The system between the documents
Legal work was always asystem. It just ran on people.
We describe our work by what itproduces. A due-diligence report. An issue list. A mark-up. An acquisitionagreement. A closing bible.
Those are the things we send,review, negotiate and bill for, so it is a natural mistake to treat them as thework.
They are the points where thework surfaces.
The work itself happensunderneath, in the movement between them. And that movement has a structure wehave never had to describe, because it has always run inside people.
A loop nobody has drawn
A mark-up arrives from the otherside. Somebody works out what changed, which changes matter, what they cost usand what we should do about them. That becomes an issue list. The issues arediscussed and decided. The decisions return to the agreement as the next draft.
Four steps, a feedback loop, anda handoff at every one of them. Any engineer would recognise it on sight. Mostlawyers have run it a thousand times without ever seeing it drawn.
And the issue list is not theproduct. Neither is the next draft SPA, or – very likely – the one after that.Each one surfaces the state of the work for a moment before the work goes underagain, all the way to the execution version.
The same shape holds at everyscale. An NDA, a non-binding offer, a letter of intent, successive drafts ofthe acquisition agreement: the documents are where the route becomes visible,and the work moves between them.
What moves is information.
What actually moves
Somewhere in the data room of acompany being sold sits a short internal email. A plastic component that guidesproducts along a production line cracked, and fragments may have reached abatch.
It arrives labelled nothing.
Somebody notices it. Somebodyasks which batch, where it went, whether there were complaints, what theinsurer was told. Other documents support it, contradict it, or make it lessalarming. It gets placed against product-safety law, against the price, againstwhat the purchaser is actually buying.
Then it may become a finding.Then a risk. Then advice, a client decision, a negotiating position, andeventually an indemnity — or nothing at all.
The fact never changes. Itsstatus and its implications do.
Source → signal → candidatefact → validated fact → finding → risk → recommendation → authorised decision →clause or action
Every arrow is a transformation.Every transformation has an input, a method, somebody performing it, and astate it leaves behind.
That is a system.
It always was one
We did not design it. Weinherited it, and we ran it on people.
For most of professional historythe lawyer was the system: a biological computing system running on neuronsrather than parameters and attention layers, trained by law, practice andcorrection, and fuelled by coffee and peanut-butter sandwiches.
That is why it is invisible. Notbecause it is mysterious, but because a system executing inside one person'shead never has to be described to anyone. A junior asks a question, absorbs thepart they can hold, and comes back months later able to understand the nextlayer. Eventually the whole model assembles itself into something we callprofessional intuition and stop examining.
It also means the system hasnever been inspected. Nobody writes down the materiality threshold a partnerapplies without thinking about it. Nobody records why that threshold moved onthis particular deal. We keep every version of the agreement and discard thereasoning that produced it.
Ask the questions anyone wouldask of any other system, and we have almost no answers.
Which source is this based on?What has been validated, and what is still only plausible? Who was entitled tomake this call? If the output is wrong, which part failed?
We can usually answer thosequestions. We just cannot answer them from anything except memory.
Biological compute does not copy
The system works. It has beenproducing capable lawyers for a very long time. But look at how it reproducesitself.
To keep it running, you need asteady supply of new people with the same curiosity and the same appetite forbeing wrong. They spend years asking, failing and being corrected. Then, oncethey are finally good at it, we ask them to spend a large part of theirremaining time explaining it to the next ones — time they are no longerspending on the work they became good at.
Every unit of capacity has to begrown from scratch, in real time, by someone who could have been doingsomething else.
That is not a complaint. It is aconstraint. A system that reproduces only through apprenticeship is bounded bythe number of experienced people willing to sit down and explain, and by howlong they stay.
It is inherently finite.
A described system is not. Oncea transformation is written down — its input, its method, what a good outputlooks like, where it has to stop — it can be run again without being relearned.
None of which has anything to dowith machines being cleverer than lawyers.
A description can be copied. Acareer cannot.
What you start asking once you cansee it
Start with something we alreadydo. Every firm has templates for the documents it produces. Almost nobody hasanything that describes how the variables inside those templates aredetermined, established and validated.
That gap is the system we arelooking for.
Which parts of this recur? Notwhich documents recur — which transformations. The same judgment, applied todifferent facts, week after week.
Which of them can be describedwell enough to be repeated reliably? What does each one need: which sources,whose objective, which framework?
Where does experience have tobecome explicit — a materiality standard, a set of warning signs, a reasoningframework, a single validated example that carries more than a page of rulesever could?
Where can a model infer, andwhere must somebody go and gather? What should be fixed, and what should staydeliberately open, because the right answer depends on circumstances nobody canspecify in advance?
And the one that decides theshape of everything else: which steps can run on their own, which shouldproduce a draft for a person to take over, and where must a human validate,weigh, authorise — or simply pick up the phone, because the act itself belongsto a person?
Most of that is not an AIquestion. It is a question about what we already do and have never writtendown.
Why this matters now
If you are a transactionallawyer, none of this is new. You have been operating this system your wholecareer. My only claim is that you cannot currently point at it.
If you are building legal AI, itis the entire problem. You cannot automate a route nobody has described. Askfive excellent lawyers what should happen to that email and you will get fivedifferent answers, each correct inside its own unstated assumptions.
And if you are running a firm,it explains something that otherwise looks strange: why buying good tools sooften changes so little. A tool accelerates one step inside a system nobody hasdrawn. The step gets faster. The route does not change, the handoffs still losethe uncertainty attached to a finding, and the correction someone made lastTuesday still disappears at the end of the conversation.
A faster step inside anundescribed system is not a transformation. It is a faster step.
The documents are where the worksurfaces. What happens between them is the work.
And that work has always been asystem. The only thing that has really changed is that we finally have a reasonto describe it — and a way to run parts of it that does not depend on findinganother person with the same decade, the same curiosity and the same appetitefor being wrong.
I tried that once, with thesmallest step I could find. It took twenty pages, and it taught me something Iwas not looking for.
