home  /  verification & quality control  /  supervision & delegation
verification · ai for legal practice

Supervision and delegation.

Firms keep asking what new rules apply. The more productive question is which existing ones already fit — and supervision fits almost exactly.

begin here

Where is your firm?

Start a conversation with the AI Adoption Concierge, already scoped to supervision & delegation. Pick a starting point, or describe your situation directly.

AI Adoption Conciergesupervision & delegation · orientation, not legal or ethics advice
Tell me how delegated work is supervised at the firm now — who reviews what, and at what point. I'll help you extend that to AI-assisted work rather than inventing something parallel.

A great deal of the anxiety about AI in practice comes from treating it as unprecedented. In one important respect it is not: firms have always had a framework for work produced by someone who does not carry ultimate responsibility for it. Associates draft, partners are accountable. Paralegals research, lawyers sign. The duty to supervise, and the principle that delegation does not transfer responsibility, are well developed and well understood. ABA Formal Opinion 512 addresses supervisory responsibility directly among the duties generative AI engages. Approaching AI output as work requiring supervision rather than as a novel category gives a firm a working answer immediately, and one its people already know how to operate.

mechanisms

Where the analogy holds — and where it does not.

Useful frames are useful precisely where their limits are understood.

Responsibility does not transfer

The signing lawyer owns the work product regardless of what produced the draft. This holds exactly.

Review is proportionate to risk

Firms already calibrate scrutiny to consequence and to the drafter's reliability. Also holds.

Competence to supervise

You cannot supervise what you do not understand — which is why the failure modes must be taught.

It does not learn from correction

Where the analogy breaks. A junior corrected once improves; the tool makes the same error tomorrow.

It has no sense of its own limits

A junior says "I could not find anything on this." A model produces something anyway, with equal confidence.

Confidence is not calibrated

Fluency reads as reliability to a reviewer. With a junior, hesitancy is itself a signal; here there is none.

methodology

What the evidence shows — and what we examine.

How firms operationalise it.

Named accountable lawyerEvery piece of AI-assisted work has someone who owns it, exactly as delegated work does.
Failure-mode literacySupervisors trained on what these tools get wrong, since you cannot review for a risk you do not know exists.
Calibrated scrutinyMore review for tools and tasks with worse track records, less where the corpus is constrained.
Incident feedbackNear misses shared, because the tool will not learn from them and the firm has to instead.
what's at stake

Why the frame matters.

It converts an unfamiliar problem into one the firm already has machinery for.

clear accountability for output alignment with existing duties a frame people already understand a training need made obvious a defensible supervisory record faster adoption, less debate

Where the analogy breaks is where the danger is.

A junior who found nothing says so. A model produces something regardless, in the same confident register as when it is right. Reviewers reading for hesitancy — the signal they have relied on for a career — will not find it.

common questions

Supervision — practical questions.

Does using AI need disclosing to a supervising partner?

Most firms that have thought about it conclude yes, at least for work that will be relied on, and several now require it as a matter of internal policy. The reasoning is practical rather than ethical: a reviewer who knows a draft is AI-assisted reads it differently, and reads for different failure modes. Some courts have also adopted standing orders requiring disclosure or certification regarding AI use in filings, which varies by court and needs checking case by case.

Can a junior lawyer be responsible for an AI error?

Allocation of responsibility within a firm is a matter for the firm and for the applicable conduct rules, and generalising is unwise. What the reported decisions suggest is that courts have looked at the whole chain — who used the tool, who reviewed, who signed, and what the firm's process was. Sanctions have fallen on individual attorneys and, in some matters, on more than one lawyer in the same filing. A firm relying on the idea that responsibility sits entirely with the most junior person in the chain would be taking a position the record does not obviously support.

What does competence to supervise actually require?

Enough understanding of how the tools fail to know what to look for. That does not mean technical knowledge of how models work — it means knowing that citations can be wholly invented, that a real case can be cited for a holding it does not contain, that quoted text may appear nowhere in the source, and that none of this will be signalled by any hedging in the output. A supervisor who believes the tool "sometimes makes mistakes" in the way a search engine returns imperfect results has the wrong mental model and will review for the wrong things.

How is this different from supervising an outsourced provider?

It is closer to that than to supervising an employee, and the comparison is instructive. With an outsourced provider a firm considers what the provider does with the material, what its quality process is, what contractual protections apply, and what independent checking is warranted. Those are precisely the questions that matter for an AI vendor — and firms with mature outsourcing diligence often find they already have most of the framework, applied to a different kind of supplier.

related

Related specialization areas & resources.

Use the frame you already have.

Describe how supervision works at your firm today. The Institute will help you extend it.

AI adoption conciergeorientation · not legal or ethics advice
Tell me how delegated work is supervised at the firm now — who reviews what, and at what point. I'll help you extend that to AI-assisted work rather than inventing something parallel.