Approval Gates in Automation: When to Require a Human
Not every step in an AI workflow should run unattended. Here is how to place approval gates so high-stakes actions pause for a human without killing throughput.
The goal of automation is to run without you. But a few steps in any serious workflow should never run without a human saying yes: sending money above a threshold, deleting data, emailing a whole customer list, signing something. The skill is not choosing between full automation and full manual. It is placing approval gates precisely, on the handful of steps where the cost of being wrong is high, and letting everything else fly. Put gates in the right places and you get speed and safety. Put them everywhere and you have just rebuilt manual work with extra steps.
I run automation across roughly twenty companies, and almost all of it runs unattended. The exceptions are deliberate. A gate is a place where I have decided the downside of an automatic mistake is worse than the cost of waiting for a person. That is a business judgment, not a technical default, and treating it that way is what keeps the automation both fast and trustworthy.
Why blanket automation and blanket approval both fail
Automate everything with no gates and eventually the system does something expensive and irreversible on its own: a bad refund at scale, a mass email with a typo, a deletion that cannot be undone. The blast radius of an unguarded workflow is every action it can take.
Gate everything and you have destroyed the value. If a human has to approve every step, the automation is just a slow assistant that asks permission constantly. Nobody keeps up with that queue, so approvals get rubber-stamped, which is worse than no gate because now the review is theater. This is the failure mode behind compliance theater versus real governance: a control everyone bypasses is not a control.
The answer is selective gating based on stakes.
Where to put approval gates
On irreversible actions. Deleting data, sending money, publishing publicly, anything you cannot cleanly undo. If a step cannot be rolled back or compensated, it is a candidate for a gate. This connects to how you handle partial failure in a multi-step workflow: where you have no clean compensation, a human check before the action is the safer control.
Above a value threshold. A $20 refund runs automatically. A $2,000 refund pauses for approval. Same action, different stakes, different treatment. The threshold turns a binary decision into a graded one, so the common low-stakes case stays fast and only the rare high-stakes case waits.
On low-confidence outputs. When an AI step is unsure, gate it. If the model's confidence is high, let it proceed; if it is low or the input is unusual, route it to a person. This targets human attention exactly where the automation is weakest, which is a far better use of a reviewer than making them check confident, routine outputs.
On anything that touches the outside world irreversibly. Outbound to customers, regulators, or partners. Internal steps can be retried and corrected quietly. External actions are seen, so the bar is higher.
Make the gate fast or it becomes the bottleneck
A gate is only good if it is quick to clear. The approver needs the full context in one place: what the workflow wants to do, why, what it will affect, and a single action to approve or reject. If clearing a gate means digging through logs to reconstruct what is being asked, the gate becomes the slowest part of the system and people start avoiding it.
The workflow also has to park cleanly while it waits. A gated run pauses, holds its state, and resumes exactly where it left off when approved, hours or days later. It does not time out, lose its place, or have to restart. This is real human in the loop behavior: the pause is a designed state, not a crash.
And the gate must be logged. Who approved what, when, and on what information. That record is what makes the whole thing auditable and is the same reason AI decisions need to be defensible after the fact. An approval with no record is an approval you cannot stand behind later.
Gates are a business decision expressed in the workflow
I decide where gates go by asking one question: if this step runs and it is wrong, how bad is it, and can I take it back? High cost and irreversible gets a gate. Everything else runs free. That single question, applied honestly across a workflow, gives you a system that is aggressive where it is safe and careful where it is not.
We build approval gates, thresholds, and confidence-based routing into Girard AI, with clean pause-and-resume and a full approval log, so you can automate boldly without automating recklessly. When you evaluate a platform, ask how a workflow pauses for a human and resumes days later without losing state. If it cannot, your only options are all-automatic or all-manual, and neither one is how a real operation runs.