Skip to main content
AI Stack
Security and oversight

Human approval where it matters.

AI Stack agents work inside business systems with clear rules, visible logs and approval steps before sensitive actions.

  • Fixed price, fixed scope
  • Human approval on anything client-facing
  • Silent test before go-live
  • 60-day guarantee
Approval neededNeeds approval
Lead
Example Lead A
Issue
Quote has had no response for seven days.
Suggested
Send a follow-up asking whether they would like to proceed with the quote.
ApproveEditReject

Example of an approval request

Supervised by design

The agent is supervised from the start.

The agent is set up around agreed rules.

Before go-live, we confirm what the agent can check, what it can draft, what it can update and what it must never do.

Approval rules

Approval rules are set before go-live.

Sensitive actions stay under human approval unless the client has clearly approved a different rule.

Actions that need approval

  • Sending client-facing messages.
  • Changing deal status.
  • Marking a handover complete.
  • Updating important client records.
  • Escalating issues to external contacts.
Logging

Agent actions are logged.

The agent's work should be visible.

Logs should show what was checked, what was drafted, what was updated, what failed and what needs human attention.

Statuses used in the log
  • Checked
  • Draft ready
  • Needs approval
  • Waiting on client
  • Waiting on manager
  • Sent
  • Failed
Task log
  1. CheckedEnquiry matched to existing contact
    Checked
  2. DraftedReply to quote question
    Draft ready
  3. FlaggedDeal status change requested
    Needs approval
  4. SentApproved follow-up message
    Sent
  5. FailedTask tool sync, retry queued
    Failed

Example log

Permission boundaries

Access should match the work.

AI Stack should use the minimum access needed for the agreed workflow.

Silent test

Nothing goes live without a silent test.

The agent runs in test mode before go-live.

During silent testing, it does not take live action. We check whether the summaries, drafts, contact matches, alerts and task updates are accurate enough to approve.

Silent test, day 4
Enquiries detected
22
Contacts matched correctly
21
Drafts prepared for review
14
Task updates proposed
17
Live actions taken
0

Next action. Review one mismatched contact before approving the workflow.

Example of a silent test summary

Client responsibilities

The client still owns the business rules.

AI Stack can install the operating rhythm, but the client must approve access, decision rules, sensitive action boundaries and escalation paths.

Privacy

Privacy law is changing on 10 December 2026

From 10 December 2026, Australian businesses that use personal information in automated decisions have to say so in their privacy policy. We built a free ten-question check so you can see whether your current automation is in scope.

Set the rules before the agent goes live.

The fit check starts with your tools, workflow and approval boundaries.