The AI platform a 1,000-person agency now runs on

The problem

Brainlabs, a PE-backed media agency of about 1,000 people, wanted Claude available across the business. Employees would author skills in Notion; the platform had to publish them to Claude, keep them safely in sync, and remain viable as usage grew.

The original architecture had latent latency and run-cost problems at scale. Fixing those before rollout avoided £1.2M in forecast cost and latency, according to the client's Finance Director.

Skills platform

Founded in April 2026, the platform keeps roughly 800 company skills across 82 plugins in sync between Notion, where people author them, and Claude, where they run. A compiler pipeline syncs from Notion every 10 minutes.

USQ's Principal Architect is the founder and lead architect of the platform.

Token architecture

The company-wide plugin's session-start listing fell from 32,399 to 17,448 characters, a 46.1% like-for-like cut. A blind 44-probe selection evaluation held discovery at 95.5%.

Safe publishing

Immutable skill identities prevent a Notion rename from deleting the wrong folder. A two-stage deletion quarantine and CI validation protect the sync without blocking on pre-existing issues.

Enforced guardrails

Eight of 11 security rules moved from advisory prompt text into deterministic Stop hooks. Three block the turn outright. The controls were tested on a 59-case simulator and through mutation testing.

Governed context

USQ built a production Go MCP server that gives Claude sessions governed, real-time access to the company directory, org chart, client assignments and client context. USQ's Principal Architect built it solo; 52% of the codebase is tests.

USQ's Principal Engineer also designed the memory and semantic layer that gives agents the relevant conversations and account context when work begins.

Authorisation

The service has its own OAuth 2.0 authorisation server, with PKCE and an encrypted token store, plus a dedicated data plane.

Default-deny

Entitlements resolve from SCIM group data. Field-level sensitivity tiers apply before data is returned, and every call—including refusals—writes an audit record.

The data platform

USQ's Principal Engineer re-architected the enterprise data-integration platform using async Python and FastAPI: 12× throughput, ETL runtimes down about 90%, Snowflake compute and storage down 87%, and about $36k a year of SaaS licensing brought in-house.

Multi-tenant OAuth integrations across Xero, HubSpot, Clockify and Greenhouse land in a Medallion model in Snowflake.

Orchestration and CI/CD

USQ's Principal Data Engineer built the Airflow orchestration and warehouse structure, then deepened the CI/CD pipelines that support self-healing operations under Claude-managed agents.

Access, cost and operations

Claude access for the 1,000-person organisation is provisioned by job function against the Anthropic Admin API, with drift correction. At contractor go-live, 23 contractors and five owners were provisioned with zero duplicates.

USQ also prepared an independent audit of billed Claude spend for the Finance Director. On the Snowflake warehouse, the team first-authored 18 specialised Claude Code subagents and 20 of 24 production skills; teammates now extend them.

During a NetSuite migration, the team reconciled 191,632 documents against support tickets and surfaced material sandbox data defects before cutover. A loader fix shipped; bulk repairs remain staged.

The published result

Anthropic's customer story covers the deployment: 1,000+ marketers, roughly 400 employee-authored skills in four weeks, and 91% adoption in North America. The platform has been live since April 2026.

"If you're doing something more than three times, you should try to automate it." — Ben Vincent, CTO, Brainlabs.

"It's one skill, refined across thousands of iterations from our SEO team, that captures the full universe of what to look for and what good looks like." Dan Gilbert, Global CEO, Brainlabs.