Surinch
页面加载中Surinch
页面加载中Define Identity, Data Scope, and Approval
Before an AI Agent connects to any enterprise data source, an authorization profile must be defined. This profile answers: who owns this agent, what data can it see, what actions can it take, what evidence must it produce, and what conditions cause its access to be revoked. This is not a compliance form — it is an operational boundary document.
Define five elements before the agent connects to any data source: (1) Business Identity — the role the agent serves, owning team, and what decisions it influences versus makes; (2) Data Scope Tiers — which tables, fields, and documents the agent can access, with explicit exclusions; (3) Action Permission Levels — what the agent can do with accessed data (read, suggest, execute-with-approval); (4) Audit Evidence Requirements — what must be logged for every interaction; (5) Stop Conditions — what triggers automatic access revocation. Document these five elements in a single profile that business, data, and security owners all review before the agent goes live.
Complete each section. No agent should connect to data with blank fields.
Unique identifier for this agent instance
What workflow or business problem this agent supports
Team and specific human responsible for this agent
What business decisions the agent informs (does not make)
How often the profile is reviewed (recommended: monthly)
When this authorization profile expires if not renewed
List every data source, specific tables/views, and explicitly excluded columns or rows.
Tables and fields the agent can read without restriction. List specific table.field names.
Example: sales.orders, sales.order_date, sales.order_amount
Tables and fields the agent can access only under specific conditions (e.g., date range, row filter).
Example: customers.email (only for orders within the last 30 days)
Tables and fields the agent must never access. Including PII, financials, and regulated data.
Example: hr.employee_salary, finance.bank_accounts, legal.contracts
Data scope must be reviewed and re-approved whenever new tables, fields, or data sources are added to the agent's access list.
| Level | Action Type | Requirement | Audit |
|---|---|---|---|
| L0 | Read data | Within defined scope | Access log |
| L1 | Suggest analysis | Human review before sharing | Suggestion log + review record |
| L2 | Suggest actions | Business owner approval | Proposal + approval + rationale |
| L3 | Execute approved | Dual human approval | Full chain: proposal, approvals, before/after, result |
| L4 | Execute auto | Not allowed in initial profile | N/A — should be blocked |
Timestamp, agent ID, query, data scope, rows returned
Proposed action, confidence score, data referenced, reviewer assigned
Reviewer ID, decision, rationale, timestamp, linked proposal
Boundary violation attempts, unusual access patterns, cost spikes
Agent access is automatically revoked when any of these conditions are triggered.
Agent attempts to access data outside its defined scope (Tiers A, B) or in the Denied tier (Tier C).
Agent executes or attempts to execute an action above its authorized permission level without approval.
Audit log shows 3 or more boundary violations within any 24-hour period.
Quality review shows agent suggestions are incorrect more than 30% of the time over 10 consecutive reviews.
API cost or token usage exceeds the monthly budget limit set in the profile.
Authorization profile has expired and has not been renewed by the owning team.
At minimum: the business owner (validates purpose and workflows), the data owner (validates scope and exclusions), and a security representative (validates controls and audit requirements). For agents with execute-level permissions, add a second business approver.
Update the authorization profile, get re-approval from data and security owners, and log the change. The agent should not connect to the new source until the updated profile is approved. Incremental scope expansion is safer than starting with broad access.
An API key or service account controls authentication — proving the agent is who it says it is. This profile controls authorization — defining what the authenticated agent is allowed to do. Both are needed. A key without a profile is like a badge without access rules.
InchStack provides the control plane for defining data scope, action tiers, human approval checkpoints, audit logging, and stop-condition monitoring. It integrates with your existing database permissions and agent frameworks to enforce profile boundaries at the workflow level.
Start with a controlled trial and profile template.