This post uses hypothetical scenarios for illustrative purposes only. It does not describe any actual client, transaction, or representation, and is not legal advice.
The 2026 version of this story plays out the same way in businesses and firms of every size. Someone is asked to update an agreement for a new arrangement. The AI tool produces a polished rebuild in four minutes — relocated clauses, new protective provisions, modernized structure, market-standard insurance, a tidy schedule. It reads like the work of a careful senior associate. The polish is the problem. Nobody asked for most of it, one of the “new” clauses is the original document’s own language moved to a different section, and the redline the tool generated is colored text rather than tracked changes, so nobody can accept or reject a single mark. The draft gets forwarded as “the updated version” by someone who has not read past page two.
Generative AI has made contract drafting faster than at any point in the history of the profession. It has not changed who answers for the result. The person who sends the draft owns the draft — every relocated clause, every invented cross-reference, every confident summary of a provision that says something else. Here are the guardrails that make AI-assisted drafting reliable, for the person whose name is on the work and the client whose deal is on the page.
Know the failure modes before you trust the output
AI drafting tools fail in characteristic ways, and none of them look like failure on the screen. They relocate provisions to where provisions conventionally belong, which reads as tidiness and functions as change. They add every protective term they can justify, because thoroughness is what their training rewards. They produce comparisons that are imitations — text colored red and blue — rather than genuine tracked changes a reader can accept, reject, and verify. They describe their own output with total confidence, including the parts they got wrong. And they occasionally invent things outright: a cross-reference to a section that does not exist, a defined term that was never defined, a summary of a clause that says the opposite. Public court dockets already carry examples of professionals who filed machine-invented material without checking it; the pattern is not hypothetical. None of this means the tools are unusable. It means the workflow around them has to be built for exactly these failures, the way good deal hygiene is built for human ones.
Guardrail one: real tracked changes, and tests decide
A contract turn from an AI tool should be delivered exactly as a turn from a careful associate would be: a clean draft and a redline against the version the client last saw, with the redline made of real tracked insertions and deletions. Then two mechanical tests get run before anyone reads a word. Rejecting every tracked change must reproduce the base document exactly. Accepting every tracked change must produce exactly the clean draft. If either test fails, the pair is misrepresenting what changed, and it does not go anywhere.
The point of the tests is that they replace trust with proof. A tool will describe every output as “the updated version” with perfect confidence, and a reader moving fast will believe it. The tests do not care about confidence. The same goes for the structural sweep — numbering, cross-references, brackets, leftover internal notes — which is precisely the layer where machine errors hide, because a document can be grammatically flawless and structurally broken at the same time. The full protocol is the one laid out in the version-control discipline most deals skip, and an AI-heavy workflow needs it more, not less.
Guardrail two: the instruction is less, not more
Left to its defaults, a language model drafts the way a nervous first-year drafts: it adds every provision it can justify, because each addition is individually defensible and thoroughness reads as competence. On a working form a business has used for years, that instinct produces exactly the wrong document — one the owner no longer recognizes, with three load-bearing changes buried under thirty that are not. The corrective is a discipline the tool has to be told, explicitly and every time: the form is the baseline; every mark must be required by the structure, requested by the client, or fixing a true defect; a clause that looks like an addition gets traced to the source before it is touched, because it may be the original document’s own term relocated; suggestions go in as bracketed notes to confirm, never as silent operative text; and rates, limits, and other economics do not move without a decision by the person who owns them. The reasoning behind that approach is laid out in a companion piece on what a good redline looks like, and it applies with double force when the drafter is a machine that never gets tired of adding.
Guardrail three: data handling is a configuration decision, made once
The question “is it safe to put this document in the AI” has to be answered before the matter starts, not paragraph by paragraph. Consumer-grade tools that train on what users type are a different category from enterprise deployments with contractual no-training and retention commitments, and anyone using these tools on other people’s documents should be able to say which category they are in and why. Client documents, deal terms, and anything that could identify a transaction do not go into a tool whose data handling you cannot describe. Frameworks like the NIST AI Risk Management Framework formalize this instinct — map where the data goes before you rely on the system — and the mapping takes an hour, once, instead of a judgment call every afternoon. The same care extends to what the tool leaves behind: generated drafts and intermediate output live in a clearly separated working area, not commingled with reviewed final documents, so that six months later nobody has to guess which version a human actually adopted.
Guardrail four: every correction becomes doctrine
This is the guardrail with the highest return and the one most users skip. When a reviewer corrects an AI draft — cut the unrequested provision, put the original clause back where it was, restore the coverage requirement the client had chosen, turn the silent addition into a bracketed note — that correction is a rule surfacing. Left in the chat, it evaporates, and the next session makes the same mistake with the same confidence. Written down once, in plain language the tool loads before it touches the next document, it compounds. The operations getting real leverage from these tools keep those rules somewhere they own — a repository under their control, versioned like any other document, readable by every tool and every person in the practice — rather than locked inside a single vendor’s product. The result over a year is a body of drafting doctrine that reflects how the practice actually works, applied consistently by a tool that no longer needs to be re-taught, and improved every time a human catches something new.
Guardrail five: a human reads every page
The last guardrail is the oldest one. Someone reads the document — not the summary of the document, not the tool’s description of its changes, the pages. That reading is where the garbled clause that passed every structural test gets caught, where the relocated fee clause gets recognized as the original document’s own, and where the judgment a client is actually paying for happens. The division of labor that works is the same one that works in agentic M&A diligence: machines for the volume, humans for the decisions. A tool can produce five turns in an afternoon; only a reader can decide which one should exist.
What a client should ask
Clients are entitled to ask about all of this, and increasingly do. The useful questions are concrete. Does the firm or team use AI tools on my documents, and which ones? Where does my information go, and does the tool train on it? How is an AI-produced draft verified before I see it — is there a real redline, and what tests does it pass? Who owns the playbook the tool follows, and does it reflect how my documents should be handled? Anyone who has done the work will answer all four without hesitation. Anyone who cannot is telling you something important about the four-minute draft you are about to sign.
The technology is not going away, and it should not. Used with these guardrails, it lets a small team deliver turns at a pace that used to require a floor of people, without lowering the standard the work has always demanded. Used without them, it produces beautifully formatted documents that nobody read — and in most cases, the market eventually finds out which kind of document it was handed.
If you are evaluating how AI is used on your contracts, or building a drafting workflow that has to survive scrutiny, feel free to reach out to our firm manager, Magda, at Magda@montague.law, or fill out our contact form. Mention you read this post.
