The death of expensive SaaS: build your own with AI

Everyone talks about AI taking jobs. Almost nobody talks about AI taking the subscription.

Here’s the thing. For twenty years, the deal was simple: software is expensive to build, so you rent it. A CRM, an HR tool, a project board, a reporting dashboard, each with a login, a per-seat price and a sales rep who calls every January to explain why it costs more this year. That deal rested on one assumption. Building was hard.

That assumption just died. And a lot of very expensive software is about to find out.

The rent goes up every year. The value doesn’t.#

Software vendors have quietly become the most reliable inflation machine in your company.

Vertice, a firm that processes more than $75 billion of software spend, measured SaaS inflation at a record 16.4% in June 2026. In August it was still 12.7%. Your national statistics office would call a central bank that delivered that a failure. Software vendors call it a “pricing update”.

The big names do it in the open. Salesforce raised list prices for its Enterprise and Unlimited editions by an average of 6% from August 2025, the second increase in about two years. Microsoft followed with higher prices for its commercial Microsoft 365 suites from July 2026, Microsoft 365 E3 from $36 to $39 per user per month. Each time, the justification is the same: we added AI features. Did you ask for them? Doesn’t matter. You’re paying for them.

Compound interest is wonderful when you’re the one collecting it. With SaaS, you’re not.

You’re also paying for seats nobody sits in#

Guess first

What share of a company’s SaaS licences sits unused, on average?

The answer: 36%. Zylo’s 2026 SaaS Management Index, based on real spend data: an average of 36% of licences are unused, measured against recommended usage levels. More than a third of the bill buys nothing.

Zylo puts the median SaaS spend at $9,455 per employee per year. Take a 50-person company. That’s about $473,000 a year, and on Zylo’s numbers roughly a third of it pays for licences nobody uses. Shelfware. A gym membership for software.

And the stuff that is used? Usually 20% of the features, wrapped in 80% of complexity that exists because the vendor has to serve ten thousand other companies at once. You don’t run your business the way the software wants. You bend your business until it fits the software. Then you hire a consultant to bend it a bit more.

What changed: building got cheap#

AI didn’t make software free. It made the expensive part, writing and changing code, a fraction of what it was.

Retool surveyed 817 builders in late 2025. 35% of them had already replaced at least one SaaS tool with a custom build, and 78% expected to build more of their own tools in 2026. The first targets: workflow automations and internal admin tools, then reporting and even CRMs. Retool sells a platform for building internal tools, so take the enthusiasm with a pinch of salt. The direction is still clear.

The markets noticed before most IT departments did. In early February 2026, Anthropic released plug-ins for its Claude Cowork agent covering legal, sales, marketing and data work. On 3 February, Thomson Reuters lost nearly 18% in a single day, its biggest daily drop on record. RELX fell 14%, Wolters Kluwer about 13%, LegalZoom almost 20%. Traders named it the “SaaSpocalypse”.

Even the people selling the stuff saw it coming. In December 2024, Microsoft’s CEO Satya Nadella told the BG2 podcast that the “notion that business applications exist” could “collapse” in the age of AI agents, because “all the logic will be in the AI tier”. When the biggest software seller on the planet says that out loud, I listen.

The most famous example is also the most misquoted one. In 2024 Klarna said it had dropped Salesforce and planned to drop Workday, and the internet read “replaced with AI”. Klarna’s CEO Sebastian Siemiatkowski later set the record straight: “we did not replace SaaS with an LLM”. Instead, Klarna pulled data out of many SaaS systems into an internal tech stack of its own. Less sexy headline. Same direction. And he doubts that every company will copy it. Fair enough. Most companies don’t have to replace Salesforce. They have to stop paying for five tools that do one job badly.

What I build instead of renting#

I don’t just write about this. I build with AI every day, and more and more of what I use, I own.

