How I burned months and hundreds of dollars building a Salesforce replacement

How did one Claude invoice change the way I think about software?

The moment that cracked everything open was a Salesforce invoice. I had already spent $50,000 across five developers and multiple years building a CRM I genuinely loved — 7,000 contacts, 10,000 units, custom permissions, custom fields, transactions, leads. It was a chef's kiss, honestly. Then the invoice arrived, and I was staring at Anthropic Claude AI assistant on one screen and a bill I couldn't justify on the other.

That collision — a recurring SaaS cost I resented versus an AI tool I was already living inside — is what sent me down a 5-month build spiral I did not see coming.

What are Claude's code artifacts, and why did they shock me?

Claude's code artifacts is a feature inside Claude that lets you describe an application and receive a rendered, interactive prototype — not just code, but a live preview you can publish with one click.

I had been using Claude daily for months before I noticed the feature. At [13:32] I said: "I skipped over that left menu option and I clicked on it. And then there was a little dialogue box at the top… do you want to build a website? Do you want to build an app?" — that moment of accidental discovery is what started everything.

I typed one prompt: build me a Salesforce. Claude ran through its lines of code, and a CRM slid out from the right side of the screen. Contacts. Buildings. Transactions. Emails. Leads. A publish button. I remember saying out loud, "You gotta be kidding me" — three times.

Why did the first two builds fail despite looking so good?

This is the part nobody warns you about, and it cost me months.

Build 1 ran from September to October. The interface was functional. I published it. I clicked around. It had everything Salesforce had, visually. Then I asked how to import my actual Salesforce data — I had exported every file — and Claude told me something that stopped me cold.

The interface wasn't wired to a database. It looked like a CRM. It operated like a CRM. But nothing was actually reading from or writing to persistent storage. The data I saw on screen existed only in the rendered UI layer.

Build 2 ran from October to December and was even more polished. The buttons were better. The layout was cleaner. Same problem. I had exported everything from Salesforce, imported it into the new build, and the interface displayed it beautifully. But as Claude explained, the interface wasn't connected to the codebase or the database. It was, as I put it at the time, "just pretty on the outside."

You might also like

Pivot

Why my 80-minute live stream vanished and what I learned

Pivot

Why prompting skill beats coding skill in the AI era

Pivot

How I cancelled $1,800 in SaaS on day 24 of live vibe coding

Whiteboard

How I learned to think objectively after 10 years of unlearning

Here is how the two failed builds compared:

Both builds were scrapped entirely. I spent hundreds of dollars across them. I hadn't sniffed a real estate deal the entire time.

What did 12 hours a day of coding actually look like for a non-developer?

Build 3 started in December and pushed into February. This time I threw everything away and started over with a different approach. I was coding every day for 12 hours, learning GitHub documentation for version control from scratch.

I am not a developer. I was a real estate agent in 2009. I was making $32 for an eight-hour shift at a diner in West York before that. The idea that I would spend half a year learning version control, database architecture, and application structure would have been unthinkable to the person who couldn't make his own website as a real estate agent in 2014.

But that's what Build 3 demanded. And that's what I gave it.

How did personal development and faith lead me to AI engineering?

The path from Tony Robbins to Claude code artifacts is not as strange as it sounds, but it does take a while to trace.

I signed up for Tony Robbins' four-day Ultimate Experience in 2009 — I walked in wanting a refund and walked out with a growth mindset that ran my life for the next decade. I spent $25,000 on that program. I hired a weekend coach for $10,000. I paid $1,200 a month for a real estate coach for years. I read hundreds of books, attended hundreds of conferences, and saw hundreds of speakers across the US.

Then I hit what I now call the personal development hamster wheel. Marketing. Sales. Public speaking. Cold calling. Cold email. Physics. Psychology. Evolutionary biology. Jordan Peterson led me to Joseph Campbell. Dawkins and Hitchens and Sam Harris sent me into philosophy. Plato's Republic sent me into history. By 2021 I was deep in rabbit holes I couldn't name. By 2023 I was asking whether God exists.

That question got answered. I had a coming-to-faith moment. A few videos hit 100,000 to 200,000 views. Then I hit another ceiling.

By mid-2025, the ceiling was AI. I built Claude projects to replace my $1,200-a-month business coach and my $250–$300-a-month triathlon coach. Claude knew all my metrics. It printed out a weekly plan. That worked. Then I found the artifacts tab.

What does the original Salesforce build reveal about the real cost of custom software?

Before I tried to replace it, I had to understand what I was actually replacing. The Salesforce official help and product documentation describes a platform built for enterprise-scale CRM — and my build reflected that complexity.

5 developers. Multiple years. $50,000. Custom permissions. Custom fields. Conditional visibility. 7,000 contacts. 10,000 units. Transactions. Leads. It took a lot of planning, a lot of whiteboards, a lot of money, and a lot of frustration. Some developers did not deliver what they promised. I won't go into that.

The point is: I had a working system I loved, and I was paying ongoing licensing costs on top of the build cost. When Claude showed me a CRM in a single prompt, the math changed instantly.

What questions do builders ask when they hit the database wall in Claude?

Can Claude build a CRM that actually stores data? Yes, but not with the default artifacts workflow alone. The rendered preview Claude produces is a front-end interface. For data to persist — meaning it survives a page refresh and can be queried — the build needs a connected backend and database layer. That requires either an external service or a full-stack deployment, which is what Build 3 was attempting.

Why does the Claude-generated CRM look real but lose all data on refresh? The artifacts feature renders interactive HTML, CSS, and JavaScript in a sandboxed preview. Without a backend API and a database like PostgreSQL or SQLite, any data entered into the interface exists only in browser memory. Closing the tab clears it. This is the exact problem that killed Builds 1 and 2.

How much did it cost to attempt these builds? I spent hundreds of dollars across the first 2 builds before scrapping them. The original Salesforce build, by comparison, cost $50,000 in developer fees over multiple years. The Claude builds were cheaper by an order of magnitude — but they also consumed months of time and zero income from real estate during that stretch.

Is learning GitHub necessary to build a real AI-assisted CRM? Based on my experience, yes. Build 3 required me to learn GitHub from scratch while coding 12 hours a day. Version control became necessary once the project grew beyond a single-file prototype. Without it, tracking changes and recovering from errors would have been unmanageable.

What would I do differently starting Build 1 again? I would ask Claude to explain the database architecture before writing a single line of UI code. The interface question — what does it look like — is the wrong first question. The right first question is where does the data live, and how does the interface read and write to it. Answering that first would have saved at least one full build cycle.