How We Built a Custom GPT That Books 40% of Our Discovery Calls

If you run a marketing agency, a service business, or any company that depends on discovery calls to close deals, you've probably wondered the same thing we did: what if the first conversation with a prospect didn't have to involve a human?

Six months ago, we built a custom GPT — essentially a specialized AI assistant — to handle our intake calls. It now books roughly 40% of the discovery meetings we schedule, qualifies leads while we sleep, and has saved our sales team roughly 14 hours of work every week. This is the honest, no-fluff walkthrough of how we built it, what went right, what broke, and what we'd do differently.

Why we built it (and why most agencies should not)

Before I share the playbook, a warning: this only works if your sales process is well-defined. If you don't know what makes a good lead vs. a bad lead, or your team disagrees on qualification criteria, no AI will fix that. It'll just produce confusion at scale.

For us, the trigger was simple math. We were receiving ~80 inbound enquiries per month. Our two-person sales team could only meaningfully respond to about 30. The remaining 50 either went cold or got auto-replies that converted poorly. Real money was leaking from the funnel.

The question wasn't whether AI could replace our salespeople. It was whether AI could handle the parts of the conversation our salespeople didn't have time for.

The architecture (in plain English)

Here's what we actually built, with no technical jargon:

1. A custom GPT hosted on OpenAI's platform, trained on our service catalogue, pricing logic, past discovery call transcripts, and qualification framework.
2. A chat widget on our contact page that connects to the GPT via API.
3. A Calendly integration so the GPT can book meetings directly into our team's calendar — but only after specific qualification thresholds are met.
4. A Slack notifier that pings our sales lead the moment a meeting is booked, with a summary of what was discussed.

That's it. No fancy RAG setup, no vector databases, no agent swarms. The hardest part wasn't the technology — it was the prompt engineering and the qualification logic.

The prompt we obsessed over

We rewrote the system prompt 23 times before we shipped it. The early versions were too friendly (too many meetings booked with bad leads), too pushy (visitors abandoned the chat), or too vague (the GPT made up pricing).

The version that actually worked has four key sections:

• Identity and tone: Who the GPT is, how it should speak (warm, direct, slightly informal — like a knowledgeable colleague).
• Qualification framework: The specific questions it must ask — about budget, timeline, current marketing setup, what success looks like.
• Disqualification triggers: Clear rules for politely declining to book a meeting (budget below threshold, looking for free advice, location mismatch).
• Hard guardrails: Never quote specific pricing without confirmation. Never promise outcomes. If the visitor asks something off-topic, redirect.

What broke in the first month

Three things went wrong almost immediately:

1. The GPT started inventing case studies

When prospects asked "have you worked with brands in [industry]?", the model would confidently mention clients we'd never worked with. We fixed this by adding an explicit rule: only mention case studies from the approved list, and if asked about industries not on the list, say so honestly.

2. It booked meetings outside business hours

India has 28 states and 8 union territories, and visitors came from all of them. The GPT scheduled calls at midnight because Calendly's "available hours" weren't being respected. The fix involved adding explicit timezone logic to the prompt and tightening Calendly's availability rules.

3. Conversion dropped on mobile

The chat widget worked beautifully on desktop. On mobile, visitors got frustrated because the keyboard kept covering the chat input. We rebuilt the widget UI to use a full-screen mobile mode with auto-scroll on input focus.

The results, six months in

Here's what the numbers look like today, with placeholders rounded for clarity:

Metric	Before	After
Meetings booked / month	~32	~58
Average response time	4.2 hours	3 seconds
Sales team hours / week	22 hrs	8 hrs
Show-up rate to booked calls	61%	73%
Lead-to-customer conversion	18%	21%

The most surprising finding wasn't the volume increase — it was the show-up rate. Prospects who had a longer pre-call conversation with the GPT were significantly more committed to actually attending the call. They'd already invested 7-10 minutes engaging with our process; they weren't going to ghost.

What we'd do differently

If we were rebuilding this from scratch today, three changes:

• Start with the disqualification logic, not the qualification logic. It's more valuable to confidently turn away bad leads than to enthusiastically engage every lead.
• Use Claude or Gemini, not just GPT. Different models have different strengths. We now route certain conversation types (technical questions, pricing logic) to different models based on what they're best at.
• Build a feedback loop from day one. Our sales team rates every meeting that came from the GPT on a simple 1-5 scale. That data flows back into prompt improvements quarterly.

Should you build one?

If you have more inbound enquiries than your team can handle, if your qualification criteria are clearly defined, and if you have someone on your team who can iterate on prompts every week — yes, this is genuinely transformative. The build took us 11 days. The cost (OpenAI API + Calendly + our time) is under ₹4,000 per month.

If your sales process is still informal, or if your team can comfortably handle all your inbound leads — wait. AI is a multiplier. If the thing you're multiplying isn't working, multiplying it just creates more of nothing.

Either way: experiment. The barrier to building useful AI tools has collapsed in the last 18 months. The agencies and startups who'll win the next decade are the ones treating these as core infrastructure, not curiosities.