“Can AI write my will?” is a query your future clients type before they type your name. The answer they get is a document that looks right, which they print and sign at the kitchen table, sometimes with a neighbour as the only witness, sometimes with no witness at all. Then they die, and the neighbour, the adult child who typed the prompt and the chatbot transcript all arrive in probate together.
That scene, described on a California estate firm’s blog, is the case for AI for estate planning lawyers as much as the case against consumer AI wills. The same tools that produce the kitchen-table disaster can, inside a practice with a clause library and a verification habit, take real hours out of drafting, review and client explanation. The difference is who checks and who bills.
What AI for estate planning lawyers does well: clause libraries, review flags, calendars
The vendor list from Spellbook’s estate-planning page is a fair map of the lawyer-side work, as long as you read its numbers as marketing: drafting wills, trusts, powers of attorney and advance directives from the firm’s clause libraries; review that flags “conflicting beneficiary designations, ambiguous distribution language, missing clauses”; probate calendars and creditor-notice tracking; trust accountings; plain-language client explanations. One attorney is quoted “saving up to two hours per day”.
ACTEC adds intake, risk evaluation, asset valuation and tax planning, and answers its own safety question: “verification, data security, and human oversight still matter.”
| Task | Risk | Tool tier | What the lawyer still does |
|---|---|---|---|
| Intake notes to instruction sheet | Medium | No-training tier, anonymised | Confirms the family tree; records capacity observations personally |
| Drafting from the firm’s clause library | Medium | Claude Project or custom GPT with the clause bank attached | Reads every dispositive clause; checks defined terms |
| Review of a draft trust for conflicts and gaps | Medium | Same | Opens every flagged clause; decides gap or choice |
| Probate calendar and creditor notices | Low | Any no-training tier | Verifies each deadline against the statute |
| Plain-language client letter | Low | Any tier, no client identifiers | Diffs the rewrite for lost obligations |
Start with intake, where the model’s contribution is completeness rather than drafting:
From the (consented and anonymised) meeting notes below, produce an instruction sheet for a will and trust package under [jurisdiction] law.
Output: a family tree with roles (executor, trustee, guardian, beneficiaries) as a table; assets by category and ownership; specific gifts; the residuary scheme; contingencies the client has not addressed (predecease, simultaneous death, minor beneficiaries, digital assets, pets); tax questions to raise, each tagged [VERIFY]; and blank fields headed "capacity and undue-influence observations" for me to complete.
Then list the questions for the follow-up call. Do not draft dispositive clauses. Do not invent any fact not in the notes.
Notes:
[paste]The plain-language letter is the safest use of all. Ask for a rewrite of your standard explanation at a level a lay reader manages on a phone, then a two-column table of original against rewritten sentence so you can see what was lost; the client communication guide has the prompt.
The consumer wave: “ChatGPT, create my estate plan”
Clients are not waiting. Clio’s 2025 consumer research found 28% of people were told by their AI to contact a lawyer and 12% were talked out of a claim; a Washington solo told the same Legal Trends Report: “They’d rather have ChatGPT do it and have a lawyer review it.”
A North Carolina estate firm’s warning lists what the review finds: special-needs, Medicaid spend-down and digital-asset issues missed entirely, recent law changes the model never saw, and “A will that looks correct may end up being unenforceable in your state.”
Why AI wills fail: two witnesses present at the same time
The AI will usually fails on execution, not prose. Ridley Law’s analysis of the California position starts with Probate Code section 6110, which requires two witnesses present at the same time. People “print an AI-drafted document and sign it at the kitchen table, sometimes with a neighbor as the only witness, sometimes with no witness at all.” The harmless-error route exists but demands clear and convincing evidence, which means litigation.
Content fails too: Ridley notes the community-versus-separate-property analysis the tool “routinely gets wrong”.
Interested-witness and drafter-beneficiary presumptions
Two further California traps follow. Section 6112 raises a presumption against a gift to an interested witness, the neighbour or relative who did sign. Section 21380 raises a presumption of fraud or undue influence against a gift to the person who drafted or transcribed the instrument. Ridley’s example is the one estate lawyers now see: “an adult child prompting ChatGPT to draft a trust amendment favoring themselves” and putting it in front of the parent to sign. The transcript shows who typed what.
The transcript in the will contest
A will contest is an argument about what the testator intended, and the chatbot conversation is that intention in the testator’s own words, timestamped. Ridley puts it plainly: “They can read your own words, typed to a chatbot, laying out exactly what you were thinking and why.”
Two 2026 facts make that transcript reachable. In United States v. Heppner (S.D.N.Y., bench ruling 10 February 2026) Judge Rakoff held that a person’s own consumer-Claude exchanges were protected by neither privilege nor work product: “Because Claude is not an attorney, that alone disposes of Heppner’s claim of privilege” (Orrick’s summary). And under the preservation order in the New York Times litigation OpenAI kept every deleted consumer chat from 13 May to 26 September 2025. “Deleted” is a setting, not a shredder.
Second-opinion pasting and Evidence Code 912 waiver
The client-side habit that most directly hurts your work is the second opinion: the client receives your letter of advice, pastes it into a chatbot and asks whether the lawyer is right. Ridley flags the consequence under California Evidence Code section 912: disclosing a significant part of a privileged communication to a third party can waive the privilege, and a consumer chatbot is a third party whose terms permit access. The UK Upper Tribunal said the same of ChatGPT in 2025, holding that pasting client letters into an open tool places them “in the public domain”. The privilege guide has the case law.
