Scaling Claude Across a Law Firm Without the Chaos

The first team to adopt Claude in a law firm almost always succeeds. A motivated group, a few well-chosen tasks, a partner who gets it — the pilot looks great. The trouble starts at team number five. Suddenly there are six different ways to prompt the same contract review, three incompatible sets of confidentiality rules, an associate quietly pasting client data into a consumer account because no one told her not to, and a knowledge base of useful prompts trapped in private chat histories. Scaling Claude is not a bigger version of piloting it. It is a different problem — one of standardization, platform, and governance — and treating it like the pilot is how firms end up with sprawl instead of leverage.

This post is about going from one team to many deliberately, so the second hundred users are as effective and as safe as the first ten.

Why the second team is the hard one

A single team self-organizes. Its members talk daily, share a partner, and converge on a way of working without anyone designing it. The moment a second and third team adopt, the implicit coordination breaks. Each team reinvents prompts, redefines what counts as an approved matter, and builds its own ad-hoc connectors. The result is not just duplicated effort; it is divergence in exactly the places — confidentiality, citation discipline, output quality — where a firm needs consistency. Scaling without a deliberate platform produces a dozen incompatible micro-deployments that are individually fine and collectively a governance nightmare.

The reframe that helps: stop thinking of Claude as a tool individuals use and start thinking of it as infrastructure the firm operates. Infrastructure needs an owner, standards, and a way to evolve. That shift in mindset is the real prerequisite for scale.

The platform layer that makes scale sane

The technical answer to sprawl is a thin internal platform that every team uses, rather than each team touching Claude directly however it likes. The platform centralizes the things that must be consistent — matter-based access control, model-tier routing, audit logging, content filtering — while leaving teams free to innovate on the things that should vary, like the specific Skills and prompts for their practice area.

flowchart TD
  A["Any legal team"] --> B["Shared Claude platform"]
  B --> C["Central access + matter control"]
  B --> D["Shared Skill & prompt library"]
  B --> E["Managed MCP connectors"]
  C --> F["Audit log"]
  D --> G["Team-specific Skills"]
  E --> H["DMS, billing, docket"]
  F --> I["Firm-wide oversight"]
  G --> I

The Claude Agent SDK is the natural foundation for this platform, because it lets you build on Claude Code's primitives — subagents, Skills, hooks, MCP — while wrapping them in your firm's controls. The platform team owns the connectors so that a new practice group does not have to reimplement access to the document-management system; they own the audit pipeline so every team is logged the same way; and they curate the Skill library so the best work of any one team becomes available to all.

Shared Skills are how knowledge scales

The single highest-leverage scaling move is a curated, firm-wide library of Agent Skills. An Agent Skill is a packaged folder of instructions, examples, and resources Claude loads when a task is relevant, and at firm scale it becomes the mechanism by which one team's hard-won expertise reaches everyone. When the M&A group perfects a due-diligence checklist Skill, the platform makes it discoverable to the whole firm; when litigation builds a deposition-summary Skill that handles your jurisdictions correctly, every litigator inherits it.

This solves the trapped-knowledge problem that kills informal scaling. Without a shared library, every team's best prompts live in private histories and die there. With one, the firm builds a compounding asset — a codified, searchable version of how this firm does its best work — and each new Skill makes the next task across the whole organization a little cheaper. Govern the library with light review so quality stays high, but make contribution easy so it actually grows.

Governance that scales without strangling

The governance that protected the pilot must scale too, and the temptation is to centralize everything into a slow approval committee. That kills momentum. The pattern that works is paved roads: the platform makes the safe path the easy path. Because matter-based access, model routing, and logging are built into the platform every team uses, individual lawyers do not have to make security decisions — they make them implicitly by using the supported workflow. New connectors and high-risk tool permissions still go through review, but routine use does not, so governance scales without becoming a bottleneck. A small platform-and-risk team can safely oversee a large user base precisely because the controls are structural rather than procedural.

The complementary discipline is observability. At ten users you can ask people how it is going. At three hundred you need the audit and eval pipeline to tell you which task types are reliable, where outputs are being discarded, and which teams have drifted from the standard. Scaling blind is how quiet quality problems become firm-wide ones.

Sequence the rollout, do not flood it

Finally, scale in waves, not all at once. After the first team proves a workflow, the right move is to harden it into the platform and onboard the next two or three teams whose work most resembles the first — not to throw the whole firm at it simultaneously. Each wave stress-tests the platform, surfaces the next set of needed Skills and connectors, and produces internal champions who help the following wave. By the time you reach the teams with the most idiosyncratic or sensitive work, the platform is mature and the firm has a deep bench of people who know how to deploy it well. Rushing to firm-wide access on day one skips the learning that makes scale safe.

Frequently asked questions

Why does scaling Claude across a firm fail when the pilot succeeded?

A single team self-organizes; multiple teams diverge. Without a deliberate platform, each team reinvents prompts, redefines approved matters, and builds ad-hoc connectors, producing inconsistency in exactly the areas — confidentiality, citation discipline, quality — that need to be uniform. Scale is a standardization problem, not a bigger pilot.

What should a firm-wide Claude platform centralize versus leave to teams?

Centralize what must be consistent: matter-based access control, model-tier routing, audit logging, content filtering, and managed MCP connectors. Leave to teams what should vary: the practice-area-specific Skills and prompts. The Claude Agent SDK is a natural foundation because it wraps Claude's primitives in your firm's controls.

How do you keep governance from becoming a bottleneck at scale?

Use paved roads. Build access control, routing, and logging into the platform so the safe path is the default path and lawyers make secure choices implicitly. Reserve formal review for new connectors and high-risk permissions, not routine use, so a small risk team can oversee a large user base structurally rather than procedurally.

Should we roll out to the whole firm at once?

No. Scale in waves. Harden the first team's workflow into the platform, then onboard the teams whose work most resembles it, building champions and surfacing needed Skills each wave. Reach the most idiosyncratic or sensitive practices last, once the platform is mature.

Bringing agentic AI to your phone lines

CallSphere takes these same scale-without-chaos patterns to voice and chat — a shared platform of agentic assistants that answer every call and message, use approved tools mid-conversation, and stay consistently governed as you grow. See it operate at callsphere.ai.

---

Source & attribution: This is an independent, original explainer inspired by Anthropic's coverage on the Claude blog. Claude, Claude Code, Claude Cowork, Claude Opus, and the Model Context Protocol are products and trademarks of Anthropic. CallSphere is not affiliated with or endorsed by Anthropic.