Contract work almost always ends up in Microsoft Word, and that is where most of the effort accumulates: inserting clauses, repairing numbering, restoring formatting, reconciling versions. Legal AI therefore only pays off when it operates inside the document rather than beside it. At CASUS, around 80 percent of usage happens in the Word add-in, not in the web app (as of August 2026).
This article describes three workflows that have established themselves in law firms and in-house teams: Benchmark, Proofread and Chat Actions. The focus is on what each one delivers, where it pays off, and what makes it fail in practice.
Why the place of work determines the benefit
Time in contract work is rarely lost during legal assessment. It is lost in the steps before and after. A tool proposes wording, the wording is copied into the document, the numbering shifts, the style breaks, cross-references point nowhere. The result is a second round of work that partly cancels out the first.
A legal AI workflow in Word therefore has to meet three requirements. The change must land in the right place in the document, not in a separate response. The document structure, meaning numbering, styles and references, has to survive the change. And the result must be traceable: anyone amending a clause needs to see what the finding rests on and where in the document it applies.
The share of roughly 80 percent of usage in the add-in is the practical evidence. Teams that have both a web app and an add-in available overwhelmingly choose the tool that runs where the document already is. The underlying argument is set out in the article on legal AI inside Word; this one is about the three concrete workflows.
Workflow 1: Benchmark against an internal standard
Benchmark checks a document against a reference standard, either an internal playbook or recognised best practice for SPAs, NDAs and data processing agreements. The result shows whether the expected clause topics are present and sufficiently developed, and names gaps such as a liability clause without a cap, undefined IP ownership, or a confidentiality clause without a deletion obligation. Alignment is expressed as a percentage.
The value lies less in the list than in what can follow from it: missing clauses can be inserted at the appropriate place, carrying the document's formatting and numbering, either as text or as markup.
Where benchmarking pays off fastest
The comparison works best for contract types that recur frequently and where a team holds clear preferences. NDAs, data processing agreements and framework agreements are the typical candidates, because a pattern exists there against which deviations are measurable at all.
The second effect concerns division of labour. When several people review the same contract type, standards drift apart without a comparison: one person falls back on an older template clause, the next on a slightly different version. A reference standard makes that drift visible before it settles into the contract portfolio. For new team members it also replaces the unwritten knowledge of what counts as customary in the firm.
What a usable result has to deliver
A comparison is only useful once it prioritises. Thirty deviations without weighting create the same work as a full read-through. What matters is the distinction between a missing liability cap and a stylistic deviation, because only the first carries negotiating weight.
The result is also useful internally: a documented alignment figure with reasoned deviations holds up better before clients and management than a general assessment.
Workflow 2: Proofread for consistency and references
Proofread is the quality and consistency check of the document before it goes out, not a legal review of the underlying law. It covers spelling, grammar and style without touching legal meaning, along with Swiss spelling conventions, consistent terminology and party designations.
The expensive errors in contracts are rarely orthographic. They are definitions used differently in two places, cross-references to a clause that no longer exists, numbering that stops adding up after a section was moved, open placeholders, and deadlines that contradict each other across the document. Such defects generate queries and, in the worse case, the interpretive latitude the contract was meant to close off.
Proofread accordingly checks cross-references, definitions, annexes that are mentioned but missing, numbering, heading hierarchy, conflicting deadlines and open placeholders. Because the correction happens inside the document, the act of inserting it does not create a fresh inconsistency.
The return is highest on long documents, many versions and tight time before signing, which is precisely where the final read-through tends to become superficial.
Workflow 3: Chat Actions that act on the document
The difference between a chat that answers and one that acts determines the time saved. A proposed clause is not yet an inserted clause. Agent Mode in AI Chat carries the change out in the document: inserting, rephrasing and extending clauses while respecting structure, numbering and formatting, and placing them correctly. It also checks consistency across the whole document and points to the passages affected.
Two additions to the instruction are worth making: where the clause should sit and how it should lean, for example neutral or narrowly drawn. The legal decision stays with the reviewer; the implementation work falls away.
The effect is most pronounced with recurring standard building blocks that are needed across many documents but fitted in from scratch every time.
Anyone wanting to test the three workflows is best served by setting CASUS up directly in the Word add-in and running it against a live matter. An account can be opened for free; the first 14 days cost nothing and carry no minimum term. What such a test yields depends heavily on the document chosen.
What makes adoption fail in practice
The most common reason an onboarding stalls is neither an acceptance problem nor a training problem. It is a test setup problem: the evaluation runs on a template document or an old NDA instead of a real matter from current work. Real matters are exactly what the tool is meant to help with, and that is where the difference shows. A clean template has no organically grown numbering, no conflicting deadlines, and no three rounds of negotiation behind it.
The second recurring observation concerns the order of questions. In initial conversations with Swiss law firms, the questions turn to data protection and security first, not to functionality. One of them comes up almost every time: whether documents have to be anonymised before processing. They do not. CASUS hosts in Switzerland, content is not retained after processing and not used for training, and there is no human review. Details are set out under security and data residency. Anonymize exists as a separate action in the add-in, but for sharing documents with third parties, not as a precondition of use.
In practical terms: a pilot on one real contract with one of the three workflows says more than a broad evaluation run on template documents.
FAQ
How secure is the processing of client documents?
CASUS hosts in the Google Cloud Zurich region, encrypts data at rest with AES-256 and in transit with TLS 1.2 or higher. AI inference runs through Google Vertex AI in Belgium and Microsoft Azure OpenAI in Switzerland North and Sweden Central, not in the United States. Customer data is used for training neither by CASUS nor by the model providers. CASUS staff are auxiliary persons of the lawyer under Art. 321 StGB.
Do documents have to be anonymised before processing?
No. With hosting in Switzerland, zero data retention and human review excluded, prior anonymisation is not a precondition of use. The Anonymize action in the add-in serves the sharing of documents with third parties.
How much effort does installing the add-in take?
The add-in is installed in Microsoft Word. The larger part of adoption lies not in installation anyway, but in deciding which documents and which workflows to move over first.
Does legal AI replace legal review?
No. Benchmark and Risk Review produce findings and drafting options, not legal advice. Proofread checks the consistency and form of the document, not the legal position. Assessment and responsibility remain with the lawyer.
Which teams benefit first?
Teams with high document volume, recurring contract types and existing standards. Small and mid-sized firms as well as in-house teams with heavy contract loads see returns faster than teams handling mostly one-off matters without templates.
How long does a Risk Review take?
A risk analysis of a contract takes one to three minutes and produces around twenty findings on average (as of August 2026), each assigned to a contracting party with relevance and severity.
What distinguishes Benchmark from Risk Review?
Benchmark measures a document against a defined reference standard and shows deviations from it. Risk Review analyses the risks of a contract from a party perspective, regardless of whether a standard exists. Teams with a playbook use both in sequence.
