Contract AI: What Actually Works (and What Doesn't)

Contract review is one of legal AI's most mature applications. The technology has been in commercial use for nearly a decade. Yet most of the coverage focuses on firms reviewing thousands of contracts a year — due diligence teams at large M&A houses, banks with massive supplier portfolios, corporate legal departments processing hundreds of NDAs a week.

If you run a 10-15 lawyer commercial firm in Switzerland or Germany, that is not your world. You do 80 contract reviews a year, maybe 120 in a good year. The ROI argument changes completely at that volume.

The right question is not "can I process 5,000 documents?" It is "can I reduce each review from two hours to 45 minutes?"

Start with NDAs

If you want to pilot contract AI, start with NDAs. They are the highest-volume, most standardised document type in any commercial practice. Every client relationship generates them. They follow predictable structures. And — critically — the consequences of a missed clause, while never trivial, are manageable enough that an error during a pilot will not destroy a client relationship.

NDAs give you the best environment to test whether a tool actually works on your documents, in your jurisdiction, before you trust it on a supply agreement or employment contract where the stakes are higher.

What Contract AI Does Well

Clause extraction is where these tools genuinely perform. Standard NDA provisions — confidentiality scope, term and termination, governing law, limitation of liability — can be identified and extracted with high accuracy on well-formatted documents. The same applies to commercial contracts: payment terms, renewal clauses, change of control provisions.

Deviation flagging is the natural extension. Once you define your preferred playbook — the clause language your firm considers market-standard under Swiss or German law — the AI can flag deviations. At 80 reviews a year, this does not replace a lawyer. It means the lawyer spends 20 minutes reviewing what the AI flagged rather than 90 minutes reading from the beginning.

Metadata extraction — parties, dates, contract values, notice periods — is now reliable enough to feed directly into a contract management spreadsheet or DMS. The time saving on data entry justifies part of the tool cost on its own.

What Contract AI Struggles With

Civil law drafting conventions are a persistent gap. Most AI contract tools were trained primarily on English-law and US-law agreements. A tool with strong performance on New York-governed supply agreements will perform materially worse on Swiss Code of Obligations contracts or agreements drafted under German HGB conventions. The syntactic markers differ. The risk allocation logic differs. Swiss and German drafting tends to be less clause-heavy and more integrated into the body text — which confuses extraction tools trained on Anglo-American modular drafting.

This matters practically: always test any tool against a sample of your own documents, from your own jurisdiction, before signing a subscription.

Novel clause drafting remains a human task. AI tools can identify that a clause deviates from your playbook. They cannot reliably draft replacement language that is both legally sound and commercially appropriate for the specific transaction context. The drafting judgment stays with you.

Context-dependent risk assessment requires professional judgment that current tools do not provide. Whether a limitation of liability clause is acceptable depends on the commercial relationship, the client's risk appetite, and the counterparty's financial position — none of which sit inside the document.

Swiss and German-Law-Compatible Tools

Beyond the established Anglo-American platforms (Kira, Luminance, Ironclad), there are now several tools built with the Swiss and DACH market specifically in mind:

CASUS (Zurich) is an AI associate for contract drafting and review with a Word add-in and web app, hosted on Swiss servers under Swiss data protection rules. Purpose-built for law firms and in-house legal teams.

Legartis (Zurich) offers AI contract review with 160+ pre-trained clause types across German, English, and French. It targets corporate legal departments primarily but is relevant for firms doing high-volume standard contract work. Claims 90%+ accuracy on trained clause types. Swiss-hosted.

DeepLaw (Switzerland) focuses more on legal research than contract review, but its conversational access to Swiss federal and cantonal legislation is useful alongside any contract review workflow when you need to verify a statutory reference quickly.

Swiss-Noxtua (Basel/Berlin) — the most ambitious Swiss-native platform — combines legal research, document analysis, and drafting in a single workspace with exclusive access to Basler Kommentar and Commentaire romand content. Pre-launch as of early 2026, but worth watching.

For German-law work, stp.one remains a dominant practice management platform for DACH firms, and international tools like Luminance have invested meaningfully in German-language training data.

The key question for any vendor: ask specifically for accuracy benchmarks on Swiss Code of Obligations or German HGB contract corpora — not aggregate performance figures.

