Legal AI Drafting: How AI Writes Contracts, Briefs, and Demand Letters
Legal AI drafting means using generative AI to produce the first draft of a contract, litigation brief, or demand letter before an attorney reviews and finalizes it. Modern legal AI software can turn a routine NDA or a settlement demand into a working draft in minutes instead of hours, though every output still requires a licensed attorney’s sign-off before it goes out the door.
Fifty-four percent of legal professionals already use AI for drafting correspondence, making it the single most common AI use case in law practice today, according to the 2025 Legal Industry Report, which surveyed more than 2,800 legal professionals. The sections below walk through how AI drafting works for three document types — contracts, briefs, and demand letters — and how firms can use it without risking court sanctions.
What Legal AI Drafting Is and How It Works
At its core, legal AI drafting applies large language models to produce contracts, briefs, and demand letters from a prompt, a set of facts, or an existing template. The output quality depends heavily on how the underlying tool sources its answers — whether it reasons from general training data or pulls from a verified body of legal text.
From Blank Page to Grounded Drafting
There are two broad approaches to AI legal drafting. A blank-page chatbot such as ChatGPT or Claude (Anthropic) generates language from its general training knowledge, with no guarantee that a cited statute or clause reflects current law. Grounded drafting tools instead use retrieval-augmented generation (RAG) to pull language from a firm’s own clause library, contract playbook, or a curated statute database before generating a response. Vaquill reports that grounded AI drafting tools reached 93% accuracy on recently amended law in vendor testing, compared with 33% for a general-purpose AI tool answering the same prompts — a gap worth keeping in mind, since it comes from the vendor’s own benchmark rather than independent testing.
The practical difference shows up fastest on anything that changed recently: a new statute, an amended regulation, a court ruling that narrowed a standard clause. A blank-page tool has no way to know its training data is stale, while a grounded tool can be pointed at an updated source document before it drafts a single sentence. For a deeper look at how these systems process legal language under the hood, see how legal AI works.
The Junior-Drafter Frame
The most useful mental model for legal AI drafting is a fast junior associate: it produces a usable first draft in seconds to minutes but carries none of the professional responsibility for the final product. AI first drafts require attorney review — Vaquill lists verification as a “non-negotiable” fourth step in its six-step legal AI workflow. Several federal courts have gone further, issuing standing orders that require attorneys to certify that every AI-assisted citation in a filing has been checked against the actual case law, a trend Harvey has documented across multiple jurisdictions.
The frame matters because it sets expectations correctly on both ends. Treating an AI draft as finished work invites the kind of citation and cross-reference errors that show up later in this article; treating it as worthless because it isn’t perfect ignores the hours it saves on the routine parts of drafting. The workflow that holds up in practice is closer to supervising a capable junior than to either trusting AI output blindly or writing everything from scratch.
Drafting Contracts with AI
Contract drafting is where legal AI adoption is furthest along, largely because so much contract work is repetitive and template-driven rather than novel.
What AI Does Well in Contract Drafting
AI handles first drafts of standardized documents well: NDAs, data processing agreements (DPAs), MSA amendments, and offer letters are all high-volume, low-novelty documents where AI drafting delivers the clearest time savings. GC AI estimates that drafting three new NDAs a week at 45 minutes each adds up to 117 hours of attorney time a year — hours a grounded drafting tool can largely reclaim. The strongest implementations don’t generate contract language from scratch; they assemble a document from a firm’s own approved clause library and playbook. A typical worked clause might pull a limitation-of-liability provision with a 12-month fee cap under Delaware law directly from an approved template, rather than composing new language.
- NDAs and confidentiality agreements
- MSA amendments and order forms
- Offer letters and standard vendor agreements
- DPAs built around an approved clause set
The pattern holds across in-house teams as much as firms: the documents AI handles best are the ones a legal team has already standardized internally, because the AI is really assembling from that existing standard rather than inventing new language.
