How to Check an AI-Drafted Legal Document Before It Reaches a Client

AI drafting tools can turn a term sheet or a client email into a first draft within minutes. That draft still goes out under a lawyer’s name, so the lawyer answers for every clause in it. The errors easiest to miss in an AI draft sit in small details a fast read skims past, such as a party name carried over from an old template. This guide covers what to check before an AI-drafted document reaches a client, starting with the tool itself.

The Drafting Tool Sets How Much Checking a Document Needs

Every AI drafting tool pulls its language from somewhere, and that source shapes the review. A general chatbot writes from patterns in its training data, so the lawyer must test each clause against the firm’s standards. A tool that drafts from the firm’s own templates and signed agreements starts closer to language the team already trusts.

Lawyers comparing the best AI for legal writing can start with one question: where does a tool’s language come from? 

Spellbook, for example, works inside Microsoft Word and drafts clauses and full agreements from AI or from a team’s saved precedents. A team that drafts from its own precedent still reviews the result, though the review starts from familiar language instead of generic phrasing.

The tool also decides where the review happens. A draft that stays in Word keeps tracked changes and comments in one file, which makes the later checks easier to run. Copying a draft out of a chat window erases that history, so the reviewer starts from a blank record.

AI Drafts Break Down in Details a Quick Read Misses

Reviewers pay the most attention to price, liability, and termination terms. Mechanical details get less attention, so their errors can survive review. AI drafting adds a few of its own. Run these checks on every AI draft before anyone reads it for substance.

• Confirm that every capitalized term has a definition and keeps one meaning from start to finish.
• Open each cross-reference, since a draft can renumber sections as clauses move around.
• Search for names, dates, and amounts left over from the precedent the tool drew on.
• Look for bracketed blanks, drafting notes, and internal comments before the file leaves the firm.

An AI draft can phrase the same concept two different ways in different sections, so a defined term drifts by the time it appears again. A lawyer reading for substance follows the deal logic and misses that drift. A separate pass that checks only these mechanics catches what the substance read skips.

Some Word add-ins now run a proofreading pass that flags bad section references, missing definitions, and placeholder text. A pass like that speeds up the search, but the lawyer still decides what each flag needs.

Client Instructions and Jurisdiction Still Need a Lawyer’s Read

An AI draft reflects only the prompt and the documents the tool can read. It has no record of the call where the client agreed to a longer notice period or dropped a warranty. 

Read the draft against the client’s actual instructions and the latest term sheet, line by line. Any term the client changed late in the negotiation deserves a second look, since the draft may not reflect it.

Jurisdiction needs the same care. Governing law, venue, and state-specific requirements change from one matter to the next. A draft can also carry another state’s rules over from its source, so check each provision against the law that applies to this client.

Citations call for a separate check. A brief or memo with case law needs a lawyer to confirm every authority at the source. A separate guide to AI agents for legal research covers that process.

ABA Guidance Treats Draft Review as Part of the Lawyer’s Job

The American Bar Association (ABA) released Formal Opinion 512 in July 2024, its first formal ethics opinion on lawyers’ use of generative AI. The opinion ties generative AI use to duties lawyers already carry, including competence, confidentiality, communication, and fees. 

State bars have also issued their own guidance, so lawyers should check the rules in the states where they practice.

Two points apply directly to AI drafting. Lawyers using these tools still owe clients confidentiality over information about the representation, unless the client gives informed consent. So check a tool’s data terms before pasting client facts into it. Look for terms that stop the vendor from storing client data or training models on it.

The opinion also lets a lawyer bill for the time spent reviewing an AI draft for accuracy and completeness. That point frames review as legal work, so firms can plan review time into every AI-drafted matter instead of squeezing it in at the end.

A Redline Shows the Client What Changed and Why

The last check happens in the redline. Tracked changes let a supervising lawyer see every edit made to the draft. Some tools insert AI suggestions under the lawyer’s own name, so the redline may not show which edits came from the AI. A note in the file recording what the AI changed fills in that history. A clean draft with no history forces the reviewer to read every line cold.

Compare the final draft against the firm’s standard template before it goes out. Any clause that departs from the standard needs a reason the lawyer can explain to the client. If no one can explain a change, the draft goes back for another pass. Accept the AI’s suggestions one at a time rather than all at once, so you can decide on each change.

Speed Only Helps When the Review Keeps Pace

AI drafting cuts the time it takes to produce a first draft, but the lawyer still signs off on the final one. A short, repeatable review catches the errors these drafts introduce before a client sees them. Check the language source, the mechanical details, the client facts, and the redline on every document. That routine keeps AI drafting fast without sending clients a document no one fully read.