ABA Formal Opinion 512 says boilerplate engagement-letter clauses about AI are “not sufficient” for informed consent, and a warning buried in the terms is no better. When the second opinion arrives anyway, answer it properly:
A client has sent me the AI-generated analysis below, which conflicts with my advice on their estate plan under [jurisdiction] law.
Draft a reply that: thanks them; identifies where the AI text goes wrong (wrong jurisdiction, outdated law, invented authority, missing fact), one sentence each; explains in plain terms why our advice stands; and warns, in one sentence and without lecturing, that pasting our letters into public AI tools can waive privilege and become evidence in a later contest. Warm, under 200 words. Cite nothing you cannot confirm from my advice.
My advice: [paste]
Client's AI text: [paste]Reformation costs “tens of thousands”
Can the mistake be fixed after death? Sometimes. Estate of Duke (2015) 61 Cal.4th 871 allows reformation of a will to match the testator’s actual intent, but only on clear and convincing evidence. Ridley’s estimate of the cost of getting there is “tens of thousands” in fees, the number to put in front of the client who wants to save a few hundred by prompting. The saving becomes the estate’s litigation budget, and the transcript is the main exhibit on both sides.
Billing when three hours become one: the NC hypothetical
The one ethics opinion written for estate planners is North Carolina’s 2024 Formal Ethics Opinion 1. Inquiry 6: a lawyer with an estate planning practice bills $300 an hour, previously spent about three hours drafting standard estate documents, and with AI now spends one. May the lawyer bill three hours?
“No, Lawyer may not bill a client for three hours of work when only one hour of work was actually experienced.” — North Carolina State Bar, 2024 Formal Ethics Opinion 1
The opinion adds that the lawyer “may enjoy the benefit of those new efficiencies by completing more work for more clients”, and that flat fees are permitted “provided the flat fee charged is not clearly excessive and the client consents”. ABA Formal Opinion 512 is consistent: bill actual time, do not bill for learning the tool, and a flat fee “may be unreasonable under Rule 1.5” if the tool makes the work much faster.
Help me price a [simple will and power of attorney package] as a flat fee. Here is our time data for the last [20] such matters: [paste hours, rate, outcome].
Compute the mean, median and 80th percentile of hours and cost; identify the three drivers of the outliers; propose a scope with explicit exclusions (contested capacity, blended families, business interests, foreign assets); propose a flat fee at [target margin] with an add-on schedule; and draft a two-paragraph client-facing scope description. Show the arithmetic in a table.Check the arithmetic; the wider pricing debate is in AI and the billable hour.
DAV: wills are a permitted AI drafting use, with verification
For German-speaking practitioners the position is explicit. The Deutscher Anwaltverein’s Stellungnahme 32/2025 lists drafting legal texts, contracts and wills among them, as a permitted use of AI, subject to verification under § 43 BRAO, with complete anonymisation or pseudonymisation for publicly accessible tools. The BRAK is stricter about what may go into a public model at all; the differences are in KI in der Kanzlei: BRAK, DAV and ÖRAK. The practical rule is the same everywhere: the clause library is yours, the model assembles, and a lawyer reads every clause that moves money.
Private-client advisers and STEP members have attended AI Lab for Lawyers, where workflows like the intake summary above are built hands-on, in the browser tools you already have, rather than described.
Where to go next: the practice-area hub covers neighbouring fields, AI for family lawyers deals with the same chatbot-using client in a divorce, and AI contract drafting: what goes wrong explains why long dispositive documents defeat a single prompt. The prompts above are in the prompt library.
Frequently asked questions
Can AI write a valid will?
It can produce a document that reads like a will. Validity depends on execution formalities the chatbot cannot supply: in California, Probate Code section 6110 requires two witnesses present at the same time, and estate lawyers report clients signing AI drafts at the kitchen table with a neighbour as the only witness or none at all. Add interested-witness and drafter-beneficiary presumptions and a discoverable chatbot transcript, and the saving evaporates in the contest.
Why do AI-drafted wills fail?
Four recurring reasons: defective execution (wrong number or timing of witnesses); substantive errors such as community versus separate property analysis the tool routinely gets wrong; missed issues such as special-needs beneficiaries, Medicaid spend-down and digital assets; and statutory presumptions against gifts to the person who arranged the drafting. A North Carolina estate firm's summary is accurate: a will that looks correct may be unenforceable in your state.
Can estate planning lawyers use AI to draft trusts?
Yes, from the firm's own clause library in a no-training tier, with the lawyer reviewing every dispositive clause. Vendors report attorneys saving up to two hours a day on drafting, review flags and probate calendars, and Germany's DAV lists drafting wills among permitted AI uses provided the output is verified. ACTEC's position is the right one: verification, data security and human oversight still matter.
Is a client's ChatGPT conversation about their will discoverable?
Treat it as discoverable. In United States v. Heppner (S.D.N.Y., February 2026) a court held that a person's own consumer-Claude exchanges were protected by neither privilege nor work product, and OpenAI kept every deleted consumer chat for more than four months under a 2025 preservation order. In a will contest, contestants can read the client's own words, typed to a chatbot, explaining exactly what they were thinking and why.
How should I bill when AI speeds up estate documents?
Bill the time actually spent, or move to a flat fee the client has agreed to. North Carolina's 2024 FEO 1 uses precisely this case: a $300-an-hour estate planner whose three-hour drafting job now takes one hour 'may not bill a client for three hours of work when only one hour of work was actually experienced'. Flat fees are permitted if not clearly excessive and the client consents; Florida and California treat general AI subscriptions as overhead.