Where Contract Drafting Needs Human Review
AI-drafted contracts still need a careful human pass on several fronts. Cross-references are a common failure point — Vaquill documents a case where an AI draft included an indemnification cross-reference to a clause that didn’t actually exist in the final document. Word choice matters too: the difference between “fees paid” and “fees paid or payable” can shift financial exposure by months. Jurisdiction-specific rules are another blind spot; California’s AB 1076, in effect since January 1, 2024, bans most non-compete clauses outright, and an AI trained on older templates may not reflect that change. Once a contract draft is in hand, the next step is redlining and clause-level analysis — a process covered in legal AI for contract review.
| Document type | What AI does well | What still needs a human check |
|---|---|---|
| Contracts (NDAs, MSAs, DPAs) | First draft from an approved clause library, cross-referenced against a playbook | Cross-references, non-standard fee or liability language, jurisdiction-specific clauses |
| Internal memos | Summarizing facts and issue-spotting from source documents | Legal conclusions and risk framing |
| Policies | Drafting from an existing policy template or regulatory checklist | Fit with the firm’s or client’s actual practices |
| Board consents | Standard resolution language from prior filings | Corporate formalities and signature authority |
| Demand letters | Fact-to-narrative drafting from case intake data | Tone, settlement figures, and the decision to route through an attorney |
Vaquill’s breakdown of these five categories illustrates a pattern that holds across most legal AI drafting: the more standardized the document, the less human rework it needs before it is ready to send.
Drafting Litigation Briefs with AI
Litigation briefs carry the highest stakes of the three document types, because a fabricated citation filed with a court is not just an embarrassment — it can trigger sanctions.
Speed Gains and the Citation Problem
The time savings can be substantial: Harvey reports that one litigation boutique cut its complaint drafting time from roughly 10 hours to under 2 hours using AI assistance. But briefs are also the category where AI drafting has produced the most public failures. In 2023, two New York attorneys were sanctioned after filing a brief that cited six court cases fabricated by ChatGPT, none of which existed — a case that became widely known as Mata v. Avianca. The episode illustrates why Rule 11 of the Federal Rules of Civil Procedure and Model Rule 3.3’s duty of candor toward the tribunal apply directly to AI-assisted filings — the full text of Rule 11 is available through Cornell’s Legal Information Institute.
The court’s reasoning in that case matters as much as the sanction itself: the judge found bad faith not in using ChatGPT, but in failing to correct the record once the fabricated cases came to light. That distinction has shaped how many federal courts now word their AI-related standing orders — the obligation is less about banning AI outright than about requiring attorneys to catch and disclose errors before a filing goes in.
Brief-Specific Tools and Verification
A small category of legal-specific tools exists specifically to check AI-assisted or attorney-drafted citations against primary sources before filing. Clearbrief, which holds SOC 2 certification and integrates directly with Microsoft Word, cross-checks every citation in a brief against the underlying case record. Other tools in this space include BriefCatch, Lexis+ Brief Analysis, and WordRake, whose more than 35,000 algorithmic editing rules focus on writing clarity rather than citation checking. Thomson Reuters’ Drafting Assistant covers a related but separate job: building a table of authorities and formatting citations to Bluebook and state-specific rules. None of these tools replace the underlying obligation: every citation in a filed brief should be manually confirmed in Westlaw or Lexis before it goes to the court.
Firms that litigate frequently tend to layer these tools rather than pick one: a citation-checking add-in during drafting, a style pass before the brief is finalized, and a manual Westlaw or Lexis check as the last step before filing. No single tool in this category claims to replace that final manual check, which is itself a signal of how seriously the profession has taken the citation-fabrication problem since 2023.
Drafting Demand Letters with AI
Demand letters are usually the fastest AI drafting win for firms handling high volumes of similar claims, particularly in personal injury practice.
Step-by-Step: An AI Demand Letter Workflow
Clio outlines a five-step process for producing an AI-assisted demand letter:
- Gather the case facts — injuries, damages, timeline, and supporting documentation.
- Choose a drafting tool suited to the claim type and practice area.
- Enter the case-specific details into the tool as structured inputs.
- Customize the tone and the specific demands or settlement figure.
- Have an attorney review the draft in full and route it into the firm’s standard workflow before sending.
A demand letter that once took an associate several hours to draft from scratch can move through this workflow in minutes, though step five remains mandatory regardless of how polished the AI draft looks.
When a Demand Letter Needs a Lawyer
Free demand letter generators, such as the tool offered by HAQQ, produce a usable template but are not a substitute for legal advice. HAQQ itself recommends that letters involving amounts above $5,000 or any disputed liability be sent through a licensed attorney — a letter on attorney letterhead is taken more seriously by insurers and opposing counsel than one from an unrepresented claimant. Sending the letter by certified mail with a return receipt establishes proof of delivery, which matters if the dispute later goes to litigation. A demand letter typically produces one of four outcomes:
- Full payment of the amount demanded
- A negotiated settlement for a lesser amount
- Silence from the recipient, which supports moving to a lawsuit
- A counterclaim disputing the underlying claim
Which outcome a given letter is likely to produce often depends more on who signed it than on how it was drafted. A well-written AI-assisted letter on attorney letterhead can prompt a response that the same letter sent directly by a claimant would not.
