AI Agents for Lawyers

Use Cases

AI Contract Drafting: From First Draft to Negotiated Text

AI contract drafting works from your precedents, not a blank prompt. A drafting prompt, a worked SaaS example under Austrian law and the risks a redline hides.

By Dr. Niklas SchmidtUpdated 13 November 20268 min read
On this page
  1. Start from precedent, not from a blank prompt
  2. The drafting prompt
  3. A worked example: a SaaS agreement under Austrian law
  4. Jurisdictional variants: what the prompt must name
  5. Style constraints and defined terms
  6. Two risks a redline never shows
  7. Where this leaves you

AI contract drafting produces usable text when the model works from your precedent, your deal points and a named governing law, and dangerous text when it works from a blank prompt. As of September 2026 the tools write competent first drafts and tracked changes inside Word; the two remaining risks, silent deviations from the precedent and invented statutes, never show in a redline and need their own checks.

Start from precedent, not from a blank prompt

A frontier model has read millions of contracts. Most of them were American, and the model's default draft shows it: "AS IS" disclaimers, consequential-damages exclusions, Delaware governing law and an entire-agreement clause written for a jurisdiction where it does real work. Asked for "a SaaS agreement", the model gives you the statistical average of that corpus. Asked for "our SaaS agreement, adapted for this customer, under Austrian law, with these six changes", it gives you something a partner can review.

The vendors have built their products around this. Lexis+ with Protégé added Agentic Drafting on 7 May 2026, built on the Henchman clause-bank technology LexisNexis acquired in 2024, so drafts are assembled from the firm's own precedents rather than the model's memory (LexisNexis). Thomson Reuters shipped a Word drafting agent inside the next-generation CoCounsel Legal on 20 August 2026 (Thomson Reuters). Spellbook, at about $180 a month, grounds frontier models on the document in front of you; its CEO Scott Stevenson has argued that fine-tuning "encourages hallucinations" (Nonbillable interview). DraftWise generates redlines from your own past deals, and Claude for Word writes tracked changes and is included in paid Claude plans (Vaquill's July 2026 add-in comparison). The Protégé, CoCounsel and Spellbook reviews cover the trade-offs.

If you have none of these, a Claude Project or custom GPT holding twenty of your precedents does most of the same job for bespoke agreements. The prompt library playbook covers how to organise the precedents so the model can find the right one.

The drafting prompt

The structure that works in 2026 has five parts: role and context, task, sources, output format, constraints (LegalFly, 17 Sep 2026). Applied to drafting:

You are drafting for a corporate lawyer in an Austrian law firm. Use only the attached documents.

SOURCES
- Precedent: "SaaS Agreement (Provider form) v7" (attached). This is the base text. Keep its structure, numbering, defined terms and house style.
- Term sheet: "Deal points – Customer X" (attached).
- Governing law: Austrian law. Do not import clauses that assume English or US law.

TASK
Produce a first draft of the SaaS Agreement for Customer X by adapting the precedent to the term sheet. Where the term sheet is silent, keep the precedent wording unchanged.

OUTPUT
1. The full draft, with the precedent's clause numbering preserved.
2. A deviation report: a table listing every clause whose wording differs from the precedent, the reason (term sheet item or drafting necessity), and a one-line summary of the change.
3. A defined-terms table: every capitalised term used in the draft, the clause where it is defined, and any term used but not defined.
4. A cross-reference list: every internal reference ("clause 12.3", "Schedule 2") and whether the target exists.

CONSTRAINTS
- Do not cite statutes, regulations or cases. Where the precedent references a statute, keep the reference unchanged. If you believe a statutory reference is needed, write [STATUTE CHECK: describe the point] instead.
- Do not add clauses that are not in the precedent unless the term sheet requires them; mark any such clause [NEW].
- Do not shorten, "improve" or modernise precedent wording that the term sheet does not touch.
- Write in British English. Use "shall" for obligations, consistent with the precedent.

