The AI sales agent does not send outreach. It answers the reply - and only from what you gave it.

Sending the cold email is the easy half. The problem starts the moment somebody replies. The reply lands on Friday evening, a rep sees it on Monday, the context has gone cold, and a real opportunity turns into another row in the CRM.

The Leeeds AI sales agent covers exactly that gap. It is attached to one specific email in the sequence and wakes up only when someone replies to that email. It then drafts an answer from the material you connected yourself, attaches the sources it used, and hands it to you for approval. By default nothing leaves without a human.

In cold outreach, the single reply you get out of dozens of attempts is the valuable one. If a language model answers it by guessing a price, a lead time or a reference customer, you damage the relationship faster than if you had never answered at all.

So the agent runs on three fixed rules: it answers strictly from the connected material, every claim carries a pointer to the exact part of the source it came from, and when the answer is not in the material it declines to write it and hands the conversation to you. It will not buy confidence with a made-up fact.

The goal. Answer the question, check whether it fits (need, scope, timing, plus a summary for you), book a meeting, or a custom goal in your own words. The goal changes where the agent steers the conversation - not how freely it may treat facts.

The material. It is uploaded once at account level and shared across campaigns: a file, a link, a whole website, or plain text. A website is crawled up to fifty pages; PDFs are read page by page, with OCR for scanned ones. On each sequence step you then simply tick which sources that particular agent may use.

Approval. The default and recommended mode is “always require approval” - the draft waits, you can rewrite it, and only then does Leeeds send it. The second option is draft-only: the reply stays in the conversation and you send it yourself. Autonomous sending is not the default state and does not switch itself on.

Pace and length. The delay window can be set anywhere from 15 minutes to 48 hours. The agent either replies once and hands the thread over, or keeps the dialog going until the goal or the reply limit is reached (1 to 10, three by default). A checkbox keeps sending inside the campaign's active hours.

A dry run. The “Try it” tab takes a text you paste in as an incoming reply and shows the draft together with its citations and risks - without sending anything. A reply you consider correct can be saved straight away as new material.

Hard facts. Prices, lead times, capacity and guarantees are stored apart from documents - as a value with an expiry date. The agent may state a figure only from there and exactly as written. Once it expires, it stops saying it the same day. With no price on file, a pricing question still hands the conversation to you.

Calendar. For the booking goal the agent offers three genuinely free times, holds them until the contact answers, and on confirmation creates the event with an invitation - only after the reply has gone out, so a rejected draft never fills a diary. If you would rather not connect an account, you subscribe to a private link instead.

Scraping finds the companies, the sequence opens the conversation. The agent only matters at the point where there is something to save - when somebody replies and time starts working against you.

That is why we keep it deliberately narrow: one step, one trigger, explicit material, explicit approval. We would rather have an agent that sometimes says “I do not know this, take it over” than one that sounds confident and is wrong.