Most bad legal AI output is a briefing failure.
No partner hands a trainee a data room with the words "find the red flags" and walks away. They say who the client is. Buyer or seller. What the deal is. What they want back, and by when. A trainee who gets less than that produces a memo that is technically fine and practically useless.
Yet that is how a great deal of legal AI gets used. A question goes in with no client, no side and no purpose. A balanced, generic answer comes out. And the conclusion people draw is that the model is not good enough.
The model is usually fine. It was never told what the work was.
Prompting was the wrong lesson
The last two years produced a small industry in prompting advice. Magic phrasings. Templates. Courses. It is easy to see why. Prompting looks like a new skill, and new skills are easy to sell.
The useful part of all that advice is old. Say who you act for. Say what the matter is. Say what you are trying to achieve and what you want back. That is a brief. Lawyers have been writing them since their first week in practice.
What changed is not the skill. It is the discipline of applying it. A trainee briefed on Monday still knows the deal on Thursday. A new chat on Thursday knows nothing. So people skip the brief, because writing it again feels like a waste of time, and then they are surprised when the answer is generic.
A model with no goal will find one
These models work hard to give you something. Whatever goal you hand them, they pursue. Hand them none and they will invent one.
That is useful when the goal is clear and a liability when it is not. Ask for "issues" in a share purchase agreement without saying you act for the buyer, and you get a list that helps neither side. Ask for the limitation of liability across 20 contracts without saying what to do when a contract has none, and you may get a confident description of a clause that does not exist.
Disclosure, because it matters here: our own AI is not perfect either. Customers have told us it can be a people-pleaser, and we are building more guardrails into it. Every serious legal AI vendor has to. But some of the most effective guardrails cost nothing and sit with the lawyer. Quote the clause word for word and give the number. Write "not found" when a document is silent. Flag every assumption. Say what you do not want as clearly as what you do.
None of that is clever prompting. It is the instruction you would give a trainee before a due diligence review.
The gap is structural
If lawyers already have the skill, why does the output still disappoint? Because most AI tools are built around a chat, and a chat is built around a single question. Legal work is not.
A matter runs for weeks. It has dozens of documents, several workstreams and a running record of what has been checked. A deal does not end at signing. Once it closes, the employment contracts, offer letters and handbook all have to move onto the buyer’s terms. Every question in those weeks depends on the same context.
When that context lives only in a lawyer’s head, every new chat starts from zero. Better models alone will not fix that. The tools are shaped around a question, and the work is shaped around a matter.
This is the reason we built Projects into Law Insider: one place where the documents and the brief stay put, so every question starts from them. Other tools have reached similar conclusions, and some force every piece of work into a project. We chose not to, because plenty of lawyers come to us with a quick sanity check and do not need the overhead. For anything deeper than a one-off question, though, the brief has to live somewhere it persists.
The brief can draft itself
There is a useful irony here. Writing a good brief takes effort, and the tool that needs the brief is good at helping you write it.
Upload the documents and ask the assistant to help write the instructions for the matter. It reads the documents, asks what it needs to know about scope, deadlines and open decisions, and drafts a brief that is often more thorough than the one a busy lawyer would write. You edit it and save it as the standing instruction.
The judgement stays with you: who you act for, what matters, where the risk sits. What you lose is the blank page.
One habit to change
Before the next substantial question, write the brief you would give a trainee. Four lines will do: who you act for, the matter, the goal and the output. Add the rules you would insist on from a junior, especially "not found". Then keep it somewhere it persists.
Lawyers spent two years learning to prompt. They already knew how to brief.
