Check everything to the same depth and you give back the time you gained. Check nothing and you end up in a reported decision. The design question is where the line sits.
Start a conversation with the AI Adoption Concierge, already scoped to review workflow. Pick a starting point, or describe your situation directly.
Firms that make AI adoption durable tend to have worked out that not all output carries the same risk. A brief for filing, a client advice letter, an internal research memo, a first-pass document summary and a draft email are five different exposures, and reviewing them identically is neither necessary nor sustainable. A workable review design sorts work by where it is going and who will rely on it, sets a proportionate check for each tier, and — critically — makes the tier obvious at the moment the work is created rather than requiring a judgement call later, when the person making it is in a hurry.
Consequence, not effort. A short filing can carry far more risk than a long internal memo.
Court filings carry candour obligations and a public record. The highest tier, without exception.
Advice a client will act on, whether or not it is formally an opinion.
Valuations, calculations and deal terms, where an error propagates into the transaction.
Which changes the tool that may be used at all, not just the review.
Working product nobody outside will rely on — the tier where lighter checking is defensible.
Precedent and template material warrants the higher tier, because errors propagate silently.
How the tiers are made to hold.
The point is a process people follow under time pressure, which is the only condition that matters.
A review process that works in a normal month and quietly gets skipped the night before a deadline is not a process. The tier that covers filings has to be light enough to survive a hurry.
Two or three. Something like: anything going to a court or a client gets full verification by a reviewing lawyer; internal working product gets a lighter check by the drafter; and material that will be reused as precedent goes in the top tier regardless of where it started. Schemes with more granularity look rigorous on paper and collapse in practice because nobody can recall which tier applies to what, so people default to whichever they remember — usually the lightest.
It depends on the circumstances and the jurisdiction, and it is worth taking specific advice rather than adopting a general rule. ABA Formal Opinion 512 addresses communication and recommends disclosure in engagement letters in some circumstances — particularly where AI use is relevant to the basis of the fee, or where confidential client information is processed by the tool. Some clients, especially institutional ones, now impose their own requirements through outside counsel guidelines, which is frequently the more immediate constraint.
Yes, and this is under-appreciated. A retrieval-based legal research platform that cites from an actual corpus presents a different risk profile from open-ended chat, which presents a different profile again from a document-summarisation tool working over material you supplied. The proposition check remains necessary across all of them, but the likelihood of wholly invented authority varies enormously. A review policy that treats every tool identically will be either too heavy for the good ones or too light for the rest.
Mostly by making it short enough to survive the pressure, and by putting it where the work already has to pass. Review steps that sit inside an existing gate — the filing checklist, the partner sign-off that already happens — get done. Steps that require someone to remember an additional obligation at 11pm do not. It also helps considerably if the firm has been explicit that missing the check is a bigger problem than missing the deadline, because otherwise people will make the opposite trade.
Describe how work moves through the firm. The Institute will help you tier it.