New York School saves $20k by replacing SaaS with AI

Commercial Software: AI replaces three SaaS subscriptions

An IT manager at a private school in New York canceled three software subscriptions. His estimated savings: $20,000 per year. In their place, he built a complete ERP—billing, customers, suppliers, recurring invoices, accounting, multi-warehouse, stock control, tax profiles for Spain and the US—generated with ChatGPT and Codex within Visual Studio Code. It is used by 350 people daily. The question sweeping the commercial software sector remains unanswered: why pay for something an AI can build in-house?

What was built and why it’s doubted

The declared scope goes beyond a simple billing app. Besides the ERP, there is a network monitoring system that inventories equipment and tracks student usage, a campus access permission manager based on schedules, and a kind of operating system for an ESP8266 microcontroller: over 60 UNIX-like commands, cron jobs, a systemd clone, a web server with console, and a C-inspired programming language. All this, he insists, was written by AI in just a few months while handling other tasks.

The harshest criticism came from an unlikely source. A veteran with 28 years of SAP experience dismissed the core claim: this is not an ERP nor anything resembling one. This peine the rift running through the entire issue. There are ERPs for companies with 100,000 employees and management tools for businesses with 350. The word is the same. His thesis is that small businesses have been paying enterprise-level prices for management for years, and that this, more than the code itself, is what is changing.

An OS inside a 4 MB chip?

The most controversial part is the microcontroller project. A technical interlocutor asked where the kernel was, how everything fit into an ESP8266, and if he had written the 60 utilities by hand. The response clarifies things: the kernel is provided by the Arduino firmware itself; on top of it, UNIX behavior is emulated, and the tools live compressed in the file system on a 4 MB board. He accepts the correction: there is no kernel/userland separation or real processes. The code has been promised as open source.

The SME niche that never needed SAP

In Spain, he points out, there are millions of SMEs and freelancers who have never used SAP and never will. They needed invoicing, up-to-date accounting with the tax authorities (Hacienda), and little else. This niche was fed for years by products like Facturaplus and its successors. The starting argument is simple: according to his account, in twenty years supporting schools, every single client eventually requested a small adapted ERP, and none wanted to pay astronomical license fees.

The opposing side does not buy the thesis. The premise, summarizes a critic, is to pay and forget: focus on business, not code. For a company, adding, spending $50, $100, or even $200 a month on something so critical is cheap, and the expensive part is not programming but maintaining it. Something that works is not the same as something well-made, he insists. Especially when dealing with tax matters and having to answer to an inspection.

Zendesk loses a client, Odoo Community another

The examples are not hypothetical. One company just replaced Zendesk—a customer management platform with 7,000 employees—with an internally built AI application. Another describes how they modified Odoo Community using prompts, field by field, testing on a database clone, leaving the partner managing that installation without a client. On a more technical level, massive migrations are cited: Stripe reportedly moved 50 million lines of Ruby code with AI assistance.

The classic counterargument—if it’s free, who makes money?—hits an uncomfortable fact. Red Hat bills millions selling software whose code is open and free. The value was never in the file: it was in standardization, support, and problem mastery. Like McDonald's: anyone can make a burger, but no one replicates the cold chain or global logistics. Those who support this add that large language models fail more often than fairground rifles once they step off the beaten path.

How long can the subscription model hold?

The underlying debate is not technological, but about market structure. If the marginal cost of generating code tends to zero, the price of specific functionality collapses, along with the raison d'être of half the enterprise software industry. Moderates concede that these solutions work in small companies and that in large corporations, no one would gamble their accounting on a model. Aggressives respond that this "no one" is already replacing tools.

Here enters the most uncomfortable aspect. The manager works with minors under US educational legislation, coordinated with the New York State Department of Education. Student names cannot pass through ChatGPT. That data is still typed manually. AI has made cheaper the part that was never the bottleneck—writing code—and has not touched the part that truly scares: legal, fiscal, and liability issues.

With development costs plummeting and 350 satisfied users, supplier replacement should be a stampede. For now, it remains the exception talked about loudly. How many companies are doing the same and staying silent?