Choosing a Legal AI Drafting Tool
The tool landscape for legal AI drafting splits cleanly into two tiers, and the right choice depends on the sensitivity of the data involved and the document type.
General-Purpose vs Legal-Specific AI
| Factor | General-purpose AI (ChatGPT, Claude, Gemini) | Legal-specific AI (Spellbook, Harvey, CoCounsel, GC AI) |
|---|---|---|
| Typical cost | ~$20/month per seat | $100s to $500+/seat/month (e.g., GC AI) |
| Grounding | General training data | Clause libraries, playbooks, case law RAG |
| Word integration | Limited or none | Native (e.g., Spellbook in Microsoft Word) |
| Confidentiality controls | Vendor-dependent; check data-use terms | Built for legal data handling |
ChatGPT Pro, Gemini Advanced, and Claude are all priced around $20 a month and produce reasonable first drafts, but they generate language from general training knowledge, carry a higher hallucination risk on niche legal questions, and raise confidentiality questions if client facts are typed into a general-purpose chat window without a data-use agreement in place. Legal-specific tools cost more — GC AI, for instance, runs roughly $500 per seat per month — but add grounding, clause libraries, and direct integration into Word, and Spellbook alone reports more than 3,000 law firms using its platform. Any vendor-reported benchmark in this space, including the figures above, should be read as a vendor claim rather than independent verification.
Most firms don’t pick one tier exclusively. A common pattern is a general-purpose subscription for brainstorming, internal research summaries, and other low-sensitivity tasks, paired with a legal-specific tool for anything that touches client data or produces a document that will actually be filed or sent.
Security and Compliance Checklist
Before adopting any legal AI drafting tool, a firm should confirm the following with the vendor:
- SOC 2 or ISO 27001 attestation covering the vendor’s security controls
- GDPR and CCPA compliance documentation for how client and case data is processed
- A written zero-data-retention agreement with the underlying model provider
- Explicit written confirmation that firm data is excluded from model training
- Encryption in transit and at rest, typically AES-256
Entering client information into a general-purpose chatbot without one of these agreements in place can amount to a disclosure of confidential information under Model Rule 1.6, regardless of how the AI vendor’s marketing describes its privacy practices.
Ethics and Professional Responsibility
Every question about AI legal drafting eventually comes back to professional responsibility, since the attorney signing a document remains accountable for it no matter which tool produced the first draft.
What ABA Formal Opinion 512 Requires
In July 2024, the American Bar Association issued Formal Opinion 512, its first comprehensive guidance on lawyers’ use of generative AI. The opinion addresses five areas:
- Competence under Model Rule 1.1 — understanding how a given AI tool works and where it can fail
- Confidentiality under Model Rule 1.6 — generally requiring informed client consent before case details are entered into a third-party AI system
- Candor toward the tribunal, including verification of anything an AI tool cites
- Supervision of AI use by subordinate lawyers and staff under Rules 5.1 and 5.3
- Billing, since a lawyer cannot charge a client for time an AI tool actually saved
Confidentiality tends to be the area that trips up firms first, since it requires a judgment call every time a new matter’s facts get typed into a tool.
A lawyer’s use of a GAI tool is subject to the same ethical rules and principles as the lawyer’s use of any other type of technology or nonlawyer assistant, and lawyers must understand to a reasonable degree the benefits and risks associated with the technologies they use.ABA Formal Opinion 512
Practical Guardrails for a Firm
Put the AI policy in writing. A short, written policy naming approved tools, prohibited uses, and a client-data rule gives associates and paralegals a clear line not to cross, rather than leaving the judgment call to each individual.
Keep an approved-tools list. Restricting AI use to vendors the firm has already vetted for SOC 2 status and data-retention terms closes off the biggest source of accidental confidentiality breaches.
Make citation verification mandatory, not optional. Every AI-assisted citation in a filed document should be checked against Westlaw or Lexis before submission — a rule that would have prevented the sanctions in Mata v. Avianca.
Train the team, not just the partners. More than 6,000 lawyers have completed continuing legal education courses focused specifically on AI use, reflecting how quickly bar associations have moved to formalize training expectations.
AI does not replace a lawyer’s professional judgment on any of these documents. It compresses the mechanical part of drafting, but the analysis, the strategy, and the responsibility for what gets filed or sent still rest with the attorney of record.
FAQ
This article is for informational purposes only and is not legal advice. Consult a licensed attorney for guidance on your specific matter.