Governing AI Usage: Budgets, Allowlists, and Approval Gates

Governed AI is a first-class citizen of the EchoFoundry core, not an add-on bolted onto individual modules. This episode covers the three levers you actually configure: budgets, model allowlists, and approval gates.

Why governance lives at the core

Because every module's AI usage routes through the same governed layer, a budget or allowlist set at the core level applies consistently everywhere that layer is used. You are not setting spend limits separately in EchoAgents and separately again in EchoAI and hoping they add up correctly. Set it once, correctly, and every module inherits it.

Budgets: caps that actually cap

A budget in EchoFoundry is a hard spend ceiling tied to a scope: an organization, a workspace, a specific agent deployment, or a specific user. Configure budgets at the tightest scope that makes sense for the risk: an experimental agent deployment should have a tight, cheap budget; a production workflow with proven ROI can carry a larger one, reviewed periodically rather than left open-ended.

Two settings matter beyond the raw number:

• Alert thresholds. Configure a warning at (for example) 75% of budget so you have runway to react before a hard stop interrupts real work.
• Reset cadence. Decide whether a budget resets daily, weekly, or monthly, matched to how bursty the actual usage pattern is: a support-agent workload with steady daily volume wants a daily reset; a periodic batch job wants a cadence matched to when it runs.

Model allowlists: choosing what's allowed to answer

An allowlist restricts which underlying models a given scope is permitted to route to. This protects you two ways: it prevents an unreviewed model swap from silently changing behavior in a workflow you've already tested and trust, and it lets you enforce provider-level requirements (no-training guarantees, zero-retention agreements, specific compliance certifications) at the point of use rather than trusting every downstream caller to remember them.

Review the allowlist whenever a new model becomes available. Don't add it by default. Treat allowlist changes with the same care you'd apply to a dependency upgrade in production code: test in a lower-risk scope first, then promote once it holds up.

Approval gates: risk-tiered human checkpoints

Not every AI-assisted action deserves the same scrutiny. Approval gates let you require human sign-off for actions above a defined risk tier (a large financial commitment, a customer-facing communication after a complaint, a change to a live production configuration) while letting low-risk, high-volume actions proceed without a bottleneck.

Design your tiers around consequence and reversibility, not just dollar amount:

1. Low tier, no gate. Reversible, low-consequence, high-volume (routine status drafts, internal summaries).
2. Medium tier, async review. A human reviews before or shortly after action, but doesn't block real-time flow (most customer-facing drafts).
3. High tier, hard approval gate. Action cannot proceed until an authorized human explicitly approves it (large financial actions, irreversible account changes, anything already flagged as a trust-critical moment).

Observability: the fourth lever

Budgets, allowlists, and approval gates are the levers you configure; the governance dashboard is where you confirm they're working. Review spend trends, allowlist violations (attempted, not just successful), and approval-gate throughput on a regular cadence, the same discipline as the capability-grant review in the previous episode.

Field exercise

For one AI-governed workflow in your organization, write down its current budget scope and amount, its model allowlist, and which approval tier it falls into. If any of the three is undefined or defaulted, that's your action item: define it explicitly before next week.