Legal Services: Liability Cannot Be Delegated
Law has a feature no other exposed sector shares. However much a system drafts, reviews or searches, responsibility for the result stays personal, professional and enforceable. That asymmetry — the machine produces, the lawyer answers for it — makes legal services the clearest case for why augmentation is not an ethical preference but the only defensible configuration.
What is genuinely exposed
A large share of legal work is volume processing. Document review across thousands of contracts in a transaction. Locating applicable case law. Comparing versions of the same text. Extracting clauses across a portfolio. First drafts of standardised documents.
These tasks are labour-intensive, poorly rewarded relative to the hours they consume, and have traditionally been carried by the firm's junior ranks. They are also, genuinely, what an automated system absorbs best.
What is not, and why
Three things sit outside a system's reach — and not as a matter of time.
Strategy. Whether to litigate or settle. What to claim and what to concede. How to sequence a defence. It depends on a specific client, a specific opponent, and an objective rarely written down anywhere.
Judgement under ambiguity. Law is not a closed system with one retrievable correct answer. Much of the work is sustaining a defensible reading where several are available.
Liability. This is the decisive one. A system does not appear in court, does not answer to a bar association, and does not carry the consequences of a badly founded filing. The signature still belongs to a person — and with it, the duty to have verified what they signed.
That last point has already produced known incidents: filings submitted containing case citations that did not exist, generated by a system and never checked by the person signing. The failure was not the tool's. It was a working arrangement that treated automated output as if it were a reviewed draft.
Who can do it, and who answers for it
| Task | Can a system do it? | Who answers for the result? |
|---|---|---|
| Bulk document review | Yes, with a clear advantage | The lawyer who signs |
| Locating case law | Yes, if sources are verifiable | The lawyer who cites it |
| Drafting a standard document | Yes | The lawyer who files it |
| Choosing litigation strategy | No | The lawyer |
| Defending a reading in court | No | The lawyer |
The right-hand column never changes. That is the whole argument.
What augmentation looks like here
If liability cannot be delegated, then the only sensible configuration is one where the system prepares the professional's work rather than substituting for it — verifiably.
A system that finds, across ten thousand contracts, the twelve clauses departing from the standard, and points to exactly where they are so the lawyer can read them. That assembles relevant case law with traceable sources, not a summary you have to take on trust. That detects a redraft changing a definition affecting four downstream clauses. That flags an approaching procedural deadline.
The difference from the replacement model is precise: the system does not produce the conclusion. It produces verified material on which the professional concludes. And everything it contributes must trace back to source, because otherwise the lawyer cannot discharge their duty to verify.
The pipeline problem, again
Law firms face what software faces. The tasks being automated are the ones junior associates did — and they are also the tasks by which the craft is learned: reading a lot, seeing many matters, making mistakes under supervision.
A firm that cuts those roles saves now and has no partners in fifteen years. No single firm can fix that alone without putting itself at a disadvantage, which is precisely the sort of collective problem this association exists to address.
What industry teaches here
One industrial concept translates cleanly: traceability.
In a regulated plant, any decision affecting the product must be reconstructable afterwards — what was decided, on what data, and who answered for it. The assistance systems that succeeded in that environment are the ones that leave a trail, not the ones that hand down an answer you have to accept.
Applied to law, the conclusion is direct: useful legal assistance is assistance whose output can be verified back to source. Anything else transfers a risk to the professional that they cannot carry.
What we are looking for in this sector
We need practising lawyers who can specify:
- What share of their time goes to processing volume, and what share to deciding.
- Which verifications are non-negotiable before signing.
- How a junior lawyer is actually trained today, and what happens if that work is removed.
- What a duty of care would require of automated assistance for it to be acceptable.
If you practise law, join the association. Without lawyers at the table, these tools get designed by people who do not answer for what gets signed.