What a good AI-assisted first draft actually looks like

Ask a room of lawyers what they want from an AI drafting tool and most of them will describe the same thing: a draft so good they barely need to touch it. It’s an understandable wish, and it’s the wrong target. The draft that looks most finished is very often the most dangerous one to receive, because polish hides problems and invites a lighter review than the work deserves. When I was in practice, the best first drafts I got from trainees were rarely the slickest. They were the ones I could check quickly and trust where I’d checked. That’s the standard an AI-assisted draft should be held to, and once you hold it there, a good one starts to look quite different from a good-looking one.

The wrong benchmark is polish

A first draft has one job: to get the matter to a signed-off document faster and with less risk than starting from the precedent bank. Whether it reads beautifully on first pass is almost irrelevant to that job. What matters is the total time from draft to sign-off, and the number of problems that survive the review.

Polish works against both. A clean, confident, fully populated document lowers the reviewer’s guard. They read it like something already settled, and the clause that quietly departs from the firm’s position sails through because nothing about it looked unusual. I’d rather receive a slightly rougher draft that tells me where to look than a gleaming one that makes me find out for myself. The question to ask of any drafting tool is not “how good does this look?” It’s “how much of my attention does this let me spend in the right places?”

It starts from the right place

The single biggest determinant of draft quality isn’t the model. It’s the starting point. A draft built from the firm’s own precedent, its house positions and the agreed terms of this particular deal will beat a more fluent draft built from general knowledge of what a lease or a facility agreement usually says, every time.

This sounds obvious, and it’s where a lot of tools go wrong. A generic draft can be perfectly competent law and still be wrong for this client, because the firm has a settled view on rent review mechanics or a client has a standing instruction on guarantees. A good AI-assisted draft knows which precedent it started from and says so. If it can’t tell you what it was built on, you can’t tell what you’re reviewing against, and the review becomes a fresh drafting exercise in disguise.

It leaves the square brackets in

Every lawyer knows the convention. Square brackets mean this isn’t settled yet: a figure to be confirmed, a party not yet named, an option still open. A good junior uses them liberally, because a flagged gap is a two-second fix and an invented fact is a professional problem.

AI tools have a strong tendency to do the opposite. Faced with a missing completion date or an unconfirmed break clause, they fill the space with something plausible, and plausible is precisely the problem. The draft I want leaves gaps visible. Where the instructions don’t say, it brackets. Where two readings are possible, it offers both. A draft that is honest about what it doesn’t know is worth far more than one that has papered over every hole, because the holes are exactly where the review time should go.

It tells you what it changed

Experienced lawyers review against something. They compare the draft to the precedent, to the heads of terms, to the other side’s last turn. A good AI-assisted draft makes that comparison easy rather than leaving the reviewer to reconstruct it.

In practice that means a redline against the starting precedent, so departures are visible at a glance. It means short notes against the clauses where the tool has done something non-standard, explaining why: “amended to reflect agreed cap in heads of terms”, “alternative wording - heads of terms silent”. It means a reviewer can spend their first five minutes on the handful of clauses that actually moved, instead of re-reading forty pages of boilerplate to find them. That’s where the real time saving lives, and it has nothing to do with how quickly the words were generated.

It’s the right size

There’s a particular failure I see constantly: the over-drafted document. Asked for a side letter, the tool produces something with the architecture of a full agreement. Asked to add a clause, it adds three, each hedged with belt-and-braces drafting nobody asked for.

Length isn’t thoroughness. Every unnecessary clause is something the reviewer has to read, the other side has to negotiate and the client has to pay for. A good draft is proportionate to the transaction. It matches the register and length of the firm’s own documents for this kind of matter, and it resists the temptation to show off. If the output is consistently longer than what your best associate would have produced, that’s not a bonus. It’s a cost you’re paying in review hours.

How to judge one

So when a firm asks me how to evaluate an AI drafting tool, I suggest ignoring the first impression entirely. Take a real matter that’s already closed. Run the tool on the original instructions. Then hand the draft to a senior associate and time the journey to something they’d sign off, while counting the issues they catch and, more importantly, the ones they nearly missed.

Do that across a handful of matters and you’ll learn more than any demo can tell you. You’ll see whether the tool starts from your precedent, whether it brackets or invents, whether it shows its changes, and whether it drafts like your firm or like the internet.

The goal was never a draft that doesn’t need a lawyer. It’s a draft that respects the lawyer’s review: one that starts in the right place, admits what it doesn’t know, shows what it did and stays the right size. That kind of draft rarely looks impressive in a demo. It just quietly gets more matters across the line, with fewer surprises, and that’s the only kind of impressive that counts.