Risk classification and human oversight, designed before a single agent was written.
Fernbridge Business Alliance is a fictitious membership association, used throughout this project for illustrative purposes. No real organisation, individual, or dataset is represented here.
The Sequencing Story
Most agentic AI projects start with the agent. This one didn't. The risk classification matrix, the three-pillar governance structure, and the data foundation were designed and validated before a single line of agent code existed, the same foundation-before-capability discipline behind real transformation programmes, applied here to a working system.
Every task an agent might handle is scored across five independent dimensions, reversibility, data sensitivity, public exposure, reputational/legal exposure, and decision novelty, then routed to one of three oversight models based on the highest-scoring dimension, not an average.
Three Tiers, Three Real Oversight Models
Low — Agent 1
Human-above-the-loop. Acts autonomously within a real-time budget cap; a human monitors and can pause or override.
Medium — Agent 2
Human-in-the-loop. Proposes an action, holds for a single approval before executing. Rejection is isolated from failure tracking.
High — Agent 3
Dual confirmation. Two independent, sequenced approvals required. A live Claude API call sits behind this tier, with a hard cost ceiling.
Governance Principles
Proportionality — control scales with consequence, not activity volume or general caution.
Verifiable Trust — trust is earned through an audit trail, not assumed at deployment.
Data Minimalism — one authoritative source per fact, never a cached or reconstructed copy.
Accountable Escalation — every tier has a named human role accountable for it, even at full autonomy.
Foundation Before Capability — governance and data design come before agent capability, not after.
Tokenomics, Not Just a Budget Cap
Real-time spend tracking, validated against live infrastructure, distinguishes between cost control (a hard budget ceiling per agent) and availability control (rate limiting), a distinction most agentic demos miss entirely. The budget cap protects the wallet. It does not, by itself, protect against a bot exhausting an agent's capacity in minutes. Both layers exist here, deliberately, for that reason.