When I was in my CEO seat, seeing the R&D budget consumed by tech debt was absolutely the last thing I wanted – but skipping it would put the business at risk.
Now let’s be clear: no CIO or CTO wants to spend money on tech debt. They just have to.
Tech debt is an absolute monster that consumes up to 40% of the value of the IT estate, if you believe McKinsey. Even if you don’t – tech debt routinely accounts for around 20% of R&D spending as any practitioner will tell you. It involves:
- Bug fixing
- Performance improvements
- Vulnerability management
- Dependency deprecation
- Architecture modernisation and refactoring
- …
- Waaaaahhh!
It’s endless. And this is not just a problem of costs; it’s a significant barrier to innovation.
Every dollar spent on maintaining outdated systems is a dollar not invested in exploring new markets, launching new products, or enhancing customer experiences. Tech debt thus silently erodes competitive advantage and responsiveness to market shifts.
Right, Fintech hat on.
Financial Services is an industry that, yes, writes a lot of code on an ongoing basis and already benefits from the various co-pilot and “software bootstrapping” products out there.
But it’s also an industry that has billions of lines of legacy code. One that spends, in some cases, billions of dollars maintaining a legacy IT estate before being able to work on innovation; and has a big, overarching vendor and in-house complexity problem.
In a December 2024 paper published by Anthropic, about 37% of conversations were about computer programming – with many of them involving code generation. The next biggest category, in arts, design, media, and sports, was at 10%.
Code generation is the number one power app for LLMs today, with multiple startups in the space showing explosive revenue growth.
Let’s not be naive. We’re not all moving to [insert your favourite software generation tool]. We can’t pretend that we just live in a greenfield world out there.
So why ignore some of the biggest issues of our time?
At scale, tech debt represents billions in development spending that everybody hates but is “mandatory”.
At 13books, we think the impact of AI in this area will continue to be enormous, but we can’t all sit here and pretend that greenfield code generation is the only game in town.
Code generation, as a use case, has the big advantage that it can be verified for correctness and usually benefits from lots of training data.
That also means that more direct optimisation techniques like GRPO, or whatever comes next, can impact this area – as we saw with DeepSeek.
Automating vulnerability patching, rewriting code bases for performance, automated cloud migration, translating mainframe code from COBOL to more modern frameworks… these problems are HARD. They create huge problems with context size, correctness, and testing – and that means they are worth a big prize to the winner.
Maybe we don’t even have to go that far. Even just analysing the existing estate, providing visualisations and insights that were previously impossible, making recommendations, managing the decade-long transition to new architectures – moving beyond the static analysis and data lineage of the past – is still a major leap forward.
If you are working in this field, reach out for a chat!
