The hallway answer
A question asked in the hallway gets a careful non-answer: it depends. The same question asked with the file open gets a real answer, with reasons. The difference is not intelligence - it is context: who the parties are, what was actually signed, which jurisdiction governs, what has happened since.
The same is true of AI - and it is why so many firms are underwhelmed by it. A model without context answers like the hallway: fluent, general, and unaccountable to the facts of the matter. Fix the context and the same tool starts returning hours to fee earners, answers clients can rely on, and work partners can trace and stand behind. Most disappointment with legal AI is not a model failure. It is a context failure.
Prompting is context work done by hand
The first generation of legal AI use looks like this: open a chat window, paste in a clause, describe the background, ask the question, and repeat the ritual tomorrow. The lawyer is doing the context assembly by hand, one paste at a time, and the quality of the answer tracks the patience of the paster. That is senior time spent on transcription, not judgement - and it works in a demo, not across a caseload.
The future of legal AI is not better prompting. It is systems that already hold the context, structured and ready, so the question is the only thing you type.
What context means in legal work
Context in legal work is four layers, and a useful answer usually needs all of them.
- The matter. Parties, dates, deadlines, obligations, what has been agreed, what is in dispute. The living state of the case.
- The documents. The clauses as they actually read, in the version that actually governs - not as anyone remembers them.
- The law. Current, jurisdiction-correct, from sources with real authority.
- The firm's own knowledge. Playbooks, templates, positions taken before - the standards that keep advice consistent from one matter to the next.
A new associate is useful only once they hold all four. A machine is no different. Hand it one layer and it will answer from one layer - confidently.
Context is engineered, not accumulated
More context is not automatically better context. Dumping an entire data room into a model does not produce a considered answer, any more than handing an associate forty binders produces one by morning. The real work is retrieving what actually bears on this question, weighing it the way a lawyer would - authority and recency, not just resemblance - and keeping the receipts so every claim can be traced back to the passage it stands on. The question for a firm is no longer 'how do we phrase the prompt'. It is 'does the system hold our context, and can it show us what it used'.
Consolidation is a context argument
This is also the honest case for one platform instead of many. When documents live in one tool, deadlines in a second, correspondence in a third, and the AI in a fourth, context is fragmented by architecture - and every handover between tools strips a little more of it away. No integration glue restores what the structure throws away.
When the matter, the documents, the law, and the firm's knowledge live under one roof, context compounds instead of fragmenting. Everything the platform holds is context the AI can draw on - without anyone pasting anything. That is what a matter in PONS is: not a folder, but structured context, ready for any question that comes next.
What changes in practice
Context makes the machine a better reader: it does the reading, assembles what bears on the question, and shows its work, so decisions land on evidence instead of memory - and nobody is carrying the file to the machine one paste at a time.
In legal work, an answer is only as good as the context behind it. The same is now true of the tools.