How my AI pitches my next 50 customers — and I approve them with one click

Letting an AI send cold outreach on your behalf sounds like a great way to torch your reputation in a single afternoon. One hallucinated detail, one pitch to someone you already emailed last week, one tone-deaf message, and you've burned a relationship you can't get back. So when I built outreach into my system, the design started from a hard rule: the AI does all the work, and a human approves every send. No exceptions. That one constraint is what makes the whole thing safe to run.

What the AI does (everything except hit send)

Point the outreach agent at a vertical — TV stations, podcasts, suppliers, real-estate media, senior centers, whatever the campaign is — and it runs the entire pipeline:

• Find real prospects in that category, with their actual published contact info — the address they put on their own site for this purpose, not scraped personal data.
• Research each one enough to write something specific — what they do, why this pitch is relevant to them.
• Draft a personalized pitch in the right voice for the brand doing the outreach.
• Dedupe ruthlessly against everyone I've ever contacted, with a cooldown so nobody gets pitched twice or hit again too soon.
• Queue it for me — a review file of ready-to-go drafts, each one a yes/no decision.

By the time it reaches me, the work is done. My job is judgment, not labor: read the draft, approve or kill it. That's the trade that makes AI outreach actually usable — it removes the 95% that's tedious and keeps the 5% that should never be automated.

The approval gate is a feature, not a limitation

It would be easy to let it send automatically. I deliberately don't, and I'm not planning to. The gate is where my reputation lives. It's also where I catch the occasional draft that's technically fine but just off — too eager, wrong angle, a prospect that isn't actually a fit. The AI is good; it is not me, and outward-facing messages with my name on them get a human read. Every time.

This generalizes across the whole system into a tiered model: routine internal actions run on their own, things that change configuration wait for review, and anything that leaves the building — an email, an external post — requires an explicit human yes. Autonomy where it's safe, a gate where it isn't.

From pitch to sale, tracked the whole way

Sending isn't the end — knowing what happened is. Every approved pitch gets tracked through its lifecycle: pitched, replied, approved, sale. That lives on a dashboard so I can see the funnel at a glance instead of digging through an inbox trying to remember who said what. The batch goes out on a regular cadence — a set day each week — rather than dribbling out at random, which also keeps the volume sane and the follow-up organized.

And the moment someone says yes, the system can fire a ready-made onboarding kit automatically — because that action (responding to an inbound yes) is low-risk and time-sensitive in a way cold outreach isn't. The gate is on the cold first-touch, not on warm follow-through.

Why this is the same philosophy as everything else I build

Readers of my other posts will notice the pattern. I don't trust automation to report its own success — I verify the live page. I don't trust a process to report its own health — a separate watchdog checks. And I don't trust an AI to be the final word on something that touches another human — I put a person in the loop. The thread through all of it is the same: let the machine do the volume, keep the human on the decisions that are expensive to get wrong.

That's also why it scaled across totally different verticals in a single day without me rewriting it each time — the pipeline is universal and the per-vertical part is just configuration. The same engine that pitches podcasts can pitch senior centers; only the targeting and the voice change.

Steal the model

You don't need my stack to use this. If you're going to let AI help with outreach, draw the line in the same place: automate discovery, research, drafting, and dedup — the soul-crushing parts — and put a one-click human approval on the actual send. Track replies to outcomes so you learn what works. Never let it auto-send a cold first message with your name on it.

That approval-gated design is built into the platform I'm working on — a self-hosted AI workforce that finds and pitches your next customers while you keep your hand firmly on the wheel. I'm building it in the open, and "it does the prospecting, you approve with one click" is one of the parts I'd hand to anyone tomorrow.