AI for CRM and Follow-Up Automation Webinar May 7th

This session was about CRM from the ground up: what a CRM actually needs to track, how to build a very simple one with an agentic coding tool, and how to make it sync itself so no lead or follow-up ever falls through the cracks again. It is aimed at solo operators and small teams who keep customer knowledge scattered across five channels and their own memory. I keep noticing that the way most people structure their sales leaves a lot of money on the table, and this build is the fix.

Tools for the Day, and Why Codex on macOS (3:22)

The usual suspects were on the slide, Claude, Gemini, Perplexity, ChatGPT, but the build ran on the heavier tools: Claude Code and, for the first time in these sessions, Codex by ChatGPT. One thing I insisted on up front (3:55): the value right now is not in Claude Code or Codex specifically. It is in any agentic tool that can build code for you. I love Claude Code enough to have its stickers on my laptop, but if it ever becomes too expensive, you migrate. Do not marry the tool.

I ran the whole session from the Codex application for macOS (4:29), and it is insanely well executed. The team that built it worked heavily on native macOS implementations and was acquired by OpenAI, and it shows: the app integrates deeply with the system and gives you computer use at a very high level. The barrier that used to scare people away from these tools, the terminal, is collapsing. You now get a proper UI with the raw power of an agentic tool underneath, and alternatives like OpenCode exist too. Codex takes connectors just like Claude Code (7:03): I have mine wired to Fireflies, Notion, Google Drive, Vercel. It ships Fast mode inside the paid plan, where Claude Code charges extra for its equivalent, and it has scheduled automations that mirror Claude Code's scheduled tasks (7:53). Everything I did in this session works the same on Claude Code.

File Over App: Own Your Data (8:28)

Why these tools and not something like Manus? One reason: you are building on local files that you actually own. There is a short blog post called File over App that captures the idea. The application layer is not where the value is. The file, the data itself, is. Once you own the data, you can build any application on top of it, and that has never been more true than in the generative AI era, where a lot of SaaS is going downhill precisely because the app adds little over the data. With Codex and Claude Code you own the databases on your literal computer and even the conversations, stored locally in your workspace (9:05).

What a CRM Actually Does (10:12)

CRM stands for Customer Relationship Manager, a needlessly complex name for something essential. It does three things: it keeps track of leads, of successful projects, and of potential customers. My definitions (12:40): a lead is someone you already contacted, a potential customer is someone you have not contacted yet.

The Follow-Up Problem (11:22)

Who has never forgotten to contact someone who said reach out in three months? It happens all the time, and it is literally money left on the table. I confessed live: there was a guy waiting on a message from me for a week, and I only realized that day. In the old way of doing business, the customer relationship is locked in two places (13:02). First, the owner's brain: he knows every customer, where the relationship stands, who will pay. Second, the channels. Any professional has at least five: WhatsApp, SMS, calls, LinkedIn, and email. Five customers means twenty-five channels to track, ten means fifty. You never remember when you last contacted someone, on which channel, what you said, or what the next follow-up is. That is exactly the job of the CRM.

Run the numbers (46:24): most of us have at least five people we never followed up with. At a twenty percent close rate and an average ticket around two thousand, one forgotten follow-up is roughly two thousand left on the table.

Core Fields for a Simple CRM (14:52)

Keep the schema boring: status, last method of contact, next follow-up date, and whether the person is an active customer. The catch with CRMs (18:01) is that everyone treats their customers and their sales funnel differently, so a one-size-fits-all CRM is very hard. That is why I invite you to build your own, within the limits of what is possible.

Live Build: Let Codex Ask the Questions (18:45)

Inside my usual workspace I asked Codex to create a new one-off project called CRM from the ground up. Then came the trick that makes this work: instead of telling it to build a CRM, I told it we would build one together and that it should ask me as many relevant questions as possible about how I work with customers and handle leads. For the first version I capped it at ten questions, no automations, no integrations (20:01).

The questions were exactly right (20:21): the CRM is only for me, three contact types (cold lead, warm lead, current customer), what basic info every contact carries including messages and last contact method, one person per record, which statuses exist, what to track over time (last conversation, purchase history, next step, preferences), what interaction data to capture (channel, summary, sentiment), and which views I want: tables, a profile page, Kanban, and a dashboard.

