“Claude can build this.”
If you run a software company, you’ve heard that line by now. About a competitor. About one of your vendors. And eventually, usually late at night, about your own product.
If you lead the engineering team, you’ve probably had the CEO send you some version of it, followed by a question mark.
I’ve been building software for 17 years, and I’m still starting software companies. So I’ve had to answer that line for myself, with my own money and time on the line.
Here’s the uncomfortable part. It’s often true.
A while back, a team of three I was leading spent a month on a tool to analyse sales quotations. Backend, frontend, a designer. It still wasn’t done. Later I rebuilt the same idea on my own with AI in two or three days. To be fair to that team, my version proved the workflow and skipped a lot of what production needs. Still.
And then? That fast version reached a handful of businesses. That’s it.
So, is SaaS dead? Parts of it, yes. Mostly the kind of thing I rebuilt in three days. Whether your business sits in that part comes down to what your product holds for customers, and how they find you.
What AI can build, and what it can’t own
The way I’ve come to see it, every software product is three things bundled together:
The screen. The forms, dashboards and buttons people click. Usually priced per seat.
The rules. Some are the customer’s own, like lead routing, approvals and discount limits. Some belong to the world, like tax tables and payroll law, and somebody has to keep those current and stand behind them.
The record. The data that has to be right, with permissions, an audit trail, and someone who picks up the phone when it isn’t.
Honestly, AI can build a first version of all three now, database included. What it can’t do is own them. Somebody still has to notice when the tax rules change, fix the bank feed when a bank changes its API, and answer when the numbers don’t reconcile the night before a filing deadline.
So when a customer says “Claude can build this” about your product, and they mean it, they’re signing up to own all of that themselves. For a lot of screens and in-house rules, that’s a fair trade. For the world’s rules and a record their auditor trusts, I don’t think most will want it. (Sometimes they just mean a cheaper vendor could build it. That’s the distribution fight further down.)
The spend data points the same way. Ramp tracks the bills of more than 50,000 businesses. Across the 200-plus traditional software vendors it follows, seat contracts are still 65 to 75% of spend, and that’s barely moved in a year. Salesforce still took 60% of first-time CRM buyers.
The movement is in the layer on top. Granola, the AI meeting-notes app, went from almost nothing to 20% adoption in about 18 months. Ramp’s own summary is that AI-natives are spiking in workflows layered above systems of record.
Why I still build software companies
Partly because cheap code helps me as much as it scares everyone else. If you’re the one building, you can try a lot more things for the same money.
At Toshi, the AI calendar I co-founded last year, my co-founder runs the business side and I’m the only one building the product. Recent versions were 80 to 90% written with Claude Code, and that’s how one person shipped three different versions of the product in six months. I wouldn’t have managed that pace without it.
We also didn’t try to rebuild the calendar. Toshi syncs with the Google and Apple calendars people already have and builds on top. That leaves our real work where I think everyone’s is now: how people find it, and why they keep coming back. It’s early, and that’s the part I think about most.
That isn’t new, either. In 2021 I co-founded TinyWhale, business tools for tutors, coaches and teachers. Scheduling, payments, class bookings, reviews. By the table further down, that’s mostly screen. We hired five developers at the start and cut back to two by month three, because we were building things nobody had asked for. If we’d had AI coding tools back then, I suspect we’d just have built the wrong things faster.
After that we built alongside our customers. Within about a year and a half it had handled seven figures in bookings and payments, and it was eventually acquired. I can’t tell you what would have happened if someone could have cloned it in a weekend. What I know is that it grew mostly because educators told other educators.
AI made features cheap to copy and first versions cheap to ship. Getting found, trusted and hard to leave still takes about as long as it always did.
Distribution wins the first sale. Defensibility earns the renewal.
How exposed is yours?
Here’s roughly how I’d rate common kinds of software right now.
Very exposed (mostly screen)
Dashboards and reporting. Customers can build it, or ask an agent.
Forms, intake and booking pages. Unless payments or compliance ride on it.
AI tools in the layer on top, like meeting notes or sales assist. Customers come fast and can leave just as fast.
Partly exposed (the seats, not the record)
CRM and help desk. The seats and the replies get automated. The customer record stays.
Hard to replace (record and rules)
Accounting and the ledger.
Payroll. Record plus rules that change every year.
ERP and inventory. Agents get built on top of it.
And two questions I’d ask about anything you’re building now.
What does it hold that would hurt to move? If it went wrong and the customer shrugged, it’s mostly screen.
How do new customers find it, and why would they pick it first? Salesforce’s 60% of first-time buyers is what a good answer looks like at scale. Educators telling educators is what it looked like for us.
Your customers are going to rent the record and build the screen. Make sure you’re the record.
I might be wrong on timing, and the market might just be early. But here’s the exercise I’d run. List everything customers pay you for and write screen, rules or record next to each line. Be honest about how much of it is screen, because that’s the part Claude really can build. Then look at how your last ten customers actually found you.
If you’re the engineering lead who got that message, that list is a better reply than “it’s complicated.” And if you’re the founder who sent it, I’d still build. That’s what I’m doing.