Two things about this prompt are deliberate. The statute ban is there because models misattribute law at a measurable rate: in HAQQ's June 2026 benchmark of 3,000 answers, 24% "cited or misapplied law that didn't support claims", and Mistral Large had a 64% citation-error rate (HAQQ). The deviation report is there because the alternative is reading the whole draft against the precedent line by line, which is what the tool was supposed to save you.

One habit to drop: instructions such as "add a final verification step". Anthropic's notes on Claude Opus 5 (24 July 2026) say the model "verifies its own work without being told to" and writes longer deliverables by default (Anthropic docs), so 2025-era prompts now produce over-verified, over-long drafts. Module 2 of the AI Lab for Lawyers course runs this prompt structure on real precedents and shows how the output changes model by model.

A worked example: a SaaS agreement under Austrian law

Take a common brief: an Austrian software provider licensing a subscription platform to a German enterprise customer, Austrian law, provider's own precedent, six negotiated deal points from the term sheet. Run it through the prompt above and the draft is workable. Run it through a bare prompt ("draft a SaaS agreement between an Austrian provider and a German customer") and the first draft typically contains the following, which a red-team pass and a clean-up then have to remove:

  1. A blanket liability exclusion for gross negligence. Imported from US paper. Austrian courts strike such exclusions from standard terms under § 879 para. 3 ABGB, and a German customer's lawyers will know that § 309 no. 7 BGB does the same at home. The precedent's tiered clause (cap for slight negligence, no cap for gross negligence or intent) is what you wanted; the bare prompt discarded it.
  2. "Consequential damages" language. The English second-limb concept has no clean Austrian equivalent. The draft reads fluently and means nothing determinate in a Vienna courtroom.
  3. A misattributed data-protection obligation. The draft placed the processor obligations in a section of the Austrian Datenschutzgesetz rather than in Art. 28 GDPR, where they live. This is the classic pattern: a real statute, the wrong home for the rule.
  4. Defined-term drift. "Customer Data" in the definitions, "Client Data" in the security clause, "customer data" in the schedule. The model treats these as synonyms; a court may not.
  5. A forum clause with no form. "Exclusive jurisdiction of the courts of Vienna" appears, but the precedent's written-form language for the jurisdiction agreement, which Austrian procedure requires, has gone.

The red-team pass that catches items 1 to 5 is a second prompt: "Act as the customer's German counsel. List the ten clauses you would attack first, and for each say whether the attack is commercial or a matter of Austrian or German law." The red-teaming playbook has the full pattern. The clean-up is the deviation report plus a deterministic compare, covered below.

Jurisdictional variants: what the prompt must name

PointAustrian lawGerman lawEnglish lawUS (Delaware/New York)
Standard-terms control§§ 864a, 879 para. 3 ABGB; "grossly disadvantageous" clauses void§§ 305–310 BGB; strict even between businessesUCTA 1977 reasonableness test for standard termsLargely freedom of contract between businesses; UCC Art. 2 for goods
Gross-negligence exclusionVoid in standard termsVoid (§ 309 no. 7 BGB); cardinal duties protectedPossible in principle, subject to reasonablenessGenerally enforceable
"Consequential loss"No native concept; describe the heads of lossNo native concept; use Folgeschäden with definitionSpecific meaning (second limb of Hadley v Baxendale)Enforceable exclusion, usually mutual
Forum clauseWritten form for jurisdiction agreementsWritten form between merchants; § 38 ZPOExclusive-jurisdiction clause; Hague 2005Choice of forum clause; consent to jurisdiction
LanguageGerman or bilingual; state which prevailsGerman or bilingualEnglishEnglish

The table is not legal advice; it is the checklist for what the prompt has to say and what the reviewer has to look at. The corporate practice page covers the same problem for board papers and shareholder agreements.

