automation

Approval workflows

Review prepared actions before they are authorised, understand risk levels, and design workflows that are safe to run unattended.

· OpenbirdAI documentation

An approval workflow is the checkpoint between "OpenbirdAI could do this" and "OpenbirdAI did this". Designing it well is what makes automation safe to leave running.

This guide describes a capability that depends on the connected provider, the permissions granted to it, and your current plan. If the provider cannot supply the required data, OpenbirdAI reports it as unavailable rather than estimating.

How approvals work

When OpenFeather prepares an action, the workflow shows what would happen, against which account, with which content. You approve or reject it. A rejected action produces no external change and no cost for the execution it never performed.

Risk levels

  • Low risk — read-only or internal preparation, such as drafting a message for review.
  • Elevated risk — anything that leaves OpenbirdAI or consumes a provider capability.
  • Not permitted — actions the product will not perform regardless of approval.
An action that is not permitted is a product decision, not something approval can unlock.

Designing a safe approval

  1. Make the prepared action easy to read in full — do not truncate the content you are approving.
  2. Show the target account explicitly.
  3. Require approval for every external action, even in an automated run.
  4. Log the decision so you can audit it later.

What approvals do not cover

  • Provider-side actions you take outside OpenbirdAI.
  • Actions performed by tools you configured elsewhere.
  • The accuracy of the reasoning that led to the action — approval is your judgement about the plan, not a guarantee of the outcome.

What a good approval screen shows

  • The complete content of the action, not a truncated preview.
  • The exact account or destination it would affect.
  • What will be sent, created or changed, in plain language.
  • Whether anything irreversible is involved.

If any of those are missing, do not approve. An approval you cannot fully read is not an approval.

Rejection is a normal outcome, not an error. It records that you chose not to act, which is exactly what the checkpoint is for.

Auditing approvals over time

Approvals are only useful if someone looks at them. Once a workflow is trusted, approvals become rubber stamps and the checkpoint stops doing its job.

  1. Review recent approvals monthly at the start.
  2. Ask for each approved action to be justified in one line.
  3. Revoke standing approval for anything nobody remembers approving.
  4. Tighten the workflow when the same wrong action keeps getting approved.
A workflow where you no longer read the approvals is an automation, not an approval process. Treat that as a signal to slow it down.

Common questions

I approved an action but nothing happened externally. Why?

Approval authorises the attempt. If the provider rejected it, the product reports the failure rather than silently retrying. Check the execution timeline.

Can approvals be skipped for low-risk tasks?

Keep approval for anything external. Skipping it removes the only checkpoint that is independent of the model.

Was this page helpful?