AI Agents for Lawyers

Use Cases

Writing Legal Memos with AI: A Method That Survives Partner Review

An AI legal memo survives partner review when the model drafts only from sources you supply, every authority is checked three ways and the checks are logged.

By Dr. Niklas SchmidtUpdated 13 November 202610 min read
On this page
  1. Why AI memos fail review
  2. The six-stage pipeline
  3. A worked example: cross-border withholding
  4. The faults to expect in a first draft
  5. The verification table the partner actually wants
  6. Partner-review annotations and the research log
  7. What to do this week

A legal memo written with AI survives partner review when the model drafts only from sources you have supplied, every authority is checked three ways (it exists, it says what the memo claims, it is still good law), and the checks are recorded in a verification table the reviewer can audit. The model supplies structure and speed; the citations remain the lawyer's job, and courts have said so.

Why AI memos fail review

The failures are well documented, because a memo that goes wrong tends to become a filing that goes wrong, and filings get judgments.

In Cork v Smith [2026] EWHC 1199 (Ch) (4 June 2026), a junior associate at Pinsent Masons asked an AI tool to research a point of insolvency law and it produced "Insolvency Rule 12.37(5)", which does not exist. The tool had flagged that its answer might be inaccurate; the associate did not check, the judge described a "cavalier attitude", and the firm was referred to the SRA and bore the extra costs (UK Human Rights Blog, 4 June 2026). In Germany, the Amtsgericht Köln (312 F 130/25, 2 July 2025) found a family-law brief citing invented BGH decisions, non-existent commentary Randziffern and fictitious journal articles, called the sources "frei erfunden" and pointed to § 43a Abs. 3 BRAO; the Kammergericht Berlin (17 WF 144/25, 20 November 2025) coined the phrase "fantasierende KI". Austria's OGH (14 Os 95/25i, 7 October 2025) dismissed an AI-drafted appeal unread because it cited OGH decisions that did not exist.

The more instructive failure is the one that looks right. The HAQQ benchmark of 5 June 2026 graded 3,000 answers across 51 practice areas and found that 24% cited or applied law that did not say what the model claimed; the best model for citation accuracy, GPT-5.5, still fabricated 3% and was the slowest and most expensive (HAQQ, June 2026). Damien Charlotin's database counted 854 misrepresented real authorities alongside 1,690 fabrications as of 14 September 2026. A real case cited for a proposition it does not support is the error a partner is least likely to catch by reading, and the hallucination sanctions page shows courts punishing it just as hard.

Firm review does not reliably catch it either. Sullivan & Cromwell apologised to Chief Judge Martin Glenn (Bankr. S.D.N.Y.) in April 2026 after an 8 April filing contained AI-generated fake citations spotted by opposing counsel at Boies Schiller; the firm said its review "did not identify the inaccurate citations generated by AI" (Bloomberg Law). The method below assumes that reading the memo is not a check; only opening the source is.

The six-stage pipeline

The pipeline separates what the model may do from what only a lawyer may do; the model touches sources only after a lawyer has found and read them.

StageWho does itWhat happensArtefact produced
1. Issue framingModel, then lawyerModel proposes an issue tree and search strings; lawyer prunes itIssue list with sub-questions
2. Source gatheringLawyerSearches in Westlaw, Lexis, RIS, juris, the firm library; downloads primary sourcesSource pack (PDFs or text) with pinpoints
3. Grounded draftingModelDrafts from the source pack only, citing by the labels you gave each sourceFirst draft with source labels
4. Adversarial passModel, then lawyerModel argues the other side from the same pack; lawyer decides what to addressCounter-arguments list
5. VerificationLawyerThree checks per authority: existence, support, treatmentVerification table
6. Partner reviewReviewerReads the memo against the table and the research logSigned-off memo

The safe-research pattern behind stages 1 to 3 is now standard advice: never ask a general model for citations; use it for issue-framing and search strings with an explicit "Do NOT provide case citations"; run the search yourself; then feed the verified authorities back with "work only with the authorities I have provided". The three checks in stage 5 come from David Kemp's analysis of Q1 2026 sanctions, which records courts describing verification as a "nondelegable duty" that AI "in no way abrogates". The legal research page explains why the pipeline is built this way; the prompt engineering guide covers the five-part prompt structure used below.

A worked example: cross-border withholding

Take an illustrative question of the kind a tax associate receives on a Friday afternoon: a German operating company pays licence fees to its US parent for software and trademarks; does German withholding tax apply, does the treaty reduce it, and is relief available at source or only by refund. No client, no matter; just the shape of the problem.