Style constraints and defined terms

Ask for these with every draft, and read them before the draft itself:

  • Defined-terms table. Every capitalised term, its defining clause, and any use without a definition. Fix the definitions clause first; the body follows.
  • Cross-reference list. Every "clause", "Schedule" and "Annex" reference with its target. Insertions renumber silently.
  • Precedence clause. Which document wins on conflict: body, schedule, order form, the customer's purchasing terms. The model will not add this unless told.
  • Modal consistency. "Shall" throughout or "will" throughout. Mixed drafts read as assembled, because they were.
  • Bilingual handling. If the contract is German with an English convenience translation, say which prevails and have the model draft only the binding version.

Two risks a redline never shows

Silent deviations. The agent "improves" precedent wording that nobody asked it to touch: a shorter warranty, a reordered indemnity, a dropped carve-out. Tracked changes against the precedent would show it, but many tools produce a fresh document, not a comparison. The fix is the deviation report plus a deterministic document compare between precedent and draft. Litera's argument for its 30-year-old comparison engine, "AI is table stakes. Trust is the moat", is a marketing line, but on this point it is right: the compare tool is not probabilistic, and the agent is (Litera relaunch, 15 Jul 2026). Anything in the compare that is not in the report means the agent deviated without saying so; read the whole draft.

Invented or misattributed statutes. A model writes "pursuant to § 12 of the Austrian E-Commerce Act" with the same fluency whether or not § 12 says that. In Cork v Smith [2026] EWHC 1199 (Ch) a Pinsent Masons associate filed a document relying on an AI-invented "Insolvency Rule 12.37(5)"; the tool had flagged its own possible inaccuracy and nobody checked (UK Human Rights Blog, 4 Jun 2026). The drafting prompt above bans new statutory references for this reason. For the references the precedent already carries, open the statute: RIS for Austria, gesetze-im-internet.de for Germany, legislation.gov.uk for England. The research page covers what the grounded tools can and cannot reach.

Vals AI's finding that lawyers still beat every tool on redlining, 79.7% against 65.0% in February 2025, is the same point from the other side: the judgment about what the words should say remains yours (Vals AI). The review page covers the response to a counterparty's draft; the question page on whether AI can draft a contract gives the short answer for clients who ask.

Where this leaves you

  • Put your ten most-used precedents into a Project, a custom GPT or a drafting tool's clause bank this week, and stop drafting from blank prompts.
  • Save the drafting prompt above with your firm name, governing law and house style filled in, and use it every time.
  • Treat the deviation report and the defined-terms table as the review, and the draft as the appendix.
  • Run one red-team pass from the counterparty's side before the partner sees it.
  • Keep a deterministic compare tool in the loop for as long as the agent produces fresh documents rather than tracked changes.

Frequently asked questions

Can AI draft a contract from scratch?

It can produce a plausible document from a one-line prompt, and that is the problem. Without your precedent and the governing law it defaults to US-style boilerplate, invents or misattributes statutory references and drifts on defined terms. Drafting from an attached precedent with explicit deal points and a deviation report is the method that produces usable text.

Which AI tools are best for contract drafting in 2026?

For Word-native work the candidates as of September 2026 are Lexis Protégé Agentic Drafting, Thomson Reuters' CoCounsel Word drafting agent, Spellbook, DraftWise and Claude for Word. Frontier chat models with a Project holding your precedents work well for bespoke agreements. The differentiator is precedent access and tracked-change output, not the underlying model.

How do I stop AI changing my precedent without telling me?

Require two outputs on every run: the draft itself and a deviation report listing each clause where wording differs from the precedent, with the reason. Then run a deterministic document comparison between precedent and draft. If a change appears in the comparison but not in the report, the agent has deviated silently and the draft needs a full read.

Is AI drafting different under Austrian or German law?

Yes, mainly in standard-terms control and liability language. Austrian courts strike gross-negligence exclusions from standard terms, German law blocks them under section 309 no. 7 BGB and protects cardinal duties, and the English concept of consequential loss has no clean equivalent. The prompt must name the law and the clauses that change.

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