Most legal AI still assumes a simple shape: document in, prompt in, answer out.
That pattern works, up to a point. You get something coherent, often convincing, sometimes genuinely useful. The surface quality has been enough to drive adoption because it gives people an immediate sense of progress.
It also hides where legal work succeeds or fails.
The real work is rarely just answering a question about a document. It is deciding what needs to happen next, who is responsible, what can be relied on, what must be checked and what should happen before anything moves forward. That makes legal work a planning problem before it is a question answering problem.
That point becomes sharper in a world of agentic AI. Once systems can plan steps, call tools, update records or trigger follow-on actions, weak coordination stops being a UX problem and becomes an operating risk.
The illusion of completeness
A well-formed answer can create the impression that the task is finished, even where the underlying work has barely moved forward.
A clause summary might tell you what the clause says, but it does not tell you whether the clause matters in the current phase of the deal, whether it conflicts with another obligation, whether it has already been addressed elsewhere, or whether it needs escalation before anyone relies on it.
The answer feels complete because it is self-contained. Legal work is not.
Most problems in matters are not caused by misunderstanding a clause. They come from missed steps, unclear ownership, unresolved dependencies or decisions made without full context. Optimising for better answers helps, but it does not fix the structure around the work.
Where things break
Take a typical workflow. A document is reviewed, risks are identified, a summary is produced, someone reads it and moves on.
The missing piece is usually not more detail in the summary. It is the surrounding coordination: who owns the risk, whether it has been resolved, accepted or deferred, whether it affects another part of the matter, whether it should block the next step, and whether anything has changed since the summary was generated.
None of that sits neatly inside the answer. It sits in the coordination of the work, which is where many AI systems are still silent.
Documents are inputs, not the system
Legal tech has historically treated documents as the centre of gravity, which made sense when documents were the main artefact being drafted, reviewed, stored and searched.
Once AI is introduced into live workflows, that assumption becomes limiting. Documents become inputs into a broader system: tasks are created from them, obligations are tracked beyond them, decisions are made in response to them and matter state changes independently of them.
If the system only understands documents, it cannot understand the matter. That gap is where risk accumulates, because the system can produce good answers while remaining blind to the operational position those answers sit within.
The missing layer
What is missing is a system of record for the work itself.
Not just a record of which documents exist or which answers have been generated, but a live view of what stage the matter is in, what decisions are pending, what has been agreed, what remains at risk and what is allowed to happen next.
Without that layer, every interaction with AI starts from a partial view. The model may be accurate within the slice of context it receives, but the slice itself may be incomplete, stale or disconnected from the current position of the matter.
Orchestration as a first-class concern
Orchestration is not a technical detail sitting beneath the legal work. It is the structure that determines how the work moves.
It answers questions such as what needs to happen before an output can be used, who is allowed to approve it, what context must be present for it to be valid and what changes once it is accepted.
That sits alongside model choice, not beneath it. A more capable model does not solve a coordination problem. In some cases, it makes the problem harder to spot because the outputs look more fluent, more complete and more authoritative than the surrounding process can justify.
Why this matters now
As AI becomes embedded in legal workflows, the failure modes change.
The obvious concern is an incorrect answer, but the more difficult risks are often correct answers used in the wrong way, incomplete outputs relied on too early, or decisions made without visibility of the full state of the matter.
Those failures are harder to detect and harder to unwind because they appear after the model has done its part. They sit in use, reliance and downstream action, which is also where professional accountability starts to attach.
Firms that continue to optimise for answer quality alone will see diminishing returns. The gains are real, but they plateau if the surrounding system cannot manage ownership, timing, dependency and control.
A different way to think about it
If legal work is modelled as a planning problem, a different set of priorities emerges.
The focus shifts towards how tasks are defined and sequenced, how state is captured and updated, how decisions are gated and recorded, and how context is carried across steps. AI still plays a significant role, but as part of a system rather than the centre of it.
The question changes from "Can the model answer this?" to "Should this answer be used, by whom, in what context, and what happens next?"
That second question is harder, but it is the one that determines whether the system can be trusted.
Where this leads
Once planning becomes the focus, several things follow naturally:
- matter state becomes a first-class concept, not an afterthought
- evaluation moves from prompt testing to scenario testing
- governance shifts from policy documents to enforced controls
- routing decisions consider context, not just cost or capability
At that point, legal AI starts to resemble other mature operating disciplines. Not because legal work becomes software, but because coordination, sequencing and control start to matter more than individual outputs.
Better answers will continue to improve legal workflows, but they are no longer the only constraint. The harder problem is how those answers are integrated into the work, governed at the point of use and connected to the live state of the matter.
Until that is addressed, many systems will remain impressive in isolation and fragile in practice.