Your standard services agreement has gone out for years and come back with a handful of comments. This time it returns with 40 tracked changes. Half of them reword clauses nobody has questioned before, a few delete a word and add the same word back, and the liability cap now appears in four places. Somewhere in there are the two or three changes that actually affect your client.
Finding them is the problem. A markup like this takes far longer to read than it took to produce, and every hour spent reading it is an hour your client either pays for or you absorb. The temptation is to work through it clause by clause, answering each change on its own terms. That is how a two-round negotiation turns into five.
The likeliest explanation is that the other side ran your template through an AI tool before a lawyer looked at it. An AI reviewer given no instructions doesn't know who it acts for or what that party cares about, so it treats every clause as open and moves each one toward the worst case for its side. It has no way of knowing that your limitation of liability has been accepted by counterparty after counterparty, or that your payment terms are standard for the industry. Without that context, every clause looks like a negotiating opportunity.
You can't control the prompt the other side used. You can control how quickly you sort what came back, and whether your own tools send the same thing in return.
Unreviewed AI markup has a recognizable pattern
No single sign proves anything, but a few together usually do.
Minor edits arrive with long explanations or citations attached. A lawyer changing "promptly" to "within five business days" rarely writes a paragraph justifying it, but a tool asked to explain its changes will. Words are deleted and then reinserted unchanged, which tends to happen when a tool rewrites a sentence and the result happens to match the original. Whole sections are struck and restated where a few words would have done, because restating is easier for a model than making a precise edit.
Structural changes are another giveaway. Cross-references are replaced with bullet points, and liability limits are added clause by clause until the carve-outs are hard to follow. Each addition might look defensible on its own. Taken together, they produce a contract where it is no longer clear which cap applies to which claim.
The clearest sign comes later, when you call to ask about a change and the person on the other side can't explain it. That isn't necessarily a criticism of them. They may have been handed the tool's output with an instruction to send it on.
When you see this pattern, assume nobody on the other side decided which changes were worth negotiating. That decision now falls to you, and the faster you make it, the shorter the negotiation.
Most of the markup can be dealt with in bulk
Read the markup yourself before handing it to any tool. You need your own view of what has changed before you can judge whether anything else has classified it correctly. Then sort each change into one of three groups.
The first group is changes that move a position you won't give up. Reject them and give a one-line reason. A short reason shows the point has been considered and isn't going to move, which makes it less likely to come back in the next round.
The second group is changes that land inside a range you would accept if pushed. Accept them, or counter at your fallback. If your standard cap is 12 months' fees and you would accept 18, a counterparty asking for 18 is asking for something you have already decided you can give. Treating it as a fight costs a round for no benefit.
The third group is changes that reword a clause without altering its legal effect. These can be rejected together with a single short note saying you have not accepted drafting changes that don't affect either party's position. Where a rewording does no harm and accepting it is quicker, accept it.
In most AI-generated markups the third group is the largest. Once it is cleared, the real negotiation is usually a handful of points.
The sorting does need care, because some changes that look stylistic are not. Moving from "shall" to "may", deleting "reasonable", or changing "and" to "or" in a list of conditions can each shift who bears a risk. These are the edits that are easiest to wave through when you are working quickly through a long markup, so read the third group twice before you send it back.
AI is useful for the sorting once you have read the draft yourself. Ask it which edits change your position and which only change wording, which of the changed clauses were already inside your fallback range, and how the remaining edits rank by risk to your client. It can then draft a response that accepts, rejects or counters each change, for you to check. Used this way, the tool does the clerical part of the work and the decisions stay with you.
Sometimes a call is quicker than another round
When the markup is plainly excessive, pick up the phone. Three rounds of AI-generated redlines can take weeks, while 20 minutes on a call can settle the same points in an afternoon.
The person on the other side is often overloaded and using AI to keep up. They may not have strong views on most of the changes they sent. Asking which changes they actually need tends to shorten the list quickly, and it gives them a way to drop points without having to concede them in writing. Go into the call with your three groups already sorted, so you can say which points you will accept straight away and which you need to discuss.
An AI with no brief produces the same markup on your side
Your own AI will behave exactly like theirs if you give it nothing to work from. "Review this agreement" produces a generic, worst-case markup, and sending that back invites another round of the same.
Before any review, the tool needs four things. It needs to know who you act for and what the deal is, in a sentence, because a supplier reviewing a services agreement wants different things from a customer reviewing the same document. It needs the points you won't give up, so it knows where to hold. It needs your fallbacks, the positions you would accept if pushed. And it needs to know what to leave alone, such as spelling, formatting and stylistic edits with no legal effect, so it doesn't fill the markup with changes nobody will negotiate over.
The fallbacks do the most work. Give an AI only your standard position and it will redline every clause back to that position, including clauses already within your acceptable range. That is how a reasonable markup becomes an over-lawyered one. A tool that knows your fallback can leave an 18-month cap alone when your range runs to 18 months, and spend its attention on the clauses that are actually outside it.
Rules work best when a tool can apply them consistently, so keep to one topic per rule and write it as an instruction starting with "Ensure that". A liability rule might read:
| Preferred | Ensure that our liability is capped at 12 months' fees. |
| Fallback | A cap at 18 months' fees. |
| Never accept | Unlimited liability for general breach. |
| Escalate | Anything beyond the fallback goes to the person who signs off. |
The escalation line is easy to leave out and worth keeping. It tells both the tool and whoever reviews its output what happens when a counterparty's position falls outside anything the rule covers, instead of leaving that to be decided case by case.
Each rule also needs a decision about what the tool does when a draft falls short: propose your wording as a tracked change, or flag the issue for someone to decide. A redline suits points where you know exactly what you want, such as governing law or a notice period. A flag suits points where the answer depends on the deal, such as an indemnity whose scope depends on what is being supplied.
A clearer template gives the other side's AI less to question
The other side's AI reads your template before their lawyer does, and some drafting habits invite more markup than others.
Refer to the substance of a clause rather than its number. "Subject to the limitation of liability" survives renumbering and is easier to follow, for a person or a tool, than "subject to clause 14.2". Use each defined term the same way every time, because a term used in two slightly different senses is an invitation to redefine it. Keep to one obligation per clause and name the party that owes it, so there is no ambiguity for a tool to resolve in its own side's favor.
Put the key commercial terms in a single table at the front. A reviewer can see the deal at a glance, and the operative clauses can refer back to it instead of restating figures in several places. Drop blanket "notwithstanding anything" overrides, which make it harder to tell which provision governs and tend to attract carve-outs of their own.
None of this requires a redraft from scratch. Most templates can be tightened clause by clause, and a clause that says the same thing in half the words gives an AI half as much to question.
Where Law Insider fits
Simple AI, the Word add-in included with every Law Insider subscription, runs playbooks inside the document you are reviewing. You can start from one of 55 standard playbooks built from more than 1,000 agreements per contract type, or build your own rule by rule, from a template you like, or from up to 10 files, including past redlines and spreadsheets.
Each rule can hold a preferred and a fallback position, a guidance note explaining your reasoning to whoever reviews, and a choice between a tracked change and a flag. When a counterparty marks up your own paper, first-party paper mode compares the draft against your template and flags every change, tracked or not.
Start a free trial, open the last over-redlined markup you received, and run your playbook against it.