Approval gates: how to keep humans in the loop without slowing down

An approval gate puts a named person’s sign-off on the step that matters. How Mercury Digital places gates, what they cost, and how they map to NIST AI RMF.

The short answer

Short answer

You keep a human in the loop by putting an approval gate on each consequential step, not on every step.

With Mercury Digital, a gate is a named person's sign-off at a specific point in the workflow. It's closed by default, the agent can't open it, and the work waits there with its evidence attached.

Definition

Approval gate

A point in an agent workflow where a consequential action (posting, paying, sending, deleting, changing a record of truth) stops until a named person approves it. The gate is closed by default and opens only on a recorded human decision.

Gate, review or escalation: which one you need

MechanismWhat it isWhen to use itWho actsCost
Approval gateThe action is blocked until a named person approvesThe step is consequential and hard to undo (money, external sends, records of truth)One named gate holderLatency plus the holder's attention, on every item that reaches it
ReviewThe action happens; a person checks a sample or a summary afterwardsThe step is reversible and low-stakesReviewer (often rotating)Attention only, and errors are caught later
EscalationThe agent stops and hands over when its rules don't cover the caseExceptions the rules can't decideOwner of that exception typeOnly on exceptions

Many workflows need only one gate. A gate on every step teaches the reviewer to stop reading, which is the failure mode gates are meant to prevent.

How to place gates, step by step

01

List the steps and mark the consequential ones

A step is consequential if it moves money, sends something outside, or writes to a record of truth.
02

Name one gate holder per gate

The gate holder is a person, not a team inbox. Name a deputy for absences.
03

Define what reaches the gate

Either every item reaches the gate, or only items above a threshold you set (amount, vendor type, confidence). Make the threshold visible inside the workflow.
04

Attach the evidence

Each item waits with its source documents, the agent's reasoning and the check results, so the holder decides from one screen.
05

Run in shadow mode first

The agent proposes, people still do the work, and both outputs are recorded. Under Mercury's standard policy, the agent acts on its own only after it has run in shadow for at least 50 items of each type and at least 2 weeks, with 95%+ agreement with the human reviewer and zero critical errors. See shadow mode.
06

Open (or widen) the gate on a recorded decision

That decision is a human one, made against the shadow-mode record. The agent can never open its own gate.
07

Log everything and review the gate itself

Track what reached the gate, what was approved or rejected, and how long items waited. A gate that never rejects anything, or one where work piles up, gets revisited.

Speed

Keeping gates from slowing things down

One gate per consequential step

Not per step.

Thresholds: below the line, items go to review instead of the gate, if the gate holder decides that.

Batching

The holder approves a batch of items from one screen, each with its own evidence.

Complete packages

No back-and-forth, because the evidence arrives with the item.

Visible queues: waiting work shows who it's waiting on.

Where this fits

Concept

Every Mercury Digital deployment ships gated by default. Autonomy is opened gate by gate, only on your decision. This is the Deployment and Governance part of the four phases: Audit → Design → Deployment → Governance. See How It Works and Agent governance.

What to measure. Gate wait time (how long items sit at the gate before a decision), the share of items approved without changes, the rejection rate, and how many items reach the gate versus going to review below the threshold. Mercury sets a baseline for these in week 1 and reports against it.

How approval gates map to the NIST AI Risk Management Framework

The NIST AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1) organizes AI risk work into four functions: Govern, Map, Measure and Manage. Govern is cross-cutting and informs the other three (AI RMF Core). Version 1.0 (January 2023) is the current version; NIST says a revision is in progress. Here's where approval gates fit. This is a mapping for orientation, not a compliance certification.

NIST AI RMF functionWhat it asks for (paraphrased)How an approval gate supports it
GovernPolicies, accountability and clear roles for AI risk, including roles for human oversight of AI systems (e.g. Govern 2.1, 3.2)Every gate has a named holder and a written opening rule. Opening the gate is a recorded decision.
MapEstablish the context and identify risks, including defining processes for human oversight (e.g. Map 1, 3.5)Marking consequential steps is a map of where the agent can do harm and where a person must decide
MeasureAnalyze, assess and track AI risks, including monitoring behavior in production (e.g. Measure 2.4, 3.1)The shadow-mode comparison and gate logs (approved, rejected, wait time) are the measurements
ManagePrioritize and respond to risks; plan response and recovery; be able to override or deactivate a system (e.g. Manage 1, 2.4, 4)The gate stops the action before it happens. Thresholds and escalations change based on the measurements.

The same pattern addresses Excessive Agency, LLM03:2026 in the OWASP Top 10 for LLM Applications 2026 (it was LLM06 in the 2025 list). One of OWASP's listed mitigations is to require human approval of high-impact actions before they're taken. With a gate, the agent never has unattended authority over a consequential action. The OWASP Top 10 for Agentic Applications 2026 covers related risks, including ASI02 Tool Misuse and Exploitation and ASI03 Identity and Privilege Abuse. It also lists ASI09 Human-Agent Trust Exploitation, where a confident, polished explanation leads a person to approve a harmful action. That's why a gate holder gets the evidence, not just the agent's summary.

Guardrails for agents that touch financial data

  • Separate read and write scopes. Write access covers only approved actions.
  • A gate on every posting and payment, held by a named person.
  • Changes to vendor or bank details always go to a person.
  • Shadow mode before any write access.
  • A full audit log: who approved what, when, and against which evidence.

See how this works in month-end close and invoice exception triage.

How It Works

Questions, answered

Put a named person's approval gate on each consequential step: posting, paying, sending outside, changing a record of truth. Let the agent do the rest. The gate is closed by default and opens only on a recorded human decision.
AI Opportunity Audit · Approval gates

Find the step that needs a gate

Bring the workflow you are least willing to hand off and find its consequential step together. The AI Opportunity Audit is fixed-price, quoted before we start.