The assembly half of this is mature and unglamorous. The review half is where the new capability actually lands.
Start a conversation with the AI Adoption Concierge, already scoped to drafting & review. Pick a starting point, or describe your situation directly.
Document assembly in this practice predates generative AI by decades — questionnaire in, jurisdiction-specific provisions out, attorney judgment applied to the dispositive scheme and the tax elections. Firms running mature template systems already have most of that benefit and should be careful about being sold it twice. What generative models added is the other direction: reading instruments the firm did not produce. Extracting the structure of an unfamiliar trust, listing fiduciaries and successor fiduciaries, identifying provisions drafted against superseded tax thresholds, spotting internal inconsistencies across a document set, and producing a plain-language summary a client can follow in a meeting. That work was previously expensive, largely unbillable, and done under time pressure — which is a good description of where AI helps most in any practice.
Separating what document assembly already did from what models added.
Mature. Questionnaire-driven assembly with maintained jurisdiction-specific provisions is a decades-old category that AI has made more conversational rather than more capable.
The genuinely new capability. Structure, fiduciaries, distribution scheme and tax provisions extracted from a document the firm has never seen.
Provisions built around superseded thresholds, deceased fiduciaries, entities that no longer exist, references to repealed provisions.
Will, trust, powers, beneficiary designations — checking they say the same thing, which is where real-world plans most often fail.
Turning a structure into something a client can follow. Increasingly what advisors are offering, which is why attorneys should too.
Useful as a first pass. The enforceability judgment, the tax judgment and the fit to the family stay entirely with the attorney.
How to use this without creating a problem.
This practice has an unusual property: most errors surface after the client has died.
A drafting failure in litigation surfaces in weeks. A drafting failure in an estate plan surfaces when the client has died and cannot explain what was intended — often years after the limitation period on anyone's memory has quietly expired. That asymmetry argues for a heavier verification posture here than the same task would justify elsewhere.
It can produce a draft. Whether that is useful depends on what the firm already has. A practice running a maintained template system with jurisdiction-specific provisions already gets a better first draft from its own precedents than from a general model, and should be sceptical of paying again for something less reliable. Where models add value in drafting is in the bespoke provisions around the standard architecture, and in restating a structure the attorney has designed. What no model does is decide the dispositive scheme, weigh the family circumstances, or take responsibility for execution formalities.
Unevenly, and the failure is quiet. Recent native-text documents extract well. Scanned instruments from the 1980s and 1990s, documents with handwritten amendments or interlineations, and poor-quality copies produce extraction errors that look exactly like correct output. The practical control is to treat extraction as a first pass that a person checks against the instrument, and to be especially careful with names, dates and dollar figures — the fields most likely to be wrong and most likely to be relied on downstream in a client-facing summary.
That is a confidentiality question governed by the tool's terms and by your conduct rules, not by the practice area. The relevant analysis is what the tool does with inputs — whether it retains them, whether they train a model, who can access them — and whether your jurisdiction requires client consent before that happens. Estate planning does add one wrinkle worth noting: the documents commonly contain sensitive information about people who are not your client, including beneficiaries and family members who have no relationship with the firm at all.
It is happening, and the useful posture is to treat it as a starting document rather than an argument. Consumer AI produces plausible-looking instruments that frequently miss execution formalities, use provisions inconsistent with the governing jurisdiction, or fail to account for the client's actual asset structure. Reviewing one is billable work of exactly the kind the review workflow above supports. Firms that treat these clients as an insult tend to lose them; firms that treat the document as a first draft that needs professional attention tend to convert them.
Describe the intake. The Institute will help you work out what to automate first.