Stage 1 asks the model for structure, not law:

ROLE: Senior tax associate, German and international tax.
TASK: Build an issue tree for the question below. For each issue, list the sub-questions a memo must answer and the type of source that answers each one (domestic statute, treaty article, EU directive, administrative guidance, case law, procedural rule).
QUESTION: [paste the question]
CONSTRAINTS: Do NOT cite any statute section, treaty article, case or guidance by name or number. Do NOT state what the law is. Output only the issue tree and, for each sub-question, two search strings for a legal database.

A good tree separates the domestic charging provision from treaty characterisation (royalty versus service fee), beneficial ownership and anti-treaty-shopping rules, procedure (exemption certificate versus refund) and documentation. The lawyer prunes it, does stage 2 in the database, and labels each downloaded source S1, S2, S3 and so on with a pinpoint.

Stage 3 is the only drafting prompt:

ROLE: Senior tax associate drafting an internal research memo for partner review.
TASK: Draft the memo answering the question below, using ONLY the attached sources S1–S9.
SOURCES: S1 [domestic charging provision, section and paragraph]; S2 [treaty article]; S3 [anti-abuse provision]; S4 [administrative guidance, paragraph]; S5–S7 [case law, with pinpoints]; S8 [procedural rule]; S9 [firm precedent memo, dated].
FORMAT: Question presented; short answer (max 120 words); facts assumed; analysis by issue in the order of the attached issue tree; open points; next steps. Every legal proposition ends with a source label and pinpoint, e.g. (S2, Art. 12(2)).
CONSTRAINTS:
- Do not cite anything outside S1–S9. If the sources do not answer a sub-question, write [GAP: no source] and say what source type would be needed.
- Do not paraphrase inside quotation marks; quote exactly or do not quote.
- Where a source is older than 24 months, add [CHECK CURRENCY].
- Do not state a conclusion with more confidence than the sources support; use "likely", "arguable" or "unclear" where appropriate.

Stage 4 turns the model against its own draft. The red-teaming playbook has the full pattern; the short form is "Adopt the role of the tax authority's reviewing officer. Using only S1–S9, give the five strongest arguments against the short answer, each tied to a source label." For a withholding question this tends to surface the beneficial-ownership and substance points a first draft underweights because the question did not foreground them.

Blue J's CEO Benjamin Alarie has said a tax research memo that took 15 hours now takes 15 seconds (VentureBeat, 18 November 2025), and a Blue J/CPA.com survey of more than 1,000 US tax practitioners found 60% using AI for tax research weekly by June 2026, up from 33% a year earlier. The speed is real; the 15 seconds do not include stage 5. The tax lawyers' practice page covers the tools tax teams use; the verification-table method is the one taught in the AI Lab for Lawyers course run by Dr. Niklas Schmidt, who heads the firm-wide tax team at Wolf Theiss and is admitted as both Rechtsanwalt and Steuerberater.

The faults to expect in a first draft

Even a grounded draft has predictable faults, and each has a documented precedent.

Fault in the draftWhy it happensDocumented precedent
A rule, paragraph or Randziffer that does not existThe model completes a plausible citation patternCork v Smith (Insolvency Rule 12.37(5)); AG Köln 312 F 130/25
A real authority cited for a proposition it does not supportThe model matches topic, not holdingHAQQ: 24% of answers (June 2026); Charlotin: 854 misrepresented authorities
Correct source, wrong title or authorFormatting a citation is itself a generative stepConcord v. Anthropic (May 2025): Claude kept the link but invented the article's author and title
Another jurisdiction's rule stated as this one'sTraining data skews to US and English sourcesVals (Oct 2025): 11–14 point drop on multi-jurisdiction questions; "Place Matters" (2025): 45% hallucination on Los Angeles law, 55% London, 61% Sydney
Agreement with the premise of your questionModels defer to the user's stated beliefStanford HAI AI Index 2026: GPT-4o fell from 98.2% to 64.4% when a false premise was framed as the user's belief
A confident URL that does not resolveSearch-augmented agents cite more, so they fabricate moreUPenn (3 April 2026): 3–13% of URLs from deep-research agents fabricated

The premise problem means the associate should never write "the partner thinks the treaty rate applies; confirm" into a prompt; the collapse recorded in the 2026 AI Index is large. The URL problem means "deep research" modes do not remove stage 5; the UPenn study found agents with live web search fabricating 3–13% of the URLs they cited, more than plain models, because they cite more. OpenAI's September 2025 paper on why language models hallucinate gives the mechanism: training and benchmarks reward a confident guess over "I don't know", and a prompt that rewards [GAP: no source] works against that incentive. The deep research guide covers when research modes are worth using at all.

The verification table the partner actually wants

Stage 5 produces one row per authority. The table travels with the memo; a partner who wants to spot-check picks a row.