My own Jarvis. One command center across all my companies, inboxes and software projects. It sorts my mail, drafts replies in my voice, briefs me in the morning, tracks what I promised to whom and shows me what’s broken across dozens of projects, on my Mac and on my phone. Nothing gets sent without my OK. The first working version took days, not quarters. Try buying that off the shelf. You’d need five subscriptions, three integrations and a project manager, and it still wouldn’t know how I work.

The aLIVE Cockpit. A management cockpit for the aLIVE companies. It started as a tool for public tenders and grew into the shared operating system of the business: tenders, finance, receivables, staff planning and a daily digest, on one database, with real roles. The classic route would have been a tender tool, a BI tool, a receivables tool and an HR tool, each with its own licence, its own login and its own idea of what a customer is. Now there is one system, and it fits the company instead of the other way round.

VEONIO. At my agency we automated large parts of website creation and marketing. A brief goes in, and a site that meets our house standard comes out, with a human check before any client sees it. Work that used to take weeks of hand-offs now takes a fraction of the time.

I won’t tell you how any of it is wired. Partly because that’s the competitive advantage. Partly because describing your security setup on a public blog is how you get to meet the people who read security setups on public blogs.

Hayek would have hated your CRM#

There’s an old idea behind all of this. In “The Use of Knowledge in Society” (1945), Friedrich Hayek argued that the knowledge that matters is spread out, the “knowledge of the particular circumstances of time and place”. A central planner can’t collect it. The people on the spot have it.

One-size-fits-all SaaS is central planning for your workflows. Somebody in San Francisco decided how your sales pipeline, your approvals and your reports should work, for every company at once. Custom software flips that. The people who know the job shape the tool, and change it the same afternoon when the job changes.

And then there’s lock-in. Your data sits in their format, the export is “on the roadmap”, and the renewal letter knows you can’t leave. That’s not a market. That’s a landlord.

What to keep renting#

Don’t burn it all down. Some software you should keep renting, happily:

Commodity infrastructure: email, cloud hosting, payments, identity. Somebody else carries the 3 a.m. pager and the security patches.
Regulated, rule-heavy systems: payroll, tax filing, banking. The rules change every year because politicians can’t leave anything alone, and you don’t want to maintain that yourself.
Network tools: where the value is that everybody else uses it, like a shared calendar standard or a marketplace.

Build the parts that make you different: how you sell, how you deliver, how you see your numbers, how your people work together. That’s where the expensive, mediocre middle of SaaS lives, and that’s the part that’s dying.

The fine print nobody puts in the hype#

Building your own isn’t free, it just moved the cost.

You own the bugs now. And the updates, the backups and the security. Budget for maintenance, not just for the build.
AI code needs review. Only 8% of builders in the Retool survey use AI-generated code without changes. Good. In a 2025 randomized trial by METR, experienced open-source developers were even 19% slower with early-2025 AI tools on large, mature codebases, while believing they were faster. Speed isn’t automatic. Judgement still matters.
Shadow IT is real. 60% of Retool’s builders built something outside IT oversight in the past year. One clever spreadsheet with an AI wrapper and no access rules is a data leak waiting for a date.
Start where it hurts. One tool, one painful renewal, one workflow you already work around. Not your whole stack in one go.

So who dies? Not software. The mediocre, overpriced middle layer: tools that charge enterprise prices for a database, some forms and a dashboard. Their moat was that building was hard. That moat is now a puddle.

The next time a vendor sends you a renewal with a 9% uplift and a slide about AI, do the maths. Then ask how long it would take to build exactly what you need. The answer might surprise you.

More on the money side of building something of your own: Your first side business, and why AI makes a great intern and a terrible oracle in AI and your money.

This article is education, not investment, tax or legal advice. No share mentioned here is a recommendation.

Sources#

Education, not advice. I don’t know your situation, and past returns promise nothing. Check my numbers, then make your own call. You’re a grown-up.

Expensive SaaS is dying, and almost nobody is talking about it