Thresholdbeta
Human approvals for AI agents

Your AI agent shouldn't act alone.
Put a human in the loop.

Threshold lets your team review, approve, and audit every critical move your AI agents make: refunds, deletions, emails, anything consequential. Before anything touches production.

Minutesto set up
Slack · Email · Webapprove from anywhere
Every actionlogged for compliance
Billing assistant wants to: Refund a customerNeeds review
Pending
Billing assistant

Refund $4,200 to Acme Industries

Details

CustomerAcme Industries
Amount$4,200.00
ReasonDuplicate charge
Awaiting review
The problem

You only find out after the damage is done.

AI agents are already calling real APIs and making real decisions on real systems. When something goes wrong, the first signal is usually a customer complaint, a support ticket, or an angry Slack ping.

A refund goes to the wrong customer.

A billing agent matches on a fuzzy email, issues a refund, and nobody finds out until the affected customer complains.

A production resource gets deleted.

An ops agent runs a cleanup script against the wrong environment. The damage is already done by the time anyone reviews the logs.

A campaign goes out to everyone, twice.

A marketing agent retries on an API error and sends a second blast. Your inbox floods with unsubscribes before the pager goes off.

Common failure modes when AI agents run without a review step.

How it works

Three steps between your agent and a mistake.

Point your agent at Threshold. Write your rules. Every consequential action is checked, routed to the right person, and recorded, without changing how your agents are built.

01

Your agent tries an action

An AI agent in your product goes to call a tool: refund a customer, delete a record, send an email. The request passes through Threshold first.

02

We check your rules

Threshold matches the action against your policies. Routine moves go through instantly. Risky ones are held for review. Dangerous ones are blocked on the spot.

03

Your team decides

Held actions land in Slack, email, or the Threshold dashboard, whichever your team already uses. One click to approve or reject, with a note for the record.

What you get

Control your team can actually run.

Threshold is built for the people who pay the price when an agent goes wrong: ops, finance, support, compliance. Not just the engineers who ship it.

Rules in plain English

Write policies the way you'd explain them to a new hire. No YAML, no code, no engineer required to change them.

Example: Pause refunds over $500

Approve from anywhere your team works

Pending actions appear in Slack, email, and the Threshold dashboard. Approve or reject with one click, without leaving your workflow.

Slack · Email · Web dashboard

Context your team can actually use

Every approval request shows who the agent is, what it's trying to do, and every detail of the request, in plain language.

Full request context, no log-diving

Stop dangerous actions cold

Some moves should never happen: deleting production, touching protected data, bulk-mailing at 3am. Block them before they run, no approval loop.

Always-on guardrails

A complete, tamper-proof record

Every request, every decision, every approver is logged to an append-only audit trail. Export to CSV for compliance reviews anytime.

Append-only · CSV export

Drop in, don't rebuild

Point your agent at Threshold and you're done. No SDK fork, no prompt changes, no restructuring your agent code.

Works with any tool-calling agent
See it in action

A rule, a request, a decision.

Here's what it looks like when a Threshold rule catches an action and routes it to your team for review.

The rule

Pause refunds over $500

Applies to all billing agents

WhenAmount exceeds $500
ThenPause for approval
Route toFinance team

Built with the no-code policy builder. Change the threshold, the team, or disable it entirely. No deploy required.

What your team sees
Threshold#approvalsJust now

Approval needed: refund exceeds $500

Billing assistant is trying to refund a customer charge.

CustomerAcme Industries
Amount$4,200.00
ReasonDuplicate charge reported
Triggered byPause refunds over $500

Full context, one click, logged to the audit trail either way.

After the human decides, the agent continues on its own. No intervention, no re-prompting, no babysitting.

Put a human in the loop before you need to.

Set up in minutes. Works with the agents you're already running. The first time a rule catches something, you'll be glad it was there.

Runs alongside your existing agent infrastructure. No rebuild required.