Entrovix AI
Automation

ReplyDesk

Support Automation

ReplyDesk is a support desk with a rule it will not bend: automation drafts, and only a person sends. The product's job is to make that approval fast enough that nobody is tempted to remove it.

Designed & developed by Entrovix

Year
2025
Discipline
Support ToolingAI ProductEngineering
Services
Product StrategyUX/UI DesignAI IntegrationFront-end Engineering
Built with
ReactNode.jsLLM APIPostgreSQL

The Challenge

The reply still needs an author

Letting a model send support replies is the cheap version of this product, and the wrong one. A confident answer that misreads a refund does its damage instantly, in the company's voice, and the first person to learn about it is the customer.

The slow failure sits on the other side of the desk. To answer one refund an agent opens the order system, the payment record, the previous tickets and a template library — four tools for a two-line reply. The brief was to remove the tabs without removing the person.

The Approach

How we worked

  1. Draft beside the conversation

    The assist panel gathers what the agent would have gone looking for — the detected intent, the linked order, the customer's history — and holds it next to the thread instead of behind it.

  2. Approval is the unit of work

    A suggested reply is a card with two pills: Approve and Edit. Nothing reaches a customer that an agent did not explicitly release, which is the whole argument of the product.

  3. Templates a person finishes

    Suggested replies open in an editable state, so the agent adjusts them where they stand. What goes out reads like the person whose name is on it, because that person touched it.

  4. Keep the clock in the queue

    SLA state travels with each conversation as a chip in the list itself, so an expiring ticket is visible from the inbox rather than discovered in a report.

The Solution

Everything the agent would have opened, already open

The inbox runs as three panes: the queue with its unread dots and SLA chips, the open conversation, and the assist panel beside it. A refund enquiry arrives with its intent detected, its order linked and a reply already drafted.

The agent reads, edits if the words need it, and approves. The suggestion never sends itself, and the panel never covers the conversation it serves — assistance sits beside the thread, not in front of it.

Inside the product

What it actually does

Every Screen

One product, whatever it is opened on

The phone layout is not the desktop one squeezed down. Density, hit areas and reading order are decided per screen, and the interface keeps its meaning at both ends.

Capability

What we brought to ReplyDesk

The same services are available on their own — each one links to what it involves, how it runs and what you get.

Built with

  • React
  • Node.js
  • LLM API
  • PostgreSQL

Disciplines

Support Tooling · AI Product · Engineering

The Outcome

What it settled

  • Every outgoing reply carries a person's explicit approval
  • Intent, order and history sit beside the conversation, not behind it
  • SLA state is visible from the queue itself
  • Suggested replies leave in the agent's voice, not the template's

Next project

DocStream

Document Processing

An invoice pipeline that leads with confidence instead of magic — uncertain fields flagged, highlighted at the exact spot on the document, and settled by a person in seconds.

View project

That was some of our work

Now let’s talk about yours.

Tell us what you’re building. You’ll get a written plan, a clear proposal, and a team that shows its work every week.