How to Set Permission Boundaries for AI Agents

Trust starts with boundaries. Before an AI agent can safely act on behalf of your team, someone has to decide — in writing — exactly what it's allowed to do. This guide walks through the process operationally, not philosophically.

Step 1: Inventory the actions the agent can take

List every tool, API, and integration the agent can call. For each, note the blast radius: does it send email, move money, write to a customer record, or just read? You cannot bound what you have not enumerated.

Step 2: Classify actions by risk tier

• Tier 0 — Read-only: queries, lookups, summarization.
• Tier 1 — Reversible writes: internal notes, draft messages, staged changes.
• Tier 2 — External or costly: sending customer emails, booking, spending under a cap.
• Tier 3 — High-impact: payments over threshold, contract changes, data deletion.

Step 3: Set explicit limits per tier

Write the limits as code, not as guidance. Per-action caps (max spend, max recipients), per-session caps, per-day caps, and hard allow-lists for domains, accounts, or record types.

Step 4: Require human approval where the tier demands it

Tier 2 actions typically warrant sampled review; Tier 3 actions warrant sign-off every time. Approvals must be logged with the approver's identity.

Step 5: Instrument, log, and review weekly

Every action, input, and outcome goes into an audit trail. Review exceptions weekly and tighten or loosen limits based on evidence.

Step 6: Expand scope only after a clean track record

Widen permissions the same way you'd promote an employee — on demonstrated judgment, not enthusiasm. This is the operating model behind our Responsible AI-by-Design Framework and the guardrails we deploy through AI Consulting Services.