Put an AI agent on 364 outbound replies, one click from human takeover

An outbound inbox where an AI agent drafts, qualifies and follows up, with a human one click from taking over.

364
replies handled by the agent
AI SDR workspace showing the conversation list with AI-active and needs-attention filters beside an open thread where the AI agent is drafting replies

The problem

Outbound sales runs on replies, and replies arrive in a heap. A rep sends a campaign, then spends the next week reading every response, deciding which ones are real, drafting an answer to each and remembering to chase the ones that went quiet. The work is repetitive but not mindless: a warm reply needs a considered answer, an out-of-office needs a nudge later, and an unsubscribe needs to be honoured immediately.

Nothing in a plain inbox tells a rep which of those a thread is. Every message has to be opened to find out, and the warm ones sit in the same list as the noise until someone gets to them.

What it cost

The cost of a heap is that warm replies get answered late. A prospect who says "yes, tell me more" on Monday morning and hears back on Wednesday has usually cooled, and the rep never finds out which deals were lost that way because there is no record of the delay.

What we built

An outbound inbox with an AI agent inside it. The agent reads each reply, tags its intent, drafts a response and, where the thread is routine, sends it and schedules the follow-up. Every thread shows its intent tags, its SLA state and whether the agent is still driving or has handed off, so a rep can scan the list and know what needs a human without opening anything.

The human stays one click from taking over. Two control modes decide how much the agent does on its own; three inbox states (AI active, needs attention, handed off) keep the list honest about who owns each thread. A side panel carries the day's counts, new leads, replies, meetings and campaigns, above a needs-attention queue ordered by how warm the prospect is.

The conversation list is the working surface. It can be filtered to the threads the agent is still driving or to the ones that need a person, and each row carries the prospect, the intent the agent read in their last message and where the thread sits against its SLA. Opening a row shows the full exchange with the agent's draft at the bottom, ready to edit or send. Taking over is a single control on the thread: the agent stops, the rep is on record as the owner, and the inbox state changes so nobody else picks it up by mistake. Handing back is the same control in reverse.

AI SDR workspace showing the conversation list with AI-active and needs-attention filters beside an open thread where the AI agent is drafting replies
The conversation list, filtered to threads the agent is still driving, beside an open thread mid-draft.
AI SDR side panel showing new leads, replies, meetings and campaign counts above a needs-attention queue of prioritised prospects
The side panel: new leads, replies and meetings at a glance, then the needs-attention queue.

What changed

The agent has handled 364 replies in the workspace. Warm threads no longer wait behind noise: they surface in the needs-attention queue with the draft already written, so the rep's job becomes reviewing and sending rather than reading and composing.

Stack

  • Next.js
  • TypeScript
  • LLM drafting and intent tagging
  • CRM sync