Sistava

AI Human-in-the-Loop Approvals | Review Risky Steps

Custom AI Team

Pause for human review on risky steps

Human-in-the-Loop Approvals let your AI Employees pause a workflow at sensitive steps and wait for a person to approve, edit, or reject before continuing. Send-money, send-email-to-VIP, post-publicly, delete-record, and contract-sign are common gates. The Employee handles everything else autonomously and only escalates when you tell it to. You configure approval gates as Duties on the Employee. When a gate triggers, the Employee posts a request with full context to the channel of your choice, like Slack, email, or the in-app inbox. A reviewer reads the proposed action, the reason, and the inputs, then clicks approve, edit, or reject. The Employee resumes from that exact point. Finance, healthcare, and legal teams lean on this pattern the most, since a documented human decision at the point of action is what most compliance frameworks now expect from any AI-initiated step with real consequences. This gives you the speed of automation without losing oversight on the decisions that matter. The Employee runs at machine speed for the 95% of steps that are routine, and a human checks the 5% where mistakes are expensive. You keep an audit trail showing who approved what and when.

Benefits

How It Works

  1. Step 1:
  2. Step 2:
  3. Step 3:
  4. Step 4:
  5. Step 5:
  6. Step 6:

At a Glance

< 30 sec
Time to surface a gate
95%
Steps still fully automated
24/7
Gates active around the clock
100%
Decisions captured in audit log

Granular gates per step

Approval gates attach to specific steps inside a Duty, not to the whole Employee. A draft email might run unattended, but sending it to a customer over a contract value triggers a gate. A research summary publishes automatically, but only after a reviewer signs off on the public post. This granularity keeps the Employee fast on routine work and careful on sensitive work, without forcing you to choose between full automation and full manual review for every workflow.

Context attached to every request

When a gate triggers, the approval request carries everything the reviewer needs to decide quickly. The proposed action, the inputs, the reason the Employee chose this path, the source data, and links to the originating ticket or thread. Reviewers do not have to dig through tabs to find context. They read the request, decide, and the Employee resumes. The same package becomes the audit record after the decision, so future reviews can see exactly what the human had in front of them.

Audit trail and accountability

Every approval, edit, and rejection is written to an immutable log with reviewer, timestamp, original proposal, and final action. This satisfies internal controls, makes incident review fast, and gives you data to tune which gates are useful versus noisy. If a gate gets approved 100% of the time with no edits, it is probably ready to be relaxed. If a gate gets rejected often, it points at a Skill or Duty that needs tightening before the gate can be removed.

FAQ

Which steps should have an approval gate?

Steps that move money, send messages to VIPs, publish public content, change contracts, or delete data. Use gates where the cost of a wrong action is higher than the cost of a few minutes of review.

What happens if no one approves the request?

You set a timeout per gate. On expiry, the Employee escalates to a backup reviewer, defaults to reject, or sends a reminder. The default behavior is configurable per Duty.

Can the reviewer edit the proposed action before approving?

Yes. Approve and edit lets a reviewer change the recipient, body, amount, or any input field. The Employee resumes with the edited values and notes the change in the audit log.

Where do approval requests show up?

Anywhere you can reach the reviewer fastest. Slack, email, the in-app inbox, or a webhook into your own tool. You pick the channel per gate or per Employee.

Can multiple people need to approve one step?

Yes. Configure the gate for two-of-three or all-of approvers. The Employee waits for the required number of approvals before resuming.

Is this required for regulated industries like finance or healthcare?

Many compliance frameworks now expect a documented human decision point before an AI-initiated action with real consequences, not just a policy on paper. Approval gates give you that decision point built into the workflow itself: a named reviewer, a timestamp, and the exact inputs they saw, captured automatically instead of assembled after the fact for an audit.