Automated cold outreach system
From target research to follow-up: a B2B lead generation process that works on its own every working day and only asks for human attention when it is worth it. Built for Matrix AI's own client acquisition, running in production.
- Client
- In-house build, for Matrix AI's own client acquisition
- Status
- In production since July 2026
Summary
Reaching small and medium businesses by cold email is research, copywriting and administration at once. A good email requires knowing the company, personalising the message, following up, watching for replies and handling undeliverable addresses. Done by hand this takes hours a day, and follow-up is usually the first thing to slip.
So we built our own automated system that organises the whole process around a single spreadsheet. Humans take part in research and in handling replies; everything else happens on its own: writing the email, scheduled sending, two-step follow-up in the same thread, reply detection, filtering dead addresses and watching the pipeline.
In one sentence: the system maintains at scale the quality a careful person could achieve by hand. Every email is built on a real fact observed on the company’s own website, yet no daily manual work is needed.
Where cold outreach fails
- Personalisation does not scale. Ten emails is fine; every day is not.
- Templates do not work. A generic email reads as bulk mail and gets deleted.
- Follow-up is skipped. A large share of replies only come after the reminder, but manual follow-up is the first thing to drop.
- Replies get lost. Someone has already replied and still receives the next reminder. That damages credibility.
- Domain reputation suffers. Repeated emails to non-existent addresses land the whole domain in spam.
The solution at a glance
The centre of the system is a Google Sheet. Each company is one row, and each row moves through a state machine. Separate Make.com scenarios handle each step on their own schedule.
| # | Step | What happens |
|---|---|---|
| 1 | Target research | Selecting the company from a database, filtering, and recording one real, observed fact from its own website. |
| 2 | AI writer | The language model writes a fixed-structure email and two follow-ups from the observed fact. |
| 3 | Scheduled sender | On working days between 9 and 15, one email every half hour, with a daily cap. |
| 4 | First follow-up | Three working days later, a short reminder in the same thread. |
| 5 | Closing email | A week later a polite, pressure-free close, then the system stops. |
| In parallel | Reply watcher, bounce watcher and a sentinel that warns before the list runs out. |
The key principles
Every email is built on a real fact. For each company we record something actually visible on its website that points to a recurring, client-facing administrative process: a form promising a callback, a note that prices change daily. If the email claims we observed something we did not, it fails on first reading. The system therefore never writes an email on an invented fact: no observation, no email.
The AI writes within a fixed structure. Five paragraphs: greeting, the observed fact in the indicative, the possible internal challenge in the conditional, a short concrete idea, then an offer of a free, no-obligation one-hour assessment. Every email states that the final decision stays in human hands.
Human-paced sending. At most one email every half hour, in working hours, from our own domain with proper SPF, DKIM and DMARC. Receiving systems treat sudden bulk sending as suspicious; human pace they do not.
Whoever replies gets no more reminders. The reply watcher matches replies to rows by thread ID every two hours and stops follow-up immediately. Proven in production: one reply arrived eleven minutes after the reminder, the system recognised it within the hour, and the closing email never went out.
At most one email to a dead address. The bounce watcher excludes undeliverable addresses every morning before sending, so the second and third emails never go out to them.
The sentinel warns in advance. If the queue of companies waiting to be contacted drops below two days of cover, it sends an email. It alerts when trouble approaches, not when it has arrived.
Results
- Since summer 2026 the system sends, follows up and watches on its own every working day; daily operation needs no manual intervention.
- Not a single respondent has received a reminder after replying.
- Since launch, no dead address has received more than one email.
- One outreach turned into an assessment meeting with a mid-sized regional company, where the managing director named in writing the process he wants AI support for.
Honest note: by the nature of cold outreach, most emails get no reply, and that is true of a well-functioning system too. So the system optimises not the number of emails but their quality and continuity: many small chances, evenly, credibly, and without a single lost reply.
Who this is for
Any business that serves other companies and wants to reach new clients regularly but has no dedicated sales team: B2B service providers, manufacturers and wholesalers, and companies where sales today lives in one or two people’s heads and inboxes. Rollout runs in six stages: assessment, target market and rules, email structure, technical setup, launch with a low daily cap, then follow-up and refinement.