Will AI eat SaaS for lunch?

BACK in 2011, American billionaire and venture capitalist Marc Andreessen famously declared that “software is eating the world”.

The ensuing 15 years proved him prescient.

Recently, a Citrini Research post on Substack grabbed headlines and triggered a market sell-off.

Citrini’s central thesis: a cannibalistic last feast where artificial intelligence – such as Anthropic’s Claude Code – devours the software industry.

The argument is simple: if anyone can prompt a large language model (LLM) and vibe-code a custom customer relationship management or enterprise resource planning system in an afternoon, the multi-billion-dollar Software As a Service (SaaS) industry becomes a dinosaur overnight.

It’s a frighteningly plausible thought experiment that put a spotlight on Citrini and its author James van Geelen, but it is fundamentally naive because it assumes a world without friction.

Prototyping is not production

AI tools such as Claude are incredible at instantaneous prototyping. But as any software engineer in the real world knows, writing code is 10 per cent of the job; maintaining, scaling and debugging it for edge cases is the other 90 per cent.

Software isn’t just a static pile of logic; it’s a living organism. Production-strength software requires auditability, 24/7 reliability, application programming interface stability and security compliance, all at scale. These are not things that LLMs, stochastic – that is, having a random probability distribution or pattern – by nature, can replicate.

The “AI eats SaaS for lunch” logic naturally leads to the conclusion that all firms will build all systems in-house, because they now can – with the help of AI. But will they, really?

I bet they will not for two reasons.

• The compliance shield

One primary reason companies buy enterprise software is to transfer risk.

When a Fortune 500 company uses a specialised SaaS provider for cybersecurity or HR, they are not just buying code; they are buying SOC2 (System and Organization Controls 2) compliance, General Data Protection Regulation (GDPR) indemnity, and ISO (International Organization for Standardization) certifications.

If firms build systems in house using general-purpose tools such as Gemini or Claude, when things go wrong and data leaks, there will be no vendor to sue, no platform to blame, and no security patch to buy.

• Scale and interoperability

A third-party provider spreads the cost of high-level security and compliance across thousands of customers. An individual firm trying to replicate that in-house would find the “efficiency” of AI quickly eaten up by the massive overhead of self-certification and liability insurance.

If all firms build all systems in-house, we will be back in the world of fragmented software, lacking interoperability – recall the legacy systems. Firms may well end up with isolated legacy piles of AI-written codes that no one understands, with the added benefit of 100 times the speed.

The auditor’s nightmare or bonanza

If every company builds their own “bespoke” AI systems, the complexity of auditing those systems becomes exponential.

If this “built-in-house” trend were to take off, auditors would become the most needed and sought-after profession on earth.

If van Geelen really believes what he says, then he might consider auditing as a new profession to hedge against the apocalypse he hypothesised.

It might be the answer to the job displacement problem.

Let’s not forget that AI is a general-purpose technology, while software companies are specialists.

It is hard to argue that generalists will replace all specialists in the modern world, which is essentially what Citrini’s scenario argues.

Everyone can buy a shovel. Not everyone shovels their own snow.

Hyper-scalers’ Achilles heel and SaaS’ true moat

The “AI eats everything” narrative assumes that LLMs have access to all the world’s intelligence.

They do not – yet.

There is the paywall barrier: the most valuable data – the “low-entropy” information required for professional-grade decisions – lives behind the moats of LexisNexis, Thomson Reuters, Nielsen and the like, as well as every company’s own data repository.

Without access to such data, generic models cannot generate deep insights; they risk “informational heat death”, where they simply recycle public-Web noise. The owners of the data, not the owners of the models, hold the ultimate leverage.

What AI will do to SaaS

SaaS will not be eaten by AI – but it will be shaken and stirred by it.

Gone will be the cosy habit of seat-based pricing and 5 per cent annual price increases based on new feature releases no one asked for.

Companies will be more discerning, and whether they will actually do it or not, they can wield a credible threat to take business away from a SaaS provider.

That threat will shape up the SaaS industry.

Those who can deliver measurable value will survive and thrive, while those who cannot will disappear.

SaaS as an industry will emerge leaner and better.

The writer is a professor of finance and dean of research and innovation at Insead