Podscan's AI-First GPT-4o-Powered CRM Runs Through Slack for 20 Cents Per Day

Arvid Kahl uses a 6-hour timer and a large-language-model prompt to decide which trial users hear from him next for his startup, Podscan.

Podscan is a SaaS that indexes 3.8 million podcasts for brand mentions. Instead of blasting every signup, Kahl lets an LLM score intent behind the scenes and then sends the outreach himself.

A 6-Hour Timer Triggers a Background Investigator

The clock starts when a visitor creates an account. While the user explores, Podscan logs every action—pages viewed, search queries, podcast categories opened, API docs scrolled. At the 6-hour mark the system packages the footprint—email domain, name, team name, search themes—into a single prompt and calls OpenAI’s GPT-4o.

The prompt instructs the model to write a short brief first and only then emit an integer 0–10. Forcing the justification ahead of the number suppresses hallucinated scores; the model cannot reverse-rationalize a digit it has not yet produced. If the score is 5–10, the same call appends extra fields: likely projects and an ice-breaker.

High-scoring trials post a Slack message to Kahl’s private channel containing the JSON. Low-scoring trials are not contacted.

The Email Button Runs a Second Model That Knows the Product Cold

3 to 6 days later Kahl reviews the shortlist. When he clicks the button inside his admin dashboard, a separate prompt—this one “quite huge”—loads the user’s full activity graph and the earlier JSON.

The prompt opens with a markdown dossier that describes every Podscan feature. The model must propose the single next action that will make the user hit an “aha” moment inside the existing feature set.

Output is a plain-text email signed by Kahl. It references the user’s last search and suggests refined keywords. The message is copied into Hey.com and sent manually.

Onboarding Wizard Scrapes the Customer’s Website in Real Time

Kahl recently moved similar logic to the front of the funnel. During signup the visitor picks a role—founder, agency, podcaster, analyst—and can type a free-form goal. While the confirmation email is in flight, a fast model turns the three data points (role, goal, email domain) into 3 starter keywords.

If the domain belongs to a known marketing agency, the model fetches the agency’s homepage, extracts the top client names, and inserts those brands as suggested alerts.

Users who accept the AI suggestions complete their first alert at a higher rate than those who must invent keywords from scratch; Kahl does not give a multiplier.

Every Customer-Facing AI Path Is Gated by a Token Budget

Behind the scenes all generative calls route through middleware that increments a counter. If a user triggers “more than a couple” calls per hour the system can block the account. A second job compares rolling usage against a dollar cap; if the projected bill exceeds a threshold the middleware pages Kahl. He recalls that abuse can burn “thousands of dollars within a couple hours,” so he keeps the limit tight.

Backend scoring and email generation are exempt from the cap because they run under Kahl’s own key and never fire more than once per trial.

A 45-Minute Monologue Became the Product Specification

To keep the prompts accurate, Kahl once recorded himself clicking through every screen while narrating what each button does. The raw transcript ran 45 minutes.

He fed the file to Claude, asked for a concise functional spec, and received a markdown manual “a couple thousand lines long.” That document now prefixes every customer-facing prompt, ensuring the model never recommends a filter or workflow that does not exist.

When he ships a new feature he appends the delta to the manual and redeploys; no fine-tuning required.

Founders Can Copy the Playbook with One Webhook and Two Prompts

The stack is replicable in an afternoon. Hook your signup form to a background worker that collects domain, name, and declared intent. 6 hours later, send a prompt shaped like:

You are a data analyst. First write a two-sentence brief, then score this trial 0–10 for fit to [your ICP]. Return JSON: {brief, score, why}.

If the score is high, trigger a second prompt that holds a miniature product manual plus the user’s activity log, and ask for a plain-text email that nudges toward one concrete next step. Rate-limit everything on the way out.

Kahl spends “maybe twenty cents” on AI per day. The return is a sales workflow that feels white-glove yet needs no extra headcount.

Hat Tip to Arvid Kahl for the insight.

Top Tweets of the day

1/

This was said over a year ago. There are people who are operating content farms and printing money for cheap especially since Local LLMs have gotten decent.

2/

Every platform will soon use LLM rankings since they are much better than traditional algorithms. X is another addition which is testing it out.

3/

Video launch playbook for Gen Z which has most people scrolling your app on TikTok/Instagram before it gets to Gen-Xers or Boomers.

Rabbit Holes

1. How the ENTIRE internet turned on Duolingo overnight by Supreme Value

2. Prophecies of the Flood by Ethan Mollick

3. Proactively DM your Instagram Followers to Lift Revenue 4% by Ilke Akcasoy

What’d ya think of today’s newsletter? Hit ‘reply’ and let me know.

Do me a favor and share it in your company's Slack #marketing channel.

First time? Subscribe.

Follow me on X.

More Startup Spells 🪄

1. 10,000 followers with 1 post (LINK)

2. AI Caps as Pricing: How Anthropic's Claude Code and OpenAI's Codex Nudge Upgrades with Usage Limits (LINK)

3. Reddit SaaS AMA Comment Growth Hack: How Base44 Supercharged Engagement (in 24 hours) (LINK)

4. Power of Good Domains (LINK)

5. GummySearch's $10 Day Pass (LINK)