An AI Mission for Contract Review

Legal queues are sorted by who asked first, not by what is at stake. The document carrying the real exposure is almost never the one with the biggest number on it — and that mismatch is where in-house review quietly breaks.

On a Monday morning, the shared legal inbox of a mid-sized company holds somewhere north of two dozen items, and almost none of them are the kind of thing anyone imagined doing when they trained. There is a mutual confidentiality agreement for a conversation that will probably go nowhere. There is a renewal of a facilities contract at a value that would not survive a rounding error on the quarterly report. There is a statement of work from a sales team that has already promised the client a start date. There is a three-seat expansion of a software subscription, on the vendor's own paper, which happens to carry an indemnity obligation running in one direction with no ceiling on it, sitting in the queue beneath a master services agreement worth two hundred times as much that has already been negotiated to a house template and is functionally finished. Whoever opens that inbox will work down it more or less in the order things arrived, because there is no other order available, and the small renewal with the open-ended indemnity will get the same fifteen minutes as the one that needed nothing at all.

That is the actual failure mode of contract review inside most organisations, and it has very little to do with speed. It is not that the queue is slow; it is that the queue is unsorted, and a queue with no sorting logic distributes the one genuinely scarce resource — a qualified lawyer's sustained attention — as though every document deserved an equal share of it. Some of these agreements need real judgement, the kind that draws on what the company has agreed to before, what it has been burned by, and what it can actually live with. Most of them need only a competent, careful check that nothing unusual has been slipped in. The problem is that from the outside, before anyone has read them, the two categories look identical.

Commercial size is the worst available proxy for legal exposure

The instinct every business has is to triage by value, and it is an entirely reasonable instinct that happens to be wrong. Deal size tells you what the organisation stands to gain and says almost nothing about what it stands to lose, because exposure in a commercial agreement is not a function of the fee — it is a function of what the company has promised to do, to warrant, and to stand behind if things go badly. A modest subscription can carry obligations that outlive the subscription by years. A large, well-papered agreement negotiated from the company's own template can be, in risk terms, the least interesting document in the pile.

This is why the "commercially trivial, legally non-trivial" agreement is such a persistent hazard. It arrives without ceremony, usually on a counterparty's paper, usually for an amount that no one wants to spend a week arguing about, and it is precisely the class of document where the pressure to just get it signed is highest and the scrutiny is lowest. Nobody escalates a small renewal. Nobody convenes a call about a supplier's standard terms. And so the risk that eventually surfaces in one of these agreements almost never surfaces at the point of signature — it surfaces years later, when someone goes looking for what the company actually agreed to and finds that the answer is worse than anyone assumed.

The organisational response to this, historically, has been either resignation or a clause-level playbook that everyone is supposed to apply consistently and nobody can apply consistently at volume. Both responses fail for the same reason. Consistency at the twentieth document of the day is not a matter of discipline or professionalism; it is a matter of human attention, which degrades in a completely predictable way across a long queue of superficially similar text, and degrades fastest exactly where the text is most familiar. The reviewer who has read four hundred variants of the same supplier terms is not more likely to catch the fifth-hundredth variant's unusual paragraph — they are less likely, because familiarity is what allows the eye to skate.

Reading at volume and deciding are different jobs

The useful move here is to stop treating contract review as one activity and start treating it as two that have been jammed together by circumstance. The first is a reading job: work out what this document actually says, where it departs from what the company normally agrees, which obligations are open-ended, which protections are absent, and how it compares to what has been signed with this counterparty and in this category before. The second is a judgement job: decide whether those departures are acceptable given this relationship, this counterparty, this moment, and this company's appetite. Lawyers are hired for the second job and spend most of their week on the first, and it is the first that scales badly and the second that cannot be scaled at all.

Separating them is what makes a machine useful here, and also what defines the limits of its usefulness. A system that reads every incoming agreement in full, against the organisation's own accumulated history of what it has agreed to, can do something no human queue can do: it can rank the queue before anyone opens it. It can identify that the small renewal contains an unusual and open-ended obligation, that the large master agreement is a clean instance of the house template, and that a third document is materially identical to one signed with the same counterparty eighteen months ago except in two respects. That output is not advice and should never be presented as advice. It is a reading — an ordered set of observations about what is in the documents and where they diverge from precedent — and its entire value lies in putting the documents that carry real exposure in front of a lawyer first, while the ones that need only a competent check are surfaced as exactly that, with the specific points that warrant a look already marked.

This is the practical form an AI Mission takes in a legal function: not a promise that software will handle contracts, but a defined piece of work with a defined output. Specialist agents read each incoming agreement against enterprise knowledge that includes the company's own prior positions, produce structured observations on what each document contains and how far it sits from precedent, and hand the queue back sorted by exposure rather than by arrival time or by fee. The reasoning layer does the comparison work at a volume and a consistency no reviewer can sustain past mid-afternoon. Every material output routes to a human, because the reading is an input to a decision and never the decision itself.

What the machine is allowed to conclude

It is worth being blunt about the boundary, because this is precisely where the current wave of enterprise AI most often oversells itself. Gartner has predicted that more than forty percent of agentic AI projects will be cancelled by the end of 2027, citing unclear value, escalating cost, and inadequate risk controls, alongside what it calls "agent washing." In legal work the temptation to overreach is particularly strong and particularly costly, because a system that appears to clear contracts is far more marketable than one that merely sorts them. Appearing to clear contracts is the one thing such a system must never do, and the distinction is not a matter of caution or tone but of what the output actually is.

So the line has to be drawn explicitly and defended. A system can read a document, compare it against precedent, and flag what a person should look at. It cannot decide that an agreement is acceptable. Advice on whether the company should accept a term is given by a lawyer, and the agreement is entered into by a person with actual authority to bind the organisation — those two steps belong to people, they are the point at which responsibility attaches, and no amount of confident-sounding machine output changes who is answerable afterwards. Human-in-the-loop is not a compliance decoration bolted onto this workflow; it is the workflow. What the machine removes is the part of the work where nothing was being decided at all: the reading, the comparing, the remembering of what was agreed with whom, the drudgery that was consuming the attention the decisions needed. This is the same reallocation that the publication tracking the shift to the autonomous enterprise documents across other functions, arriving in legal with an unusually sharp constraint on where the automation stops.

The reframe worth keeping is that a review queue is not a backlog to be emptied. It is a portfolio of exposure that happens to be shaped like a to-do list, and the only meaningful measure of how well a legal function is running is not how many agreements it turned around last month but whether the ones that mattered got the judgement they deserved and the ones that didn't stopped consuming it. Emptied queues have always been easy to produce, which is why throughput makes such a comforting metric and such a poor one. The organisations that get this right will stop asking their legal teams to go faster and start giving them a queue that arrives already sorted by what is actually at stake — which is the one question the queue itself has never been able to answer.