Automating the drafting of legal documents and pleadings: what a law firm can truly delegate to AI
Partager
A legal document is not a monolithic block.
Some parts can be delegated. Others never.
Let's break down a set of legal submissions section by section, to see where AI truly saves time.
The question "can legal drafting be automated?" is poorly posed. It calls for a binary answer, whereas the reality is graduated.
A legal document consists of very different parts. Some are mechanical and repetitive. Others involve your reasoning and responsibility. Treating them the same way would be a mistake in both directions: wasting time on the former, or delegating the latter.
The distinction that governs everything
Rephrasing
You provide the content. The model formats it, structures it, expands it, or condenses it.
The content comes from you. The model's added value is purely editorial.
Example: you dictate the chronology of facts in telegraphic style, the model produces a drafted and ordered account.
Generating
You ask a question. The model produces content it draws from its training.
This is where errors arise, because the model produces what is plausible, not what is accurate.
Example: asking "what case law supports this argument?" without providing it with a documentation base.
Anatomy of a set of legal submissions
Here are the usual sections of a legal document, with the realistic level of delegation for each.
The four real risks, and how to address them
1. Inventing references
This is the most documented and serious risk. When queried without a documentary base, a model produces references to decisions that resemble real references: credible format, plausible jurisdiction, plausible number. They do not exist.
The solution: never ask the model to provide case law from memory. Either you provide it with the decisions, or you look them up in your usual databases. Any reference produced by a model must be verified at the source, without exception.
2. Confidentiality of documents
Drafting with AI implies giving it access to case file elements. With an online tool, this means transmitting documents covered by professional secrecy to an external provider.
The solution: a local AI, installed on a firm's machine. No outbound connection is necessary for inference, so no documents leave your premises.
3. Standardization of style
A model used without proper guidance produces recognizable prose, often bland and vaguely Anglo-Saxon in its construction. In a legal document, this is noticeable.
The solution: feed the model with your own validated documents. It then adopts the firm's formalism and vocabulary rather than a generic style.
4. Dilution of responsibility
The most insidious risk is not technical. It is the habit of quickly validating a text that "looks good," because it is already drafted and reads well.
The solution: an explicit firm rule. No document leaves without a complete review by the signing lawyer. The comfort of a first draft never reduces the requirement for review.
How much time can you really save?
Adjust the values to your firm's reality. The calculation only covers drafting and formatting time, never legal analysis work.
Your drafting volume
Estimate
Estimate based on the assumptions you enter. It does not constitute a commitment to results. Actual gains depend on your cases, your organization, and the time it takes to adopt the tool. Legal analysis, strategy, and review time are not included in this calculation.
The right method in three steps
1. Start with delegable sections. Header, recall of procedure, list of exhibits. No risk, immediate gain, quick team adoption.
2. Feed the model with your documents. About ten validated submissions are enough for it to adopt the firm's formalism. This is what distinguishes a usable tool from a generic text generator.
3. Index your documentary base. Your past documents, notes, templates. The model then relies on your own writings rather than its training, which greatly reduces the risk of invention.
Required hardware
A model capable of reasoning on long legal texts requires approximately 16 GB of video memory. The choice then depends on the number of connected people and the size of the fund to be indexed.
CoreAI 16 — RTX 5060 Ti 16 GBIndividual workstation. Assisted drafting and light document base.
€1,703
CoreAI 32 — RTX 5070 Ti 16 GBHigh-volume firm. Indexing a significant document base.
€2,442
NVIDIA GB10 Mini AI Server128 GB unified. Entire firm connected, complete document base.
€3,999
CoreAI 64 — RTX 5090 32 GBMost powerful models for long and complex documents.
€6,042
Frequently Asked Questions
Can legal document drafting really be automated?
Partially. Mechanical sections—header, recall of procedure, list of exhibits—can be delegated without difficulty. The statement of facts and the structure of the discussion can be assisted. Legal reasoning, strategy, and characterization are not delegable.
Can AI find case law?
Not from memory, and this is a critical point. A model queried without a documentary base produces references that appear authentic but can be entirely invented. Any reference must come from your usual databases or be verified at the source before citation.
How can professional secrecy be preserved?
By using a local AI. Processing is carried out on a firm's machine, with no outbound connection required. Case documents are not transmitted to any external provider, which directly addresses the obligation of secrecy.
Can the model write in the firm's style?
Yes, by giving it access to your validated documents. It then adopts your formalism, your phrasing, and your vocabulary. Without this guidance, it produces a recognizable generic style.
What hardware is needed?
A workstation with 16 GB of video memory is sufficient for individual use, starting from €1,703. For an entire firm sharing an indexed document base, a compact server like the GB10 is more suitable.
Who remains responsible for the document?
The lawyer who signs it, entirely. AI is an assisted drafting tool. It does not modify professional responsibility, the obligation to review, or applicable ethical rules.
This article presents general benchmarks and does not constitute legal or ethical advice. The use of drafting assistance tools remains subject to applicable professional rules and the discretion of each firm. Radiance Systems provides the hardware and technical support.




