Why legal AI drafting has hit a ceiling, and what lifts it 

By David, co-founder, Avvoka 

I’ve sat in a lot of meetings with knowledge teams about one question. How do we get lawyers to start from the precedent instead of their last deal? 

Those teams work hard to codify judgement and record deals, pushing for rigour and repeatability. And still, under deadline, most lawyers open the last deal they worked on and start from there. 

Then generative AI arrived, and a first draft that used to take an afternoon came back in minutes. Adoption followed quickly, but with mixed results. When we ask firms why, the answer is usually some version of “it couldn’t learn our firm’s style.” 

The obvious fix was to give the tools the firm’s own documents, and over the last year the industry built the connections to do it. Almost every system a lawyer relies on has learned to talk to AI, most of it through one open standard, the Model Context Protocol, now backed by the major AI labs and the legal tech vendors firms already use. Now the assistant can search the DMS for you. But can it tell which precedent the firm actually stands behind? 

You can roughly divide the last decade of legal drafting into three eras: the template era, the assistant era and now the connected era. Each one solved the problem the last one left behind. And each one, so far, has stepped around the question of what lawyers start from. 

I think the connected era is where it can finally be answered. 

What search can’t tell you 

Plugging the DMS into your AI is real progress. A year ago it meant exports, bespoke integration work and nervous emails to the risk team. 

But think about what the DMS holds. It has the version the partner signed off, the three that came before it and nothing to say which is which. Research platforms hold the law. Neither of them is the firm’s standard. 

Every firm we work with has a standard. It’s in the precedents the knowledge team maintains, the fallback a practice group agreed after a hard negotiation, the clause a senior partner insists on because of a deal that went wrong years ago. It’s the firm’s judgement, written down. 

Search answers where is it? The standard answers which one do we use? Only the first is a search problem. Search can find every indemnity clause the firm has ever written, but it can’t tell you which one the firm stands behind. So when a lawyer asks for a first draft, the AI drafts from what it can reach, and what comes back is the firm’s last deal or the market’s average one, rarely the firm’s best. That’s where drafting hits a ceiling. 

The drafting ceiling 

Lawyers have taken to AI. More than two-thirds of legal professionals now use generative AI at work, more than double the year before, and 94% of firms are using or exploring it. 

The return is harder to find, and not just for firms. In-house, 83% of legal teams can’t tell whether their AI spending is paying off. Just 6% of firms describe the AI inside their DMS as enabled and widely used. The connections exist, but very few firms have their AI drafting from their own documents yet. 

When I read those together, I see firms have bought the tools and lawyers use them. What I don’t see is the time coming back. The AI produces a capable first draft, and then someone senior brings it back to the firm’s position, clause by clause. The time saved on drafting gets spent again on review. 

But here’s the part that bothers me most. That rework fixes the draft, but nothing gets fixed at the source. The partner’s judgement goes into one document and stops there. Tomorrow’s draft starts from the same place and needs the same corrections, so the firm pays for the same expertise again and again. 

That’s what I mean by the drafting ceiling. It’s the point where a better model stops producing a better draft, because it’s still drafting from whatever it can find, and that’s rarely the firm’s best work.  

Why agents make the source matter more 

Firms are not choosing one AI tool and stopping there. Already 16% of law firms use agentic AI, and that number will only grow. 

Agents change the stakes. They don’t wait to be asked. As the editors of a new global guide to AI law put it: “Agentic AI opens new horizons in terms of AI capability, but with that inevitably comes execution risk.” When an agent puts together a first draft with nobody watching, what it drafts from becomes a risk question as well as a quality one. 

If every firm is using the same models, the model can’t be what sets one firm apart from another. The judgement each firm owns has to do that. 

I don’t think the standard can live inside any one tool any more. It must sit underneath all of them. That means the firm’s approved wording and rules, recorded once and called by whichever tool or agent happens to be drafting, so the answer is predictable and repeatable. Without that, four tools give you four versions of roughly right, or your standard ends up tied to whichever vendor you picked. 

At Avvoka, we’ve started calling this headless drafting. The drafting logic has no single front end. It goes wherever the drafting happens. 

What it changes for the lawyer 

Most of the talk about agents describes the same shape. The lawyer sets the work up, agents carry it out and the lawyer reviews what comes back. What I think it leaves out is where the firm’s standard sits. 

Today the standard sits with the lawyer, so it gets supplied at both ends of every matter. At set-up, someone restates the firm’s position in the prompt or hunts for the right precedent to attach. At review, someone senior checks each clause against it and puts back whatever drifted. 

With headless drafting, the standard sits underneath every step. The knowledge team records it once. Set-up becomes about the matter: the client, the deal, the objectives. The tools and agents draft from the firm’s approved wording by default. Review becomes about judgement and strategy, and each correction improves the standard rather than disappearing into one document. 

Lawyers keep the pen throughout. The firm’s standard covers what the firm has already decided, and the lawyer makes the calls that need a lawyer. 

The old question, again 

Remember the question from all those meetings? It never went away. AI inherited it, and now answers it on every matter, at speed, usually with the last deal. I used to think it was a knowledge management problem. I now think it’s the most important question in legal AI. 

A lot of firms have spent the last few years choosing tools. The next few will be about what those tools know. So it’s worth asking your team one question. When our AI drafts, what is it drafting from? If the honest answer is whatever it can find, you’ve found your ceiling. 

Learn more about headless drafting: https://avvoka.com/headless-drafting