I built an AI BDR agent. Here's what the data gap taught me.

I didn't set out to build a sales tool. I was trying to understand a pattern I kept seeing in RevOps consulting work — the amount of time senior BDRs spend doing things that feel like they should already be automated. Research. Personalization. Objection prep. Every rep I talked to described roughly the same 45-minute-per-prospect ritual, and none of it felt irreducibly human.

So I built the AI BDR Agent — a 4-step workflow that takes an ICP, researches a prospect, writes email + LinkedIn + call copy, and handles objections. I built it on free tools: Google News RSS, basic web scraping, and Claude Haiku via API. The whole thing runs as a Netlify serverless function. No paid data vendor. No scraping infrastructure. No B2B database subscription.

It worked. And then I hit a ceiling.

The free research stack — what it actually does

Here's the honest version of what the agent's research step does when you're not paying for data:

Google News RSS. I query https://news.google.com/rss/search?q="CompanyName" and pull the last 5 headlines. This is surprisingly good for companies that make the news — funding rounds, product launches, leadership changes. For the other 80% of B2B SaaS companies that operate quietly and don't do press releases, you get nothing.

Homepage meta scrape. I fetch the company's website and extract the meta description tag. This gives you one sentence of context — what the company chose to put in their SEO description. It's useful, but it's their marketing copy. It tells you what they want to be known for, not what's actually happening in their business right now.

Claude synthesises from what it has. Claude Haiku takes these two inputs and produces a structured research brief: company snapshot, inferred signals, likely pain points, conversation openers. It's genuinely impressive what Claude can infer from limited context. But here's the word that matters: infer. Claude is making educated guesses. The confidence scores I return from the API reflect this honestly — 40–65 when data is thin, versus 75–90 when there's real signal.

What the free stack cannot give you: a verified email address, a real-time signal that someone just posted 8 SDR job descriptions on LinkedIn, or the knowledge that a company is migrating their CRM and therefore actively evaluating adjacent tooling. These things exist in the world. They just don't live in public RSS feeds or meta description tags.

What actually changes when the data gets better

I collaborated with Bitscale to run an experiment: take the same agent, the same Claude prompts, the same 4-step workflow — and plug in enterprise-grade enrichment at the research step. Here's what I observed.

The email problem goes away immediately. The single biggest friction point in outbound is not having a verified email address. With the free stack, the agent produces great copy and then the rep has to go find the email manually. With an enrichment waterfall (Bitscale sequences 18 providers until one returns a verified hit), this disappears. It sounds trivial. It's not. At 50 prospects a week, that's hours of lookup time every week per rep.

Signals shift from inferred to detected. "This company is probably hiring" (inferred from news) versus "this company posted 8 SDR job listings on LinkedIn in the last 14 days" (detected from live data) are the same surface-level signal. But they produce fundamentally different outreach. The second version lets you write: "Noticed you're scaling your outbound team — the timing felt relevant." That sentence only works if the signal is real. If you're inferring it, you're guessing, and the prospect will feel that.

ICP scoring changes how you prioritise. When you're running 10 prospects manually, you can eyeball fit. At 100+ prospects per day — which is where AI agents start to matter — you need a score. The free stack produces a rough confidence rating. Enterprise enrichment adds a proper ICP fit score based on your criteria: company size, tech stack, recent hiring patterns, funding stage. The top 20 prospects in an enriched list are meaningfully different from the top 20 in an inferred list.

The bigger question: is data the moat, or is the workflow?

I've thought about this a lot, because it affects how I think about what we're building at Unwind GTM.

The workflow is replicable. Anyone can build the 4-step flow I built — ICP setup, research, outreach, objection handling. The prompts are good, but they're not magic. A competent engineer with a weekend and Claude access could reproduce the structure. The barrier is low and getting lower as AI tooling matures.

The data is harder. Building a reliable email waterfall from 18 providers is years of work. Training signal detection models on live job posting data, LinkedIn activity, SEC filings, and CRM intent signals is an infrastructure problem that requires real investment. The companies doing this well — Bitscale, Apollo, ZoomInfo, Clay — have compounding advantages that are hard to close.

This doesn't mean building on free tools is wrong — it's how you validate quickly and without cost. But it means being honest with yourself about what you're building. A free-tools AI BDR agent is a workflow prototype. An enterprise-data AI BDR agent is a product. The gap between them is the data layer.

What this means for founders building outbound AI

A few things I'd tell myself if I were starting this build today:

Start free, but know the ceiling. Google News + homepage scraping + Claude is a legitimate MVP. It proves the workflow. It gets you 60-70% of the output quality at 0% of the data cost. But plan explicitly for the moment you hit the ceiling — and know what that ceiling looks like before you hit it.

The data gap shows up in the output, not the architecture. The code structure of my agent with free tools is identical to the version with enterprise enrichment. Same functions, same prompts, same step structure. The difference shows up in what Claude writes. That's a subtle distinction, but it matters: you can build the right architecture first and layer in better data later. The workflow doesn't need to change.

Inferred signals erode trust over time. The first time a rep sends "I noticed your company is growing" with nothing concrete behind it, it's fine. The tenth time the same rep reads an email that says "I noticed your company is growing," they stop trusting the output. Signal specificity is what makes reps believe the AI is actually helping rather than generating plausible-sounding noise.

The objection handling layer is underrated. Most people focus on the outreach generation step. The objection handling step — which I built as a separate Netlify function with playbook-driven responses per objection type — is where I've seen the most adoption among reps I've shown it to. It's lower stakes to test (no sending involved), immediately useful, and reveals something real about whether the AI understands the sale. Build it as a first step if you're skeptical about AI outbound.

The experiment

I've built an interactive demo that lets you see both modes side by side. You define your ICP, then watch the same 10 Indian SaaS prospects get enriched — first in free-tools mode (Google News + homepage scrape), then with enterprise enrichment (Bitscale's 18-provider waterfall). You can see the signals that appear, the emails Claude writes from each dataset, and the difference in output quality.

It's not a sales pitch for Bitscale. It's a concrete demonstration of what the data gap looks like when you run it on real prospects. Form your own view.

---

Aishwarya Agarwal is the founder of Unwind GTM, a GTM consulting and AI tooling practice for B2B SaaS companies. The AI BDR Agent is free to use at unwindgtm.com. The Bitscale collaboration is an independent experiment — no commercial relationship exists at the time of publishing.