Generative AI Will Not Replace Traditional Software – But Must Merge With It

The release of a legal workflow plugin to Anthropic’s Claude Cowork product last month erased roughly $285bn from software equities before a partial recovery. Even before this, predictions of the “death of SaaS” abounded as progress on agentic AI reached a new inflection point. In Citrini’s now viral research report scenario, “a competent developer working with Claude Code or Codex could now replicate the core functionality of a mid-market SaaS product in weeks”.

Surely the writing is on the chat interface for traditional software?

I don’t think so.

My position has little to do with frontier lab strategy or the daily gyrations of the market. It’s a technical argument about the fundamental nature of large language models.

Generative AI changed our experience of technology

Before we were beguiled by ChatGPT, we lived in a smaller, more predictable software world. Applications ran on bounded databases. Interfaces were quiet boxes rather than garrulous chatbots. Actions followed structured workflows. We had never heard of hallucinations – at least not in a systems context.

That era now feels bland and constrained – so very 2021.

But it had virtues.

Actions were predictable. Data was structured. Errors were traceable (and usually fixable). Applications were good at calculations. Systems did not resist being shut down. It was an ordered – even mundane – technical world.

But now, Iceman has been replaced by Maverick, and we are told not to worry about aberrant behaviours: hallucinations will be engineered out; scaling will yield self-correcting intelligence; and guardrails will protect users (although commitment to safety varies dramatically across labs).

Undoubtedly, there will be progress on these fronts. But here’s the problem: the generative leopard cannot change its probabilistic spots.

Uncertainty is not a flaw in transformer architectures – it is intrinsic to their design. LLMs generate outputs by sampling from probability distributions across high-dimensional latent space. That generative flexibility is their power – but also the root of their instability.

LLMs rarely do exactly the same thing twice – even under the same conditions.

Going back to the future

In contrast, enterprise systems are built on deterministic foundations. This is because payroll cannot be creative about tax rates; billing engines cannot improvise compliance; and supply chains cannot “hallucinate” inventory.

Traditional software is still essential because institutions sometimes require certainty.

However, what if we deliberately construct hybrid architectures that can follow rules, perform accurate calculations, encode models of the real world — but also chat, create, and adapt to uncertainty?¹

We can. But for some reason, we don’t seem ready to go this route yet.

It is already feasible to add generative capabilities to structured systems and call bounded functions from LLMs. Yet the industry narrative remains polarised: frontier labs want to scale their single architectures toward “AGI”, while incumbent SaaS vendors only cautiously bolt copilots to their legacy stacks.

I think both camps need to broaden their horizons and pivot hard strategically.

Otherwise, frontier labs will hit diminishing returns from scaling; safety concerns will slow the deployment of complex agents (CEOs are already saying “no” to unconstrained autonomy); and productivity gains will be absorbed by error-correction efforts. In parallel, SaaS will start to look and feel dated, driving users to take unnecessary leaps into the AI dark.

This is all avoidable because the real opportunity is not the replacement of enterprise software by generative AI; it is the integration of deterministic offerings like SaaS with their probabilistic generative counterparts to form multi-dimensional (neuro-symbolic) platforms.² Think of Iceman and Maverick sitting side by side in the F-111 Aardvark.

Integration is a mindset

However, humans are notoriously bad at “best of both” thinking. We prefer a single lens, then lock into what we see, excluding other perspectives. This is why empiricists and rationalists fought for centuries before combining their insights into what we now call science – one of history’s most powerful intellectual fusions. Imagine if it had been achieved two hundred years earlier?

On AI, we do not have centuries. Perhaps it is only a few hundred days before we devalue our legacy systems beyond recovery and place all our chips (literally) on half the future.

We can act now

The good news is that nothing prevents improvised integration today. Cloud platforms like Azure, Google, and AWS already support combining deterministic applications, such as SaaS and classical machine learning, with generative models.

That means we can combine the proven top-line benefits of machine learning with generative AI process productivity – a twin improvement strategy for EBITDA.

Similarly, agentic systems are evolving beyond our ability to govern them, but can be constrained by structured systems that filter and test their output.

So, even though business leaders cannot solve the current strategic disconnects, we can advance our local worlds by taking some practical steps:

1. Form multi-technology architecture teams to work on integration and define the strengths (and weaknesses) of each system type.

2. Ask the big business questions again from this wider perspective. Could a simple automated workflow be more reliable than an AI agent? Is a structured datastore more accurate for calculations than an LLM?

3. Rebalance investment portfolios to reflect this holistic technical approach. Over-indexing on generative projects may temporarily impress stakeholders, but it will ultimately come at a cost.

Conclusion

Iceman and Maverick were always strongest as partners, and enterprise SaaS and generative AI are no different.

Results come from a diverse architecture, not just one model – even if it does claim to be “the best of the best”.

Postscript

After this edition of The Hornet was finalised, Anthropic announced a partnership with Salesforce, with the Wall Street Journal now hailing AI as the saviour of enterprise software.

What a difference a week makes.

If you'd like to listen to an audio version of this article, please click here.

Footnotes

1 I have discovered that Gary Marcus was the first to proposehybrid AI approaches, but I came to the idea independently in my philosophical evaluation of ChatGPT.

2 More use should also be made of classical machine learning, which is mature and underutilised.