#Authority as citedProposition in memoExists (source, pinpoint)Supports (quoted passage or para)Treatment (KeyCite, Shepard's, juris, RIS)Checked by, dateStatus
1S2, Art. 12(2)Treaty caps withholding on royaltiesTreaty text from official gazette, Art. 12(2)"may be taxed ... but the tax so charged shall not exceed ..."In force; protocol checkedAssociate, 19 Sep 2026Verified
2S5, para 31Beneficial ownership requires economic control of the incomeCourt database, para 31Passage supports; note para 34 qualificationFollowed; no adverse treatment foundAssociate, 19 Sep 2026Verified, qualification added to memo
3S9, p. 4Exemption certificate available at sourceFirm precedent memo (2023)Supports as of its date[CHECK CURRENCY]: procedure amended sincePendingOpen

Three rules make the table honest. "Exists" means the checker opened the primary source, not a headnote, a secondary summary or another AI tool; a New Jersey lawyer was reportedly sanctioned a second time in April 2026 for using AI to verify AI, and Charlotin's tracker names self-validation as an aggravating factor. "Supports" requires the passage to be quoted or the paragraph identified, so the reviewer can read it in ten seconds. "Open" rows block sign-off; the memo goes to the partner with the open rows listed in the cover note, never silently. The verification playbook gives a longer form of the table with jurisdiction-specific treatment tools.

Partner-review annotations and the research log

The reviewer should be able to see, paragraph by paragraph, what kind of text they are reading. A simple annotation scheme, borrowed from the [verify] tag that Anthropic's open-source Claude for Legal plugins attach to any citation not checked against a connected research tool, works in Word comments or inline.

  • [V] Proposition and source verified; row number in the table.
  • [verify] Source supplied by the model or a secondary source; not yet opened in a primary source.
  • [M] Model-generated reasoning that connects sources but is not itself sourced; the reviewer is asked to agree with an argument, not a citation.
  • [GAP] No source found; decide whether to research further or flag the point as open.

The research log sits behind the table: databases searched, search strings, dates, sources downloaded and rejected, the model and tier used. It takes ten minutes to keep and it is the document a court or regulator asks for. The SRA's warning notice of 17 August 2026, which reported 42 AI-misuse reports in the year to July 2026, says solicitors must understand a tool "well enough to review its output critically"; the log is the evidence that the review happened. The drafting must also run on a tier that does not train on the memo: as of September 2026 Claude Team, Enterprise and API and ChatGPT Business, Enterprise and API do not train by default, while consumer tiers do unless opted out. The confidentiality checklist has the settings; a Claude Project per matter, as in the matter-project playbook, keeps the source pack, the issue tree and the drafts together.

What to do this week

Take one memo you have already written and reviewed. Rebuild it through the six stages using its own sources as the source pack, compare the grounded draft against the memo that went out, then run the adversarial pass on the original. Keep the verification table from that exercise as the template for the next real memo.

Then set the rule for the team, in writing: the model never supplies an authority; every authority gets three checks; no memo leaves without its table; open rows are disclosed in the cover note. Judge Manasco's remark in Johnson v. Dunn (N.D. Ala., 23 July 2025) that "if fines and public embarrassment were effective deterrents, there would not be so many cases to cite" is the argument for a rule rather than a reminder; Butler Snow had an AI policy from June 2023 and three of its lawyers were still reprimanded and removed from the case. A policy says what should happen; the table shows what did.

Frequently asked questions

Can I use ChatGPT or Claude to write a legal memo?

Yes, if the model drafts from authorities you have found and read, and you check every citation before the memo goes anywhere. The Divisional Court in Ayinde v Haringey said freely available generative AI tools are not capable of conducting reliable legal research. Used as a drafting and structuring engine over verified sources, on a business tier that does not train on your data, the same tools are safe and fast.

How do I stop AI from making up cases in a memo?

Do not ask it for cases. Use the model to frame issues and write search strings, run the search in Westlaw, Lexis, RIS, juris or your firm's library, then paste the verified authorities into the prompt with the instruction to cite only those. Anything the model adds on its own is tagged for verification and either confirmed in a primary source or deleted. A research log records what was checked and by whom.

What is a citation verification table?

A table with one row per authority in the memo: the citation as written, the proposition it supports, whether it exists in a primary source with a pinpoint, whether that passage actually supports the proposition, its subsequent treatment, who checked it and when. It turns the partner's question from 'did you check' into 'show me row 7', and it is the document that saves a lawyer when a court asks the same question.

Are AI memos worse for cross-border or tax questions?

Measurably. In Vals AI's October 2025 legal research study every system, including those that beat the lawyer baseline overall, dropped 11 to 14 points on multi-jurisdiction questions, and an earlier academic study found the same models hallucinating far more on London and Sydney law than on Los Angeles law. For a withholding-tax or treaty question, supply the domestic provision, the treaty article and the procedural rule yourself rather than asking the model to know them.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion