A firm that uses AI on legal work and tells you it simply works is describing its marketing, not its process. The failure modes of these systems are known, recurring and — this is the important part — controllable. The difference between serious use and careless use is not the model. It is what is built around the model. Here are the four failures we plan for, and how we catch them.
Missing context
A system reads what it is given. It does not know about last week's phone call, the history of the business relationship, or why a clause is worded strangely — a hard-won concession or a drafting accident. The typical symptom is a finding that is textually right and commercially wrong: a clause flagged as unusual that was, in this deal, exactly what the parties intended.
The control here is not technical. A lawyer who knows the matter frames the instruction before the analysis starts, and reads the output against what the deal is trying to achieve. Capturing the commercial context belongs at the beginning of the process, not at its end.
Plausible but wrong
The most dangerous failure mode: a fluent, coherent reading that fits the words but not the law — or that misses how two clauses interact. The wrong answer arrives in the same confident prose as the right one; nothing in the tone gives it away.
The classic example is a clause read in isolation: a termination right that looks broad until a definition forty pages earlier narrows it to almost nothing, or a liability cap that seems solid until an indemnity elsewhere sits outside it. Each reading is defensible on its own; together they are wrong.
Two controls answer this. First, source verification: every finding cites the exact passage it rests on, so checking means opening the source — not rebuilding the analysis. Second, structured checklists by document type and matter type — which clauses must be read together, which questions must be answered for every contract of this kind — so that review does not depend on what happens to catch the eye that day.
Overconfident summaries and outdated sources
Summaries compress; bad summaries compress away the reservations the original carefully preserved. "May be liable" becomes "is liable". The symptom to watch for is a summary that reads more decisively than the document it summarises: when many pages of carefully qualified drafting yield one page of firm statements, precision has been lost somewhere — and in legal work, the qualifications are often the substance. And the law moves: whatever a system carries from its training ages from the day it was learned.
We therefore treat summaries as navigation, not authority — decisions are made against the underlying text. And any legal position is checked against current sources before anyone relies on it.
The lawyer's job is to challenge
The controls above are procedural; the last one is personal. A named lawyer reviews the output with the express task of attacking it: questioning what is surprising, verifying what the conclusion rests on, and answering for the whole. That is a different activity from skimming and signing — it presupposes a reviewer with the time and the standing to say "this is wrong", and a process in which saying so is expected rather than exceptional.
This is also where the economics of the setup show themselves. The time the system saves on the first read is not pocketed as margin; it is reinvested in exactly this challenge. A firm that automates the reading and thins out the review has not modernised its process — it has removed the part that was protecting you.
Uncertainty is escalated, not smoothed over
Behind all four failure modes sits the same tendency: systems smooth over uncertainty. Ours are built to surface it — "this clause is ambiguous", "this document was partly unreadable", "these two sources conflict" — and to pass it to the lawyer rather than resolve it silently. Uncertainty declared is information; uncertainty hidden is a risk.
If a provider tells you their AI has no failure modes, that is the failure mode. We would rather show you the controls. If you want to see what they look like on a real matter, talk to us.