert's lever
For judgment, the paper's mechanism is this: model output inevitably contains errors; experts can verify it, push back and spot which sentence is wrong, while novices tend to accept it as given. So the design goal here is not to have AI decide, but to help the expert verify faster.
In the product, that comes down to three things:
- Every comment points back to the text. Each flagged risk is anchored to its exact place in the contract, so the expert can judge at a glance whether it holds up, instead of reading a summary detached from the source.
- Nothing moves on until a person confirms it. AI review comments are candidates; they go into approval only after the lawyer confirms, edits or rejects them.
- Changes and conclusions are read separately. When the counterparty sends back a revised draft, version comparison lists substantive changes apart from wording changes and marks which changes alter the checking result, so the expert's attention goes to the few places that need weighing.
What the three have in common is that AI sits before the expert's judgment, not after it. It lays out the material and points to the questions; the decision stays with a person.
What this means for companies
When we explain this distinction to clients, one question comes up often: can a junior lawyer use AI to review contracts at all?
Our answer is yes, and they should start with checking. Once checking is handed to AI, a junior lawyer's output gets closer to a senior colleague's more quickly; judgment calls still need an experienced reviewer before they go out. The reverse is also true. If a team hands complex judgment straight to AI output and skips review, the failure the paper describes, errors carried unchanged into the final work, shows up in a contract as clauses, and a signed contract cannot be taken back.
Roll the same tool out to a whole legal team and some people get good results while others do worse. The difference is often not training but which side of the line their work falls on.
Where the framework stops
The paper is a position paper based on field observation, not a controlled experiment, and "task complexity" is not quantified. The same holds for contracts: which provisions count as checking and which as judgment differs by company and by contract type. That is why in Zico AI the review checklist is configured by each company. Our recommendation is that anything outside the checklist defaults to the side that needs human judgment.