Our Inbox Was Becoming a Second Job

The Inbox Problem Nobody Talks About

Here's what a candidate reply actually requires: open the email, find the original outreach (sometimes buried in a thread), remember who this person is and what role we were sourcing for, assess the tone of their reply, think about what the right response looks like, write something that sounds like you, actually send it. Then the next one.

When you're doing 10 replies a day, that's manageable. When you're doing 50, it becomes your whole afternoon. And when you're tired, the quality drops — you send the same generic response you sent to the last person, and the conversation dies.

What We Built

The inbox assistant monitors our talent inbox and handles the heavy work automatically for every inbound message:

• →Reads and classifies each incoming message — candidate reply, spam, system notification, or something that needs direct attention
• →For candidate replies, fetches the full thread and the original outreach we sent
• →Understands the context — who the candidate is, what role we're sourcing for, what they said and what it means
• →Drafts a response in our voice — specific to this candidate and this conversation, not a template
• →Presents the draft for review — I read it, decide, and respond

The agent doesn't send anything automatically. Every reply goes through a human decision before it goes out. But the human decision is now fast — instead of spending 10 minutes gathering context and writing something, I spend 30 seconds reading a good draft and deciding whether to send it as-is, tweak it, or handle it differently.

Getting the Voice Right

The hardest problem wasn't the context-gathering. It was getting the agent to write in a way that actually sounds like us. Generic AI-generated email is easy to spot — it's over-formal, over-enthusiastic, or both. People can tell.

We solved this by giving the agent access to our previous outreach and sent messages as reference. It reads what we've actually sent before — the specific way we open conversations, the language we use, how direct we are — and uses that as its style guide. The drafts sound like us because they're calibrated against what we actually write, not a generic instruction to "sound professional."

What Changed

• →Time per reply: from 10–15 minutes to a quick review of a ready draft
• →Inbox management: no longer a dedicated afternoon task
• →Response consistency: every candidate gets a thoughtful, context-aware reply regardless of how many came in that day
• →Spam and noise: automatically filtered, nothing irrelevant reaches the review queue

What We Learned

Keeping a human in the loop was the right call — not because the agent gets things wrong often, but because the edge cases matter. A candidate who replies with something unexpected, a message that needs a different kind of response than the agent anticipated — those are exactly the situations where you want a human reading before anything goes out.

The agent's job is to make that human decision fast and informed, not to eliminate it. That framing — use AI to compress the work, not to replace the judgment — is one we apply across everything we build.