Scaling Claude Cowork Across Finance: One Team to Many

One finance team running three great Claude Cowork plugins is a success story. Twelve teams each building their own version of the same reconciliation plugin, with no shared standards, is a maintenance nightmare waiting to happen. Scaling agentic work across a finance organization is less about more usage and more about shared infrastructure and discipline — turning one team's wins into reusable assets without letting a hundred slightly-different copies sprout. This post covers how to go from one team to many without the chaos.

Key takeaways

• Scale by centralizing reusable plugins in a shared registry, not by letting each team rebuild the same thing.
• Separate shared skills (org-wide mappings, policies) from local config (team-specific entities, thresholds).
• Version plugins and run a shared golden eval set so an update can't silently break ten teams at once.
• Establish a platform owner or guild that maintains connectors, standards, and the registry.
• Roll out in waves — prove with a lead team, template it, then onboard the next teams against the template.

To define the term: scaling agentic finance work means moving from isolated, team-built plugins to a shared, versioned, governed library of skills and MCP connectors that many teams reuse with local configuration — so the organization compounds its work instead of duplicating it.

Why does naive scaling create chaos?

Left alone, each team will solve its reconciliation, its variance commentary, and its vendor triage independently. The result is a dozen near-identical plugins with subtle differences in account mappings, thresholds, and prompts. When a policy changes — a new chart of accounts, a different materiality threshold — someone has to find and fix all twelve, and they won't find all of them. Drift, inconsistency, and audit headaches follow. The cure is a shared core with local overrides.

What does the scaling architecture look like?

The pattern that works is a central registry of versioned plugins, a shared library of org-standard skills and connectors, and thin per-team config layers. A platform owner maintains the shared core; teams compose from it and supply only what's local to them.

flowchart TD
  A["Lead team builds plugin"] --> B["Generalize: shared skills + connectors"]
  B --> C["Publish to central plugin registry (versioned)"]
  C --> D["Team A: add local config"]
  C --> E["Team B: add local config"]
  C --> F["Team C: add local config"]
  D --> G["Shared golden eval gates every update"]
  E --> G
  F --> G
  G --> H["Platform owner ships version bump"]

The key insight is that the eval set is shared and central too. When the platform owner ships an update to the org-wide reconciliation skill, it runs against every team's golden cases before release — so a change that would break Team B is caught before it reaches Team B.

A registry layout that keeps teams in sync

Here is a directory shape that cleanly separates the shared core from local config, which is what prevents the twelve-copies problem:

finance-plugins/
  shared/
    skills/
      coa-mapping/          # one org-wide chart-of-accounts map
      materiality-policy/   # shared thresholds & rules
    connectors/
      erp-read/             # least-privilege, central
    evals/
      reconciliation.golden # shared test cases
  teams/
    emea/
      config.yaml           # entities, local thresholds, overrides
    amer/
      config.yaml
    apac/
      config.yaml
  registry.yaml             # plugin versions each team is pinned to

With this layout, a chart-of-accounts change happens once in shared/skills/coa-mapping/ and propagates to every team on their next version bump. Teams only edit their own config.yaml. The registry.yaml pins versions so no team is force-upgraded before its evals pass.

Common pitfalls when scaling across teams

• Copy-paste proliferation. Forking a plugin per team guarantees drift. Share the core and override locally instead.
• No version pinning. Pushing a change to all teams at once means one bug hits everyone during close. Pin versions and roll forward deliberately.
• No platform owner. Without a named owner for shared skills and connectors, the registry rots and teams quietly re-fork.
• Skipping shared evals. If each team tests in isolation, a shared-skill change can break teams nobody thought to check. Gate on the combined golden set.
• Big-bang rollout. Onboarding every team at once amplifies any mistake. Wave by wave, with the previous wave's lessons baked into the template.

Scale across the org in 7 steps

1. Prove the pattern with one lead team and a single well-loved plugin.
2. Generalize it: pull org-wide logic into shared skills, leave team specifics in config.
3. Stand up a central registry with versioning and a shared golden eval set.
4. Name a platform owner or guild accountable for the shared core and connectors.
5. Onboard the second team against a template; capture what was hard and fix the template.
6. Roll out in waves, pinning each team's plugin versions until their evals pass.
7. Run a recurring review of usage, drift, and access across teams.

Per-team builds vs. shared registry

Dimension	Each team builds own	Shared registry + local config
Policy change effort	Fix N copies	Fix once, propagate
Consistency	Drifts fast	Enforced by shared skills
Audit story	Fragmented	Centralized & versioned
Speed to onboard new team	Slow rebuild	Add a config file
Failure blast radius	Contained but duplicated	Controlled by version pinning

Frequently asked questions

Should every team use identical plugins?

They should share the same core skills and connectors but keep their own thin config for local entities and thresholds. Identical-down-to-the-config rarely fits real org structure; shared-core-with-overrides does.

Who owns the shared plugins?

A named platform owner or a small finance-AI guild. Someone must be accountable for the registry, the connectors, the shared eval set, and version releases, or the whole thing erodes into forks.

How do we avoid breaking ten teams with one update?

Version everything, gate every change on the combined golden eval set across all teams, and roll forward in waves with pinned versions rather than force-upgrading everyone simultaneously.

Scaling agentic AI to every phone line

CallSphere scales the same shared-plugin discipline into voice and chat across many locations and teams — agentic assistants that reuse common skills and connectors to answer every call and book work, without each site reinventing the wheel. See it live 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.