Why business AI needs permissions, approvals and audit trails

An AI assistant that only answers questions is easy to trust. If it is wrong, you notice and move on. An assistant that can also act, by creating tasks, sending messages or changing records, is far more valuable and far more dangerous. The difference is not the quality of the model. It is whether the surrounding system keeps the assistant inside boundaries your business has chosen.

Three controls do most of that work.

1. Scoped permissions

An assistant should only be able to see and do what it has been explicitly allowed to. In practice that means:

• Per-source access. Connect only the tools a workflow needs, and nothing else.
• Separate read and write rights. Reading project data and changing it are different decisions.
• Respecting the source’s own access rules. If a person cannot see a document in the original tool, the assistant should not surface it to them.
• Revocation that works. Disconnecting a source should stop access immediately.

The principle is the same one security teams apply to people: give the minimum access required, and review it regularly.

2. Human approvals

Not every action carries the same risk. Drafting an internal summary is low stakes. Sending a message to a client, deleting data, changing permissions or touching anything financial is not.

A sensible model has three levels for each type of action:

1. Read-only. The assistant can look but not change anything.
2. Draft-only. The assistant prepares the work, and a person decides whether to use it.
3. Executable with approval. The assistant can carry out the action once a reviewer approves it.

Approvals should be fast to give and impossible to skip for sensitive actions. If review is slow, people will route around it. If it is optional, it will not protect anyone.

3. Audit trails

When an AI system takes an action, someone will eventually ask what happened. A usable audit trail answers four questions without guesswork:

• What was requested, and by whom?
• Which sources did the assistant use?
• What action did it propose and what did it actually do?
• Who approved it, and when?

Audit trails are also how teams build confidence over time. When people can see exactly what an assistant did, they can safely give it more responsibility.

Questions to ask any vendor

Before connecting an AI product to your business tools, ask:

• Can I limit which sources and actions the assistant can use?
• Which actions require approval, and can I change that?
• Can I see a history of every action the assistant took?
• What happens to my data when I disconnect a source?
• Does the product explain where its answers come from?

How Opsvian approaches it

Opsvian is built on these three controls. Administrators choose which sources connect and whether each action is read-only, draft-only or executable. Sensitive and external actions wait for approval, and every request, source and outcome is recorded. You can read more on the Security page, or join the beta to see the controls in practice.