Four AI agents that leave a human approval step

#ai #saas #agents #code

This issue follows four AI agents that leave a human approval step before a write, retry, reply, or send. Their pages show what each agent reads, what it produces, and where a person can still stop the action.

Approval in the loop

These are short field notes on the queues, documents, drafts, and receipts each agent produces.

TriageDesk


	TriageDesk reads a shared support inbox overnight, classifies each message, and drafts one reply per thread. It is for teams working from Microsoft 365 or Gmail who want a morning queue instead of a wall of unread mail. The queue shows the source message beside the draft, with options to approve, edit, or kill it. Complaints, legal threats, and cancellations go to a human undrafted, while approved replies sit for 120 seconds before dispatch.

TriageDesk reads a shared support inbox overnight, classifies each message, and drafts one reply per thread. It is for teams working from Microsoft 365 or Gmail who want a morning queue instead of a wall of unread mail. The queue shows the source message beside the draft, with options to approve, edit, or kill it. Complaints, legal threats, and cancellations go to a human undrafted, while approved replies sit for 120 seconds before dispatch.

MatchRail


	MatchRail runs a three-way match across bills, purchase orders, and receipts for a month-end close. It is for bookkeepers who need to see what disagrees before a payment run. Each disputed row shows the documents and the reason, such as a missing purchase order or a second copy of an invoice. An approved correction waits 60 seconds, writes an audit row, and re-reads the documents before writing to the ledger.

MatchRail runs a three-way match across bills, purchase orders, and receipts for a month-end close. It is for bookkeepers who need to see what disagrees before a payment run. Each disputed row shows the documents and the reason, such as a missing purchase order or a second copy of an invoice. An approved correction waits 60 seconds, writes an audit row, and re-reads the documents before writing to the ledger.

LeadGrade


	LeadGrade scores every inbound lead overnight and ranks the queue highest-first. It is for sales teams using HubSpot who need the score and its reason beside each lead. Every lead starts at 20 and can add points from form intent, role, company size, email domain, industry, buying language, and completed fields. Approve holds the HubSpot write for 60 seconds, while Kill writes nothing.

CardChase


	CardChase recovers failed subscription charges by reading decline codes, staging retries on a ladder, and stopping hard declines. It is for subscription businesses that need a queue sorted by recurring charge at risk. Each row carries the decline code, issuer advice, attempts spent, and the next rung. A retry can be approved into a 30-second kill window, while a hard decline or a charge at its four-retry cap is stopped.

CardChase recovers failed subscription charges by reading decline codes, staging retries on a ladder, and stopping hard declines. It is for subscription businesses that need a queue sorted by recurring charge at risk. Each row carries the decline code, issuer advice, attempts spent, and the next rung. A retry can be approved into a 30-second kill window, while a hard decline or a charge at its four-retry cap is stopped.

One issue a week, and one shipped product taken apart every month. What each product does, what it cost to build, what the pipeline behind it looks like, and what the numbers did, read off the repository and the live site, not written from memory. Join the list.

DEV Community