When SaaS Becomes a Prompt

My experiment does not prove the end of SaaS. It shows something more interesting: AI is shortening the distance between a requirement and working software—and separating replaceable features from genuine product value.

Twenty Minutes

I write only a few invoices, yet I still pay about €25 every month for invoicing software. I do not need it to do much: manage customer data, include the mandatory information, assign consecutive invoice numbers, calculate totals and create a clean PDF. So I asked ChatGPT whether it could recreate this as an Excel model. Twenty minutes later, the file was ready.

That does not replace a legal review. Whether such a file provides GoBD-compliant archiving, generates valid e-invoices, logs changes in an audit-proof way and offers reliable support in a dispute is a different question. But that is precisely the point: the visible bundle of features for which I had been paying a subscription was suddenly no longer sufficient product value. It could be described as a prompt.

The experience changes my willingness to pay. And it raises a larger question: what happens to software as a service when customers can generate a substantial part of the interface themselves?

The Stock Market as a Seismograph

Capital markets are already asking the same question. By the end of July 2026, Salesforce shares were down about 31 percent year to date; SAP's New York-listed shares had lost roughly a third. In early February, US software and data-services companies lost around one trillion dollars in market value within a few trading days, according to Reuters. The trigger was concern that new AI agents could devalue specialized applications faster than their vendors could renew their usefulness.

This is not a death sentence. Share prices do not prove that Salesforce or SAP will disappear. Both have data, customer relationships, integrations and enormous switching costs. But the sell-off shows that the market is reassessing an old article of faith: recurring revenue is durable only when the underlying benefit remains difficult to replace.

This fits Andreessen Horowitz's report from conversations with 100 IT leaders: companies are building more AI applications themselves because model APIs and coding assistants lower the barriers to entry. For SaaS vendors, that matters at least as much as a new competitor. Their customers are becoming developers—not everywhere, but precisely where the task is clearly bounded and the data is available.

Seven Test Cases in Publishing

Scholarly publishing deserves particular scrutiny. Many systems serve small markets, were conceived before mobile interfaces and sell historically accumulated process steps as separate products. AI can already reassemble seven such packages: 1. receive manuscripts and perform formal checks; 2. suggest and remind reviewers and structure their feedback; 3. prepare language, style and copy-editing; 4. check citations, references and potential integrity issues; 5. generate XML, metadata, abstracts and alt text; 6. pre-sort rights, licenses and permissions; 7. query CRM, marketing and usage data without clicking through dashboards.

These are not futuristic examples. Research and products already document citation checks, claim-evidence mapping, simulated peer review, automated technical checks and AI-supported data extraction. An NBER field study involving more than 5,000 customer-support agents found an average productivity gain of 14 percent, and substantially more among less experienced employees. AI is not merely replacing functions. It is reducing the value of the learning curve on which many SaaS vendors built customer retention.

In publishing, a small market no longer protects a vendor automatically. Previously, it was rarely worthwhile for a new competitor to reproduce a complicated niche process. A general-purpose model can now serve the same process as one of thousands of use cases. Human oversight remains indispensable: confidential manuscripts, fabricated citations and editorial decisions are not fields for unchecked automation. The opportunity lies in assistance and orchestration, not in abandoning responsibility.

What Remains Valuable?

AI will therefore not replace every application. My Excel model knows neither the dependable history of my transactions nor does it accept responsibility. It orchestrates no bank, tax adviser or archive. It guarantees neither standards nor availability. That is where the defensible part of SaaS begins.

The systems that survive will be more than collections of forms: reliable systems of record; proprietary, well-structured data; deep integrations; identity, permissions and audit trails; regulatory responsibility; network effects; and workflows whose result matters more than the individual click. AI does not make these properties obsolete. It makes them visible.

Pricing will change as well. A vendor charging per user per month while agents perform more work without their own seat is defending the wrong unit. Pricing based on usage, transactions or verified outcomes becomes more plausible. The vendor then sells not access to an interface but a reliable service.

Standards, Hosting and Workflow

For metadata and usage standards, AI broadens the task. As answers increasingly emerge outside publishing platforms, machine-mediated, syndicated and AI-supported usage must also be defined credibly. If usage remains synonymous with a human click or download, such standards will not lose their reason to exist—but they will lose some of their explanatory power.

For hosting platforms, the risk lies in the visible interface. When search engines and answer engines summarize content before anyone visits a publisher's site, page views and traditional discovery lose weight. The structured corpus, metadata, identities, access rights, commerce, availability and secure technical delivery remain valuable. The hosting platform of the future is therefore less the most beautiful website than the most dependable, agent-ready content infrastructure.

In submission and peer-review systems, forms, reminders and formal checks are easy targets. Confidential manuscripts, reviewer identities, conflicts of interest, audit trails, integrations and accountable decisions are difficult to replace. A large installed base offers protection—but only if the submission portal evolves into an open orchestration platform. AI-supported checks and editorial tools are a beginning. What matters is whether they create a better process rather than merely another button.

AI is not ending SaaS. It is ending the convenient equation that more features automatically mean more value. Anything that is merely an interface over a standardized process becomes replaceable. Anything that organizes trust remains scarce.

At my next contract renewal, I will therefore ask more than: What can this tool do? I will ask: What does it know, connect, prove and take responsibility for that I cannot generate myself in an afternoon? The convincing SaaS response to AI is not alarmism. It is a better product.