AI Governance
An AI integration starts with permissions, not APIs.
Before connecting a system, decide what AI may read, write, trigger and escalate, and what still requires a person’s approval.
“Can we connect it?” is only part of the integration question. A successful connection does not tell you what the AI should be allowed to do through it.
Decide the authority before treating technical access as a finished design.
Use specific verbs
Read a record. Draft an update. Save a change. Trigger a process. Approve an action. These are separate permissions, even if one technical credential could enable several of them.
For a hypothetical request workflow, the system might read permitted details and prepare a draft while a staff member approves submission. That boundary must be enforced by the application and its access controls, not merely suggested in an instruction.
Design the refusal and escalation path
What happens when permission is missing or approval is denied? Which information does the reviewer receive? Who owns a request that cannot proceed?
If the answer is “the agent will work it out,” the business decision is still unresolved. AI agents need bounded authority and a clear stop condition just as much as they need tools. A tool error should not become a reason to try a broader permission.
A connection gives access. Your design must define authority.
Then define the connection
Once the task and authority are clear, technical teams can assess the available interfaces, tool definitions and controls. Tool calling lets a model request an operation; the application remains responsible for validating and executing it. MCP can standardize parts of tool and context exchange, but it does not decide the organization’s business permissions.
The business-agent guide provides a broader readiness checklist. Keep the permitted task, the actual access and the human approval process aligned as the integration evolves. A new connection is a reason to review authority, not an automatic extension of it.
A useful integration makes it clear what may happen, who may approve it and where the system must stop.