Build a custom AI cold email agent when personalization quality is your actual bottleneck and you're prepared to own the pipeline. Buy an off-the-shelf tool like Instantly or Smartlead when your bottleneck is deliverability, warm-up, and sequencing — problems those platforms have already solved better than you will in a first pass. Most founders assume they need to build; most should start by buying and only build the one piece that's actually broken.
Strip away the marketing and there are really three separate jobs:
Off-the-shelf tools are excellent at (1) and (2). They're mediocre at (3), because true personalization means researching each prospect individually, and most platforms bolt on a generic "AI write this" button that runs the same prompt template for every lead. That's the gap a custom agent is actually solving.
We run our own outreach engine for exactly this reason. It scrapes each prospect's site with a self-hosted Firecrawl instance, feeds the cleaned page content to a local LLM, and asks for two things in one shot: pull out three or four facts worth mentioning, and draft the email using them. No separate "research agent" that hands off to a "writing agent," no planning step, no critique loop — one call, one output.
That decision wasn't arbitrary. We tested the multi-step version first: a chain that scraped, then summarized, then planned an angle, then drafted, then self-reviewed. It was slower, cost more per lead in tokens, and — this was the surprising part — the output quality wasn't better. Extra steps gave the model more chances to drift from the source material or hallucinate a detail that sounded plausible but wasn't on the page. Collapsing it to a single call with the full scraped context in the prompt produced tighter, more accurate emails for less money. If you're deciding between a single call and an agent chain more generally, the tradeoffs are the same ones covered in single call vs. agent chains — this is that pattern applied to outreach specifically.
Building this yourself means owning:
Realistic cost for a first version: a few thousand dollars of build time for scrape-extract-draft, plus ongoing LLM API spend that's typically cents per lead if you stick to a single call. That number multiplies quickly if you add retries, multiple model calls for "quality," or a vector database to store scraped context you don't actually need to search later — resist adding those until you've measured that the simple version is actually underperforming.
Instantly, Smartlead, Apollo, and similar platforms give you deliverability infrastructure that took them years to tune, plus sequencing, tracking, and unsubscribe handling out of the box. Their built-in AI personalization is real but shallow — usually one prompt per lead using whatever fields you've imported (name, company, title), not a per-prospect research step. For volume outreach where the message doesn't need to reference something specific about the prospect, that's often good enough.
Where it falls short is exactly the case where personalization is your differentiator — for example, if your pitch depends on referencing something concrete on the prospect's own site, their tech stack, or a problem visible in their public-facing product. Generic tools won't do that research for you.
There's also a compliance layer that doesn't disappear just because AI is writing the copy: unsubscribe handling, sender identification, and (in the US) CAN-SPAM requirements apply the same way to AI-drafted emails as to human-written ones — see the FTC's compliance guide if you're building or buying and haven't checked this.
The pragmatic middle ground, and what we'd recommend to most founders: keep an established platform for sending, warm-up, and sequencing, and build only the personalization step as a custom single-call agent that feeds its output into that platform via API or CSV. You get proven deliverability without rebuilding it, and you get the personalization depth that off-the-shelf AI features don't provide. This also caps your build scope to the one component where custom work actually pays off, rather than reinventing infrastructure that's already commoditized.
One more lever worth testing regardless of which path you pick: what you ask the prospect to do next. In our own tool, switching the call-to-action from "book a call" — a calendar link — to a soft "reply YES" consistently lifted reply rates. Scheduling friction kills otherwise-interested cold leads before they ever get to a call. It's a cheap test to run on top of whatever drafting approach you land on.
If you're weighing whether this is worth building in-house versus scoping it out, let's talk. Originally published on the Pykero blog.