A vendor sent a draft master services agreement last week. I ran it through my own review pass — same playbook I built when we rolled out our CLM at Beyel Brothers, fed through the model that does the first read. The model flagged the usual: an uncapped indemnity, a one-way audit right, a venue clause that didn't make sense for our footprint.
The vendor's counsel sent back the redline a few hours later. I read theirs. About 70% of the changes matched my notes verbatim — same indemnity carve-out, same audit reciprocity, same venue swap. The phrasing on three different sections was almost identical to what I had drafted.
I don't think they read my version. I think their model and my model just read the same training data on what reasonable looks like.
That's the ghost in the redline.
Both sides converge
When both legal departments are using a frontier model to do the first pass on a contract, the redlines flow toward a center. Each model has been trained on roughly the same corpus of what neutral commercial language looks like. The result: neither side asks for the asymmetric concession the other side's model already accepted as reasonable. The gap between Party A's "aggressive ask" and Party B's "aggressive ask" shrinks.
That isn't a bad outcome on most agreements. Routine vendor MSAs, equipment leases, NDAs — the kind of work that makes up the majority of an in-house counsel's queue — benefit from a center-converged starting point. We close faster. Fewer late-stage surprises. The legal team isn't wasting cycles on points that were always going to land near the same place.
But "most agreements" isn't every agreement. And the convergence creates a problem you don't notice until you're three negotiations in.
What the models smooth out
A frontier model reading a vendor agreement doesn't know that this particular vendor has been trying to win your business for two years. It doesn't know that you have a fleet of 60 cranes in Florida and the indemnity language matters differently for you than it does for a tech company in Delaware. It doesn't know that the GC on the other side just lost a different contract over the same carve-out three months ago and is twitchy about it.
Models flatten that context. They produce what a neutral, sophisticated lawyer would write — which is exactly what neither side actually wants on the parts that matter.
The asymmetries you should be taking — the times you should be saying "I know this looks reasonable but I want it phrased my way because of [non-textual reason]" — get harder to find. Both sides converge on a draft that nobody would have signed five years ago, because five years ago each side's first-pass review was done by an associate who didn't know what a fair midpoint looked like across ten thousand prior deals. The associate would push harder, the counterparty would push back, and the friction would surface what each side actually cared about.
The model removes the friction. The contract is cleaner. The actual deal — the one that should reflect the relationship — gets quieter.
Where the human counsel re-enters
The fix isn't to stop using the model. The fix is to know that the first 70% of the redline is the easy part now, and to spend the time you saved on the 30% that the model can't see.
For me, that 30% looks like:
- Commercial context the model doesn't have. What's our actual leverage on this counterparty? Are they desperate to close this quarter? Do we have a relationship that matters more than this one contract? These don't go in the redline. They go in the conversation with my CEO before I send the redline.
- Asymmetric risk that I'm choosing to take. Sometimes I leave language alone that the model flagged, because I want this counterparty to feel that we trust them. Sometimes I add language the model didn't suggest, because I know what a litigated dispute with this industry looks like and the standard mutual indemnity isn't enough.
- Strategic ambiguity. Models hate ambiguity and try to resolve it. There are clauses where I want ambiguity — because resolving it now means a fight I don't want to have, and the ambiguity favors us if we ever need to enforce.
- The signals the markup sends. A redline is a communication, not just a document. Sending back 47 changes when 12 would do tells the other side something about how you negotiate. The model doesn't think about that. The human does.
A simpler way to think about it
The model is a fast, thorough, even-tempered associate who has read every contract ever filed in a public docket. It doesn't know your business, your counterparty, or your appetite for risk. It produces a starting point.
The job that used to take a senior lawyer five hours — the first pass — now takes thirty minutes. The job that's left — the judgment — is the same job it always was. The convergence means that's where the leverage moved.
If both sides' first pass is going to land at the same place, the only thing that distinguishes your work product from the other side's work product is what you put on top.
A practical close
If you're in-house and you haven't yet noticed the convergence, you will. The signal looks like this: opposing counsel returning a redline whose comments match what you would have written verbatim, faster than you would have expected, with more accepted changes than you assumed they would accept on first pass.
When you see it, that's the moment to pause and ask the question the model never will: what does this deal actually need from a human?
The answer is almost always not in the document.