Gemini for Legal vs Claude + MCP: A Small-Firm Take

By Jude Lee · · Comparison

Lawyers and a paralegal reviewing documents on a screen in a small law firm conference room

What the platform announcements mean below the AmLaw 200

Read the vendor announcements themselves before you read the coverage — and confirm dates, tiers and pricing on the vendor’s own press page or product documentation, because secondhand summaries (including this one) go stale quickly. As of 2026 the pattern across them is consistent: the big legal platforms are wiring agents into the systems large firms already run — enterprise DMS, enterprise practice and financial management. That is a real shift, and eventually it is genuinely good news for everyone.

The honest small-firm translation: if you run an enterprise DMS or an enterprise financial-management suite, this roadmap matters to you directly. If you run Clio, MyCase, Smokeball, Filevine or a shared drive, it mostly signals where the market is heading — not something you can buy next quarter at your seat count.

Three ways a small firm actually gets agentic AI

Path 1 — Platform-native agents. Your DMS or practice-management vendor ships agents that already live inside your data. Zero integration work, vendor-owned security model, and the agent understands the platform’s own objects (matters, documents, timekeepers). The tradeoff: you get the workflows the vendor decided to build, on the vendor’s timeline, and enterprise-tier offerings often assume enterprise-tier infrastructure and budget.

Path 2 — Practice-platform AI features. Clio Duo, Smokeball Archie and their peers sit closer to the small-firm price point. They are strong at summarize-this-matter and draft-this-email work and deliberately narrow beyond that — and for plenty of firms, narrow is the correct final answer. Concrete case: a nine-lawyer family law firm that keeps every document, note, deadline and time entry inside one practice-management system, and whose highest-volume AI task is drafting routine client update emails and pre-hearing file summaries. All the data the agent needs is already in the platform, a lawyer reviews the output regardless, and the volume is a few dozen items a week. A custom build there would add cost, a maintenance burden and a new security surface while reaching exactly the same data the bolt-on already reaches. Buy it, use it, stop. We compared this category in detail in Clio Duo vs Smokeball Archie vs a custom agent.

Path 3 — A general assistant connected via MCP. MCP (the Model Context Protocol) is an open standard, created and open-sourced by Anthropic, for giving an AI assistant secure, scoped access to specific data and tools. Support is deepest in Claude today; other assistants and developer tools have been adopting it, but maturity varies — verify current support before you count on portability. Instead of pasting documents into a chat window, you connect an MCP-capable assistant to your practice-management system, your document store, or your own internal database, with permissions you define. Practical walkthroughs: connecting Claude to Clio with MCP and building a custom MCP server over your matter files.

Platform-native agents (enterprise DMS and practice/financial suites)
Fastest when you already run that platform. Vendor handles security review, hosting, and updates. Agents are pre-scoped to the vendor’s understanding of legal work. Weak fit if your data lives across three systems the vendor doesn’t cover, or if your best workflow is idiosyncratic to your practice area.
Claude + MCP
Works across whatever you can expose — practice management, a folder tree, a spreadsheet of referral sources, your billing export. You define the tools and the guardrails. It also means you own the security posture: someone has to store API credentials somewhere better than a laptop, decide and periodically re-review which tools the agent may call, keep an access log you’d be comfortable showing a client or a bar investigator, and document a handover so the setup doesn’t become undocumented if the person who built it leaves.

A worked example: the “what’s the status of this matter” question

Take a workflow every firm has. A client calls; the associate needs the current posture of a matter with 400 documents, six months of email, and three deadlines.

With a platform-native agent this is close to a solved problem if everything relevant lives in that platform. For a firm running Clio for matters, Dropbox for documents and Outlook for correspondence, no single vendor agent sees the whole picture.

With an MCP setup, you expose three read-only tools — search_matter_documents, get_matter_timeline, list_upcoming_deadlines — and the assistant composes an answer with citations back to specific documents. Scope the effort honestly rather than trusting anyone’s generic timeline: ask whoever would build it for an estimate in developer-days for a three-tool, read-only server against your specific systems’ documented APIs, plus a named owner for ongoing maintenance.

Buy the agent that already lives where your data lives. Build the agent when your data lives in four places and no vendor covers all four.

Where agents break, and what stays human

Modeling whether it’s worth it

Don’t take anyone’s ROI headline, including a vendor’s. Build the estimate yourself with numbers you can defend:

(Hours spent on the task per month × people doing it) × realization-adjusted hourly rate = the ceiling on value. Then multiply by the share the agent realistically removes — and be conservative, because review time replaces some of the drafting time you saved. Subtract licenses, build cost amortized over a window you choose, and maintenance.

hrs/mo × rate
Formula for the value ceiling of one workflow
Illustrative model — use your own timekeeping data
× 0.3–0.6
An arbitrary starting placeholder for how much of that time an agent removes — replace it with what your own 30-day pilot actually measures
Placeholder only; no data behind it, not a benchmark
12–36 mo
Amortization window for a custom build — pick one you'd defend to your partners; 24 months is a common starting point
Your call, not an industry standard

If the ceiling on a workflow is small, no agent is worth building. That’s a real outcome, and it’s why we’d point most firms at ranking workflows by hours recovered before shopping.

A 30-day way to decide

  1. Pick one workflow with a measurable clock

    Matter status summaries, lease abstraction, discovery triage, intake qualification. One. Something you can time today.

  2. Ask your current vendors first

    Email your DMS and practice-management reps: what agentic capability is shipping, on what tier, when. If the answer covers your workflow at a price you’ll pay, stop here — that’s the cheapest good outcome.

  3. Run a manual baseline

    Have one person do the task with a general assistant and copy-paste, no integration. If output quality fails here, integration won’t rescue it.

  4. Prototype read-only access

    Connect an assistant to one system via MCP with read permissions only. Measure whether answers hold up against source documents — and record the real time saved to replace the placeholder above.

  5. Decide with numbers, then add write actions slowly

    Only after read-only is trustworthy should any agent create, file, or send anything — and each write action gets a named human approver.

The unglamorous conclusion

The platform announcements validate the direction: agents that act inside legal systems of record, not chatbots in a browser tab. If you’re on an enterprise stack, watch that roadmap closely and let your vendor carry the security burden. If your workflow lives entirely inside one practice-management platform, buy the bolt-on and move on. Build only when your data is scattered across systems no single vendor covers — and when you build, start read-only, keep the integration layer open, and name the person who owns it, so you’re not re-buying this decision every time a new vendor announcement lands.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides