The Real Cost of AI

August 17, 2026   I’m testing our new AI capabilities before we officially launch, and it’s kind of mind blowing. It’s so much fun it’s a distraction…which means I might be spending too much money on Claude.

Here’s a real example that I had no intention of using today. Last night, I wanted to create a map of every publicly available loan in the US. I decided to use Fable, Claude’s top of the line model, given the complexity and scope of the project.

There are about 28k loans with a balance over $10mm, not in lockout, and closed more than two years ago. 

It. Was. Slow. Claude actually warned me that I was over on my monthly allotment of credits and that it still had 3 hours to go. I went to bed a little bit nervous about the bill I was racking up.

It. Is. Awesome.

Seriously, click on the link to try it out: Public Loans 

It even included cool filters you can use to narrow the search. 

I picked Occupancy < 85% and a < 5% rate maturing in 2027 and it whittled the list down to 306 deals. 

 It provided a table for me to investigate further. 

This is part of what makes AI feel so magical, right? But about that cost JP…

Turns out, it only cost me about $258. 

That led to me investigating Claude spend across the team. MTD at LoanBoss only, we’ve spent $4,106, which means we are on track to spend about $8,200 this month.

Here are some of our top spenders. These three teammates are on track to spend $5k of the $8k on Claude this month (not including our engineers’ Claude Code). 

• Sydney is Head of Product and has 3 Claude agents running around the clock to scan the database for issues, write product tickets, test new features, etc.
• Ben is Head of LoanOps and the loan abstraction and QA process flows through him. He spent two years building this AI abstract tool. It was a brutal process, but he’s got it to 99% accuracy on the nastiest of the nasty loans.
• Tom implemented the majority of our automation processes via n8n. Tom has zero coding experience and no technical background, He was hired as a junior relationship manager but just figured it out. I’m telling you, you don’t need consultants to help you do this stuff.

Obviously, it’s not nothing. And I didn’t have $8k per month in the 2026 budget for Claude. But as you know from my emails about reconfiguring the teams and layoffs, the Claude spend is less than 10% of the personnel cost it replaced. It sucks on a personal level, but it’s impactful from a business perspective.

Plus, those three are the absolute power users here. They have entire systems set up that run 24/7. Most users, including me, never bust through the monthly allowance of tokens. And when they do, it’s usually for a few dollars. 

Those $0’s don’t mean the teammate isn’t using Claude. It means they haven’t exceeded the monthly allotment.

• Standard seats are $300 per year and are what most teammates have
• Premium seats are $1,500 per year and are what the power users have

Cost is something to be mindful of, but the ROI is insane. But here’s the punchline…

The real cost of that cool public data set graphic isn’t the $258 Claude bill.

It’s the two years we spent building out the database of all Public deals.

Everything I was able to do was only because I had structured, clean, organized, and trusted data. Once that is done, AI makes the rest easy now.

Yardi, HubSpot, LoanBoss, etc all spent years and millions of dollars on learning painful lessons. Customers benefit from those lessons. Here are some of the ones I can think of off the top of my head about those public loans.

• Figuring out how to update automatically each month
• How to decipher the string of text about whether a deal is delinquent
• How to handle deals that got added or paid off
• Do we keep historicals or simply overwrite last month’s?
• How do we validate accuracy every single month?
• How do you store a couple hundred million records without bogging the system down?
• Cost for the database and the supporting team
• Etc etc etc

We have an entire data team committed to updating data every month. Loan balances change, rates change, property financials change, prepays change, properties are bought and sold, refinanced…data integrity amplifies AI…for better or worse.

Honestly, I wish I had never built it out. I started building, kept making tweaks and fixes and changes and before I knew it, years had passed and I was pot committed. I could pay some third party provider $50k a year for the next decade and never get close to what we spent building that.

Today’s example captures my entire argument around why SaaS isn’t dead, it’s just evolving. And just because I’m talking my book doesn’t mean I’m wrong.

Do you want to spend all your energy on building out, maintaining, updating, and repairing that underlying data?

Or doing the cool stuff like turning your data into an interactive map and finding a deal before anyone else does?

I personally want my $258 per month doing the awesome stuff.