As of October 2026, the question for a managing partner is no longer whether general-purpose AI vendors will sell into legal work. LawNext reported on October 15 that Microsoft launched Copilot for Legal. It offers more than 20 legal-specific skills in Microsoft Copilot, bringing contract review, drafting and obligations tracking into Word and Cowork, along with plugins from more than two dozen legal technology partners. LawNext counts Microsoft as the fourth major AI company in five months to ship a dedicated legal offering, after Anthropic's Claude for Legal in May, Google's Gemini Enterprise for Legal in August and OpenAI's Astra for Law last month.
The more useful question for a firm is this. When legal work runs through skills and connectors inside a platform, what decides whether the firm's published expertise, entity details and reputation signals are found, cited and described accurately? And how would the firm find out? This article separates what the sources show from what is still hypothesis, and ends with a test a firm can run itself.
What the reporting establishes, and what it does not
The LawNext piece is news reporting, not independent proof of anything about performance. Its observation is that all four launches follow essentially the same model: a set of legal-specific skills or agents, a roster of connectors into the specialist products lawyers already use, and a promise that lawyers can do more work without leaving the platform. LawNext adds that Microsoft's distinguishing factor is that lawyers already work in Word, so the vendor only has to introduce a new capability inside a familiar application.
The excerpt does not tell us adoption rates, accuracy, or how any of these products sources or weights outside information about a law firm. It says nothing about whether a firm's public presence is helped or hurt. Everything below about firm visibility is therefore inference, and we label it as such.
Connectors move the point of citation
A second LawNext report, published October 19, gives a concrete architecture example. Legora customers with a Blue J subscription can now query Blue J from within Legora and receive answers with inline citations linking back to sources in Blue J. The integration is built on the Model Context Protocol (MCP), and Blue J performs the research on its own infrastructure. The companies say their shared customers asked for it. Blue J's CEO and cofounder, Benjamin Alarie, is quoted saying tax professionals can get cited analysis where they already work, check the source materials themselves and keep their judgment at the center of their advice. That is a company statement, not an independent finding.
Reading that design, three consequences follow. These are our reading of the architecture, not claims the source makes.
- The host platform sees what the connector returns. If research runs on the provider's infrastructure, the host receives an answer and its citations, not the provider's full index.
- Provenance travels as citations. A claim that has no stable, checkable address is harder to cite than one that does.
- Verification happens at the citation. The lawyer's check is whether the linked source supports the sentence, so the quality of the underlying source matters as much as the model.
Hypothesis, not finding: as research moves into connected platforms, a firm's public material may be one input among several, and its usefulness may depend on whether its claims are specific, dated and attributable. We have no first-party data showing this for any firm. It is also worth remembering that a citation is not proof of accuracy. We wrote earlier about three sites that built 215,128 fake "best software" pages that Perplexity cited, which is a reminder that an answer engine pointing to a source says little about whether that source deserves trust.
Test the agent as deployed, not a stand-in
A Microsoft Research blog post from October 7 introduces Agent Lightning v1.0, a reinforcement learning framework from Microsoft Research Asia. This is vendor research, so treat it as the authors' framing. They argue that agent capability increasingly depends on the harness that coordinates the model, meaning the tools and execution environment around it. They also say most agent training systems require developers to reimplement the agent inside the training framework, so the agent being trained is not quite the agent that gets deployed.
Most firms will never train a model. The principle still carries over as an inference of ours: evaluate through the same harness, tools, connectors and settings that production uses. A partner who pastes a question into a consumer chat window is testing a different system from an associate who invokes a skill inside Word with connectors enabled. If you want to know what an agent says about your firm, test the surface your referral sources and prospects actually use, and record which one it was.
A second pattern comes from a preprint, and a note on its limits is needed. An arXiv abstract dated October 16 describes SEER-Repro, a benchmark of 20 published papers based on a U.S. cancer registry. The authors report that their agent reproduces 77.4% of 2,190 atomic claims. They also report that an agentic pipeline flagged potentially consequential analytical choices in 9.5% of fully reproduced findings for expert review. These are author claims. We cannot confirm peer review, the domain is medical rather than legal, and the numbers say nothing about legal AI or about any firm.
What transfers is the method, not the numbers: break an output into atomic claims, check each against a source, and send the flagged ones to a human expert. That is a workable structure for auditing how an AI system describes a firm.
Cost is tokens per task, not the price list
Repeated testing costs money, so the budget matters. In a post on the llm-gemini plugin, Simon Willison reports that Gemini 4 Argon launched at $2 per million input tokens and $10 per million output tokens, with a price increase scheduled for December 1 that would match Claude Opus 5.5. Willison's point is that such prices are not directly comparable, because what matters is how many tokens a model needs to achieve a goal within an agentic harness. That is one practitioner's observation, and the excerpt we have does not include the benchmark cost data he goes on to cite.
For a visibility test, the practical takeaway is to measure cost per completed run on your own prompt set and your own harness, and to expect list prices to change. Here one change is already scheduled.
A reproducible visibility check
The following is an illustrative design, not a result from any firm. It is the kind of procedure we would expect to work given the sources above.
- Write the question set. Draft 20 to 30 questions a prospect or referral source might ask about the firm: practice areas, jurisdictions, attorney names, locations, credentials and how to engage. Record the date.
- Run the real surfaces. Execute each question in the tools actually in use, noting product, model or version, date and whether connectors were enabled. Repeat runs, because outputs vary.
- Split answers into atomic claims. Mark each claim as supported by a first-party record, contradicted by one, or unsupported.
- Trace the citations. For each cited URL, record whether the firm controls it and whether it is current.
- Assign corrections to an owner. Contradictions go to whoever controls the underlying record: the website, a directory listing, a bar profile or an attorney bio.
- Log cost and cadence. Record tokens or dollars per run, and rerun after a major model or platform release.
The output of this process is a dated set of observations about what a given tool said on a given day. It is not a ranking, and it does not predict whether any AI system will mention the firm in future.
What this does not tell you
Nothing in the supplied sources shows that any law firm is losing inquiries, leads or revenue because of how AI systems describe it. Whether that happens at a particular firm is a hypothesis. Testing it requires first-party data: intake source fields, referral tracking, inquiry logs and call records, read alongside the visibility check above. Public observation alone cannot settle it, and we would not assert it from public evidence.
What the sources do support is narrower and still useful. Four large vendors have built legal offerings around skills and connectors. At least one connector pattern returns answers with citations that users are expected to check. The sound evaluation approach is to test the deployed system, decompose outputs into checkable claims, send the flagged items to a person, and track cost per task. A firm can start that work with its own records this week, before it has any view on which platform its clients or peers will end up using.