A Practical First Four Weeks

If you have never run a contract AI pilot, here is a workable sequence:

Week 1: Pick one document type and gather examples. Choose NDAs. Pull 20 examples from your own matter files — a mix of your own drafts and counterparty forms, short and long, German and English if applicable. These are your test documents.

Week 2: Run a free trial on your own documents. Most tools offer 14-30 day trials. Upload your 20 NDAs. Do not read the vendor's demo materials first — go in cold, run the tool, and see what it produces.

Week 3: Compare AI output to your own review. Take five of those NDAs and review them yourself the way you normally would. Then compare what you found against what the AI flagged. Look specifically for what the AI missed — not what it caught.

Week 4: Make a decision. If the AI found 85%+ of what you found, and the misses are consistently in a predictable category you can check manually, the tool is probably useful. If the misses are random, the tool is not reliable enough for your practice.

Professional Indemnity — Check Before You Deploy

Before you use any AI-assisted review on a live client matter, check with your professional indemnity insurer. Most policies have not been updated to address AI use explicitly, and the question of whether AI-assisted review satisfies your duty of care — or whether a missed AI output shifts liability — is live in most European jurisdictions.

The SAV (Swiss Bar Association) AI Guidelines adopted in June 2024 establish a three-pathway compliance framework for using AI with client data. Under Swiss professional rules, you remain personally liable for AI error regardless of what the tool vendor's contract says. That liability does not disappear because the software missed something.

Document your process. If the AI flagged a clause and you reviewed and accepted it, log that. If the AI missed something and your manual review caught it, that is the system working as designed — but the log should show the human review happened.

Choosing a Tool

Four questions before signing:

1. What jurisdiction is the training data? If you practise under Swiss or German law, verify that the vendor has substantial training data from those jurisdictions — not translated US or UK contract corpora.

2. How customisable is the playbook? The best tools let you define your own clause library and your own deviation thresholds. Tools that impose the vendor's definition of "market standard" create problems for a boutique firm with specific client preferences.

3. What is the integration with your DMS? A tool that lives outside your document management workflow will be ignored within three months.

4. What is the audit trail? You need to demonstrate that a qualified lawyer reviewed and approved the AI's output. The tool should log every AI finding and every human decision. This is not optional — it is how you demonstrate professional diligence if a matter goes wrong.

The firms that succeed with contract AI treat it as a process change, not a technology deployment. The technology is the straightforward part.

Build Your Firm's Clause Library First

The biggest mistake small firms make is treating every contract as a new document. The recommended approach: digitise your firm's preferred positions into a clause library — five best examples per common section (indemnification, governing law, force majeure, termination). Store these in a structured format. Then use AI to audit incoming contracts against your library rather than against the vendor's generic playbook.

A solo practitioner running a 50-page contract through this pipeline can get a fully redlined version with suggestions and margin comments in 15 minutes. For a Basel firm handling OR-based commercial contracts, building a Swiss-specific clause library (Gewählter Gerichtsstand, Haftungsbeschränkung, Vertragsstrafe) once and reusing it across every incoming contract delivers immediate ROI.

Where AI Gets Specific Things Wrong

In live cross-model benchmarks, AI reliably catches non-solicitation clauses that don't belong in standard NDAs and asymmetric obligations in "mutual" NDAs. But it consistently misses the distinction between negligence and gross negligence in indemnification clauses — a critical legal nuance where the risk exposure differs substantially. Under Swiss contract law (OR Art. 100 Abs. 1), the distinction between grobfahrlässig and leichtfahrlässig determines whether a liability exclusion is even enforceable. This is exactly the kind of check a reviewing Anwalt must perform manually.

The Claude Legal Plugin runs 13 checks on an NDA in approximately 20 seconds. But it has six production gaps that matter for Swiss practice: no document upload (copy-paste only), no memory across sessions, no audit trail, no Word export with tracked changes, no hallucination prevention beyond a disclaimer, and no severity-level playbook. These are the gaps between a free tool and an enterprise-ready system — and understanding them is essential before advising clients on tool selection.

If you are evaluating contract AI for your firm or managing a deployment that has stalled, get in touch. I work with law firms at every stage — from tool selection through to workflow design and change management.