Delegation and Permissions: How Your Agent Can Act for You (Without Overreach)

Delegation means giving your agent explicit permissions for specific actions. It's not about handing over control—it's about defining exactly what the agent can do, under what conditions, and with what oversight.

In practice, delegation looks like:

• "Draft a follow-up email based on my meeting notes" — the agent creates a draft, you review and send.
• "Schedule a meeting with this client" — the agent checks availability and proposes a time, you confirm.
• "Update the CRM with call notes" — the agent writes to a specific field, logged and reviewable.

Each action has a clear scope, a defined input, and a reviewable output. That's safe delegation.

ind.ai organises permissions into four levels, each with increasing capability and control:

• Read — the agent can view data but cannot modify, share, or act on it. Lowest risk.
• Draft — the agent can create outputs (emails, documents, plans) but cannot send or publish them. You review everything.
• Propose — the agent can suggest actions (schedule a meeting, update a record) and present them for your approval.
• Execute — the agent can act autonomously within defined boundaries. Highest capability, highest oversight needed.

Most users start at Draft and move to Propose as they build confidence. Execute permissions should be reserved for well-tested, low-risk workflows with strong logging.

An approval gate is a checkpoint where the agent must present a preview and wait for user confirmation before proceeding.

Actions that should always require approval:

• Money movement — payments, transfers, refunds, subscription changes.
• Identity changes — profile updates, verification submissions, handle modifications.
• Legal commitments — contract signatures, terms acceptance, formal communications.
• Medical decisions — appointment bookings involving treatment, prescription-related actions.
• Irreversible actions — data deletion, account closure, public publications.

The preview should clearly show what will happen, who will be affected, and what data will be used—so the user can make an informed decision.

Every delegated action generates an audit log—a record of what the agent accessed, what it did, and what the outcome was.

Audit logs support:

• Accountability — trace any action back to the agent and the permission that authorised it.
• Debugging — when something goes wrong, logs help identify the root cause.
• Dispute resolution — if a handoff or collaboration produces unexpected results, logs provide evidence.
• Continuous improvement — reviewing logs helps you refine permissions and workflows over time.

Conceptually, well-designed systems also support rollbacks—the ability to undo an action if it was incorrect or unauthorised. This is another layer of safety.

A safe, practical approach to delegation:

1. Start with draft-only — let the agent create outputs without sending or publishing.
2. Add a review checklist — before upgrading permissions, define what "good enough" looks like.
3. Upgrade gradually — move to Propose after consistent, correct drafts. Move to Execute only for low-risk, well-tested workflows.
4. Scope by task, not by agent — give different permissions for different workflows, even on the same agent.
5. Review logs weekly — check what your agent did, catch issues early, and refine permissions.

The goal is progressive trust: your agent earns more capability by demonstrating reliability. You stay in control at every step.