Watch a transactional attorney work for an hour and the pattern is hard to miss. The deal lives in Microsoft Word. Tracked changes carry the negotiation history. Styles and house numbering carry the firm's formatting discipline. Cross-references hold the instrument together, and defined terms carry meanings that took three drafts and two phone calls to settle. The document is not a container for text. It is the workproduct itself, layered with structure the parties have been arguing about for weeks.
This is why so many legal drafting tools stall after the pilot. They ask the attorney to leave. Copy a clause into a chat window, describe the ask, paste the answer back, fix the numbering, reconcile the defined terms, repeat. Every round trip costs minutes and introduces errors, and the attorney quietly concludes that the tool is slower than typing. Any AI that requires leaving Word loses. Not because attorneys resist change, but because the context that makes drafting work stays behind in the document.
The question worth asking is not whether AI can draft. Models can clearly produce contract language. The question is what a Word-native integration has to know, and where it has to run, before an attorney can trust it on a live deal.
Grounded in your precedent, not a generic sample
A general-purpose model will produce a plausible indemnification clause on request. Plausible is the problem. That language is an average of the public internet's contracts, which is precisely what a sophisticated client is not paying for. The clause the firm wants is the one it actually negotiated last quarter: the cap structure that survived a specific sponsor's pushback, the carve-out the litigators blessed after it was tested, the fallback the team accepts when the preferred position fails.
A serious Word integration drafts from that corpus. It retrieves candidate language from the firm's own precedent, knows where each candidate came from and how it fared, and offers it with that history attached. In Reframe's case, the Word plugin is one surface of the Context Graph, the same structured model of matters, clients, documents, and clauses that powers search and chat. The clause that appears in the task pane arrives with provenance, not just formatting.
Context the attorney never has to paste
The second thing a serious integration understands is the document in front of it. Defined terms first. If the agreement says the Company and the model inserts the Target, an associate now owns a cleanup problem that will surface at the worst possible moment. Inserted language has to be internally consistent with the instrument it lands in: the right defined terms, the right section references, numbering that continues the scheme already in place.
Matter context is the other half. The deal's parties, the governing law, the prior drafts, the counterparty's last markup: all of it already exists in the firm's systems. The attorney should never have to paste any of it anywhere. The integration should know which matter the document belongs to and bring that context to bear quietly, within the permissions the firm has set for that matter and that user. This is plumbing more than modeling: the integration sits on top of the document management system, the matter file, and the email record, and it respects the walls between them.
What attorneys actually use it for
The tasks that hold up in daily practice are concrete and repetitive:
- First-cut ancillaries. Board consents, officer certificates, assignment agreements, and closing certificates generated from deal terms already in the matter file, in the firm's own forms, ready for attorney review instead of assembly from scratch.
- Conforming changes after a pivot. When the structure moves from an asset purchase to an equity purchase, the sweep across nomenclature, mechanics, and cross-references is exactly the kind of tedious, error-prone work a grounded assistant does well under supervision.
- Fallback retrieval during a live negotiation. On the call, the attorney asks for the firm's fallback positions on a survival cap and gets the three formulations the firm has actually accepted, with the deals they came from, in seconds rather than after a hallway search for the partner who remembers.
- Markup summaries organized by issue. A counterparty's redline summarized by what it does to risk allocation, timing, and conditionality, rather than walked in redline order, so the response strategy forms first instead of last.
- Consistency checks. Defined terms used but never defined, definitions that drifted between drafts, cross-references pointing at sections that no longer exist. Mechanical, thankless, and exactly where software outperforms a tired associate at midnight.
Just as telling is what stays off the list. Novel deal structures, bespoke risk allocation for a transaction nobody has done before, the judgment call about what a client will actually sign: none of that belongs to the tool. The pattern is consistent. The machine handles language that already exists somewhere in the firm's experience, and the attorney handles what has never been negotiated before.
Inserted text should land like the attorney wrote it
How generated language arrives matters more than most vendors think. Text inserted into a draft should match the document's style, continue its numbering, and respect its defined terms, because the attorney is about to own every word of it. No banners, no colored flashes, no visual gimmicks announcing that a machine was here. If the attorney wants attribution, tracked changes already exist for exactly that purpose.
The attorney reads every word and owns every word. The tool's job is to make that reading fast, not to make its own contribution conspicuous.
The measure of a good insertion is quiet. Reviewing it should feel like reviewing your own earlier draft, not proofreading a stranger's. Anything that breaks the document's conventions converts saved drafting time into cleanup time, with interest.
The plugin is a window. The intelligence stays inside the boundary.
Drafts are privileged workproduct. A half-negotiated purchase agreement reveals the client's position, its reservation price, and its fears. So the architecture question is not a technicality. The plugin should be a thin client that talks to a service running inside the firm's security boundary, in the firm's own cloud tenant, not a public endpoint that receives contract text from thousands of unrelated users under terms nobody read.
Reframe deploys this way: the Word plugin talks to the firm's own enclave, drafts never cross the boundary, and every request is logged against the matter. Permissions matter as much as location. The plugin sees only the matters its user can see, and an ethical wall in the document management system is a wall in the task pane too. The security model deserves as much diligence as the drafting quality, because the alternative is explaining to a client why its term sheet transited someone else's infrastructure.
The attorney remains the author
None of this changes who signs the work. AI-generated text deserves the same scrutiny as an associate's first draft, which is to say real scrutiny: read it against the deal, test the mechanics, check the defined terms, and follow the sources. A well-built integration makes that discipline cheap by keeping the source clause one click away, the grounding standard we describe in our piece on hallucinated citations. The review side of the same architecture, where the machine reads the counterparty's paper instead of drafting yours, is covered in AI contract review that never leaves your tenant.
Good habits form quickly when the tool supports them. Read inserted language in context, not in the task pane. Open the cited precedent when the language is load-bearing. Treat confidence as a prompt for verification, not a substitute for it. Partners already know how to supervise drafting; the skill transfers intact.
The firms getting real value from drafting AI did not find a smarter model than everyone else. They put a grounded assistant inside the surface where attorneys already work, connected it to precedent the firm already owns, and kept the attorney in the author's chair. Word is where drafting happens. That is where the intelligence has to live.