While Codex built, I answered workspace questions from the chat (16:47): I keep one single workspace with subfolders for clients, my own projects, skills, and visual assets, plus Notion and Google Drive as the two external stores. I even keep my websites in there. And because I sync skills between Claude Code and Codex (25:02), when one tool goes down I switch to the other and keep building. Codex reads an agents.md file the way Claude Code reads a claude.md.

The Demo: A Working CRM in Fifteen Minutes (29:56)

Codex produced an extremely simple CRM that already looks a lot like the big commercial ones: who is a cold contact, who has been contacted, who warmed up, calls booked, current customers, and a list of next steps so I never forget what to do. Statuses actually change when you click. Under the hood there is no database at all (32:11): it is literally a JSON file on my computer. Your CRM can be a JSON or CSV file until you reach critical mass, then you migrate into something like SQLite or a real database. Just always keep a backup. Within fifteen minutes, one prompt and eleven questions, we had something functional, and that is good enough.

Keeping It in Sync Without a Web Deployment (30:50)

Samia asked the right question mid-demo: the page is static, so how do you keep it live without constantly telling the tool to update it? The answer for solo operators (35:01) is that you do not need a live automation running on the web. The sync can be a Claude or Codex skill that runs once or twice a day, at 8 AM and 8 PM. It fetches all the meetings you had and all the emails you received, hands them to Codex or Claude Code, and says: this is everything that happened with my clients today, sync the CRM according to what you see. Deal closed, status changes. Follow-up requested for Friday, date changes. If two messages conflict, take only the last thing the customer said.

To demo it, I added one real customer, then had Codex write a one-off script that used Gemini API calls to generate ten fake emails per client as threads plus one fake meeting transcript each (37:37). A few minutes later the CRM had a communication view per customer (48:15): email threads, meeting confirmations, proposals, onboarding check-ins, and the meeting transcript itself, all in one place. You already have access to all of this for real through your Fireflies, Fathom, or Gemini note takers and your email.

Automations as Code vs AI Skills (37:58)

Rohan asked whether automations can be done in Claude Code, which deserved a proper answer. In n8n or Make, you look at a visual workflow, but every node in those bubbles is just code executing in the back. So yes, anything you could build in n8n you can build with Claude Code. The real distinction is where it runs (40:06). Deploy on the web when you need live triggers, like an automation firing the moment an email arrives, or when multiple people work on a single operation. If once or twice a day is fine, a local skill with step one, step two, step three is equivalent and simpler to build. Solo operators live happily on skills.

Enrichment, Signals, and Nurturing (45:20)

Once the foundation syncs itself, you keep adding tracking features. With a service like Apify you can check whether a target has posted on LinkedIn lately, whether that post is a signal, and whether they need your services right now. Track the company on top of the person: are they hiring, are they doing well financially, have they invested in something? And nurturing (46:49) is a CRM job too: a message every two weeks with something genuinely useful. Relationships and business relationships are built on trust, and trust takes months. The heart of all of it stays the same: the follow-up mechanism and all customer data in one single place.

For channel coverage beyond email, Unipile (48:39) connects your Codex or Claude Code to LinkedIn and WhatsApp, so you can fetch a full day of messages and feed the same sync.

Q&A Highlights

Outlook connectors (49:48): Codex has an Outlook plugin that can read email and create drafts, but not send. Claude Code has an Outlook connector as well, so use whichever you want.

Can this replace GoHighLevel (51:03)? It depends what you use it for. If you rely on its automations, that is money well spent, because rebuilding to that level takes real time. If you only use the basic functionality for a nice interface, the account adds no value over what we just built.

Cold email (54:24): Tamir has been at it four months with good deliverability and weak replies, and his call to action was to jump on a call. It is normal nobody replies. They do not trust you, you are a stranger, and no one cares about a call. What worked for me is maximum value upfront, and podcast offers pushed through cold email worked fantastically for one client. And do not send unsolicited links (57:05), that hurts your deliverability. Ask instead: I built a personal roadmap of what a few automations could do for you, would you like to receive it? When the person says yes, they requested the link, and that is a completely different conversation.

The takeaway: build the must-have CRM foundation first, then stack implementations on top. As soon as you actually follow up with people, MRR goes up by a big margin. All past and upcoming sessions are on the webinars page.