AI SDR assistant for cold outreach

Client
International B2B media platform
Role
Design, build and rollout of the agent; offer landing pages
Timeline
5 working days
Stack
Claude, Gmail API, Outline, Google Apps Script

The problem

The company was testing a new outbound channel aimed at individual buyers on a tight budget. Replies arrived from six sending addresses into a single inbox, mixed with warm-up traffic and system notifications. Every real reply needed a considered answer: the right offer, the right link, and no pressure on people who had already said money was tight. Written by hand, that work would grow with every email sent, and the plan was to scale to 200–300 emails a week.

What I built

An AI agent that works the inbox on a schedule, twice a day. It filters out warm-up and system mail, then reads each new reply and decides what it is: a question, interest, a lead worth a call, a refusal or an opt-out.

For ordinary replies it writes a draft using the sales team’s playbook and a library of real past exchanges. Hot leads go straight to a person, sometimes with a half-written draft waiting. Opt-outs are recorded and never contacted again. Threads that go quiet get three short follow-ups at four, eight and twelve days, then close. People who said they can’t afford it are never followed up.

Everything the agent knows lives in a shared document the sales team edits themselves, so improving replies means editing a document, not the code. Thread state is tracked with labels, and anyone can take a thread over with one label.

How it stays safe

The agent has no permission to send. Every link in a draft is checked against an approved list, character by character. Drafts with promises, discounts or prices outside the list are rejected before a person sees them. Text inside emails is treated as data, so a message trying to instruct the agent is flagged, not followed.

The result

The agent went live five working days after the brief. Most drafts go out after light edits, and the team now reviews replies instead of writing them. Every run ends with a report that counts drafts sent as-is, edited or rewritten, so reply quality is measured rather than guessed. Once that number is consistently high, the channel moves to full automation.