The ChatGPT Honeymoon Is Over

When ChatGPT launched, a wave of lead gen agency owners did the same thing: they started using it to write outreach emails. They’d paste a prospect’s name and company, ask for a cold email, and send the output. Some got a brief bump in replies. Most saw results plateau within a few weeks — or worse, prospects could tell the emails felt templated and reply rates dropped. The “I thought this would work but it didn’t” moment usually came when they tried to scale: the more they used ChatGPT, the more generic the outputs felt, and the less their best performers wanted to use it.

Here's why that happened: ChatGPT is a general-purpose tool trained to be helpful to everyone. Your lead gen agency is not a general-purpose business. It has a specific niche, a specific way of qualifying leads, a specific voice, a specific definition of a "good reply" — none of which ChatGPT knows about by default.

Using ChatGPT for outreach without customization is like hiring a copywriter and giving them zero context about your client's business. You'll get something. It won't be good enough.

What "Custom AI" Actually Means for Lead Gen

Let me be specific because this term gets abused.

A custom AI tool for a lead gen agency is not:

A custom ChatGPT prompt you paste in every morning
A Zapier workflow that sends text to an API
An AI that "learns" in some vague future-tense way

A custom AI tool for a lead gen agency is something that embodies the workflows and the core profit-making steps of the client. This can be in the form of anything — it can be a workflow or it can be an app or a website — but it needs to mimic what has been done manually by the client over the years.

The difference is the same as the difference between a generic employee and someone who's been working in your agency for 6 months and knows your clients, your voice, your standards.

The Lead Gen Workflow Breakdown: Where AI Actually Helps

Not every part of lead gen benefits equally from AI. Here's where the real leverage is:

Lead Sourcing and Enrichment

Finding leads is table stakes. Enriching them — pulling context that makes outreach feel researched — is where most agencies still do things manually.

For the client I built for, enrichment meant: LinkedIn activity (recent posts, job changes), company news (funding, product launches), job postings (hiring signals), and website copy (positioning, ICP signals). We pulled from public APIs and scraping, normalized it into a single context object, and fed that into the personalization step. The manual version was a VA spending 15–20 minutes per lead. The automated version: under 30 seconds per lead once the pipeline was running.

Personalization at Scale

This is the part everyone talks about but few actually solve. The challenge isn't generating personalized text — LLMs can do that. The challenge is generating text that sounds like it came from a human who actually did the research.

We solved it by giving the LLM structured context (the enrichment output) plus 3–5 few-shot examples of the client’s best-performing emails — not templates, real emails that had gotten replies. We also added tone guidelines (“direct, one short paragraph, no hype”) and a rule to always anchor the first line in a specific signal (e.g. a LinkedIn post or a job posting). The output felt genuine because it was grounded in real data and mirrored his voice. A sample opener looked like: “Saw you’re scaling outbound in EMEA — how are you handling lead scoring across the new team?” — tied to a real job posting we’d pulled.

Reply Handling and Follow-Up

The client’s tool didn’t send replies automatically; that stayed human. What we did add was a reply triage layer: incoming replies got summarized and scored (interested / not interested / need more info). The team could see suggested next steps (e.g. “Send case study” or “Offer a call”) so they could respond faster. The human stays in the loop for anything that could turn into a deal — AI handles the sorting and the first draft of follow-up options.

Reporting and Optimization

We didn’t build a full AI analytics layer for that first version, but it’s practical now: subject-line performance, reply patterns, and which enrichment signals correlate with replies. What’s possible: AI can surface “these three openers are outperforming” or “replies drop when you mention pricing in the first email.” What still needs human judgment: interpreting why and deciding what to test next. So AI-assisted analysis is valuable; full autopilot on strategy is not.

Real Numbers From a Real Client

This is from the lead gen client I mentioned. I'm sharing what I can without compromising their privacy.

Before the build:

~50 leads processed per week
~50 minutes per lead on research and first email
~20% reply rate baseline
2 people full-time on research and outreach

After the build:

500+ leads per week with the same core team
Effective time per lead under 2 minutes (review and send)
Reply rate held at ~20% while volume went up
Team focused on closing conversations and on productizing the tool for other agencies

One unexpected outcome: the client started getting inbound from other agency owners who’d heard about the tool. That accelerated his decision to package it as a product — we hadn’t assumed that would happen in the first six months.

The Productization Angle (This Is the Real Opportunity)

The thing most lead gen agency owners miss when they think about custom AI tools: the tool itself can become a product.

My client isn't just using the tool internally. He's planning to sell it to other agencies in his space.

Think about what that means: he paid for a project to build something that now has the potential to generate recurring revenue from other agencies. The ROI calculation completely changes.

Productization is realistic for agencies whose workflow is repeatable and valuable to others in their niche — lead gen, email marketing, recruiting. The gap between “internal tool” and “sellable product” is packaging: onboarding, docs, support, and pricing. We don’t turn every project into a product, but when the use case is general enough that others would pay for it, we build with that architecture from day one so the client isn’t stuck rewriting later.

This is something I actively look for when scoping projects for agency owners. If the tool we're building is general enough that others in your niche would pay for it, we should be building toward a productizable architecture from day one.

"Can't I Just Use Clay / Instantly / Apollo?"

Yes, maybe. Let's talk about the specific tools:

Clay: Great for enrichment and data stacking. Where it falls short: complex custom logic (e.g. your proprietary scoring or multi-step personalization) gets messy fast. You’re often fighting the UI and the node limits. Fine for testing; limiting when your workflow is your moat.

Instantly: Good for volume sending, warmup, and deliverability. Where it falls short: the AI personalization is shallow — it’s template-plus-fields. If you need deep, research-backed first lines that reference specific prospect signals, you hit the ceiling quickly.

Apollo: Strong database and sequencing. Where it falls short: the built-in AI is generic. It doesn’t know your voice, your best-performing patterns, or your niche. Fine for basic sequences; not enough when differentiation is in how you write and qualify.

The pattern I see: these tools work great up to a certain point of sophistication. The moment your workflow requires logic that the tool didn’t anticipate — your custom qualification criteria, your proprietary data source, your specific voice — you're hacking around the tool's limitations instead of building something that fits.

When to Build vs. When to Buy

Build custom when:

Your outreach logic is genuinely differentiated (it's your competitive advantage)
You want to own the system — no dependency on a SaaS pricing change
You're doing enough volume that the ROI on build time is clear
You're planning to productize

Stick with off-the-shelf when:

You're under 30–50 leads per day and still validating what works
You're still figuring out what works in your outreach
The existing tool does 90%+ of what you need

One nuance: some teams need a hybrid — e.g. Clay for data and a custom layer on top for personalization and routing. And sometimes the right move is to fix the process (e.g. define what “good” looks like) before automating; we’ve seen builds fail when the client hadn’t nailed that first.

The Starting Point Is Always the Same

Every custom AI build I do for a lead gen agency starts with the same question: "Walk me through what your best outreach looks like — not the template, the thinking behind it."

That conversation usually takes an hour. By the end, I have enough to spec a system that replicates that thinking at scale.

If you want to have that conversation, the audit is the place to start.

Book the free AI audit

Shubham builds custom AI systems for agency owners at Maximal Studio. He's shipped tools for lead gen and email marketing agencies, and he's currently taking on 2–3 new projects per month.

Why Your Lead Gen Agency Needs a Custom AI Tool (Not Just ChatGPT)