Trending Topics
Trending
Your AI strategy is just another form of technical debt
This article is part of our Opinions section, where we invite industry professionals to share their views on the most pressing technology questions of our time.
Modern software development has become riddled with indeterminable processes and long development chains. AI should be able to fix this problem, but it’s not actually doing so. Instead, chances are your current AI strategy is saddling your organisation with even more technical debt.
The problem is fairly straightforward. As software development matures, longer and longer chains are being created from when a piece of software is envisioned until it’s delivered. Some of this is due to poor management practices, and some of it is unavoidable as programs become more complex.
How AI doesn’t fix the problem
Enter AI, which should, according to its marketing materials, fix these issues by, well, exactly how is unclear. The issue with most AI tools is that, beyond what they promise, most of them are frighteningly single in their purpose. It’s even questionable how well they carry it out.
Where a human can operate across multiple domains with varying depth, most enterprise AI deployments today are optimised for a single, narrowly defined task, and even then require close supervision. We’ve all seen the problems when you’re trying to get AI to do what you want; you need to be very precise in what you ask of it, and it needs to be supervised at all times.
Still, though, this hasn’t prevented many managers from removing people from certain links in the development chain and attempting a one-to-one replacement with AI. However, since most AI tools are designed around a single function, organisations end up deploying different tools at different points in the chain.
These tools can’t talk to each other, though; after all, they have just one purpose, and talking isn’t one of them. The results of all this, from the perspective of maintaining a coherent value chain, are pretty grim. Results are no longer predictable. Worse yet, they are not testable or reproducible. It’s just a set of random work. Coherence is missing, and lots of ends are left dangling.
When this happens across the chain – replacing people with AIs that aren’t interconnected – and knowledge becomes fragmented. While the chain still exists on paper – with Claude the coder replaced by Claude the clunker – in reality, you’ve only made an already unstable system even more fragile. Problems are bound to arise.
Technical debt
If this wasn’t bad enough, using all these different, single-purpose tools adds another problem, namely that you’re fragmenting all your data. Because these tools don’t talk to each other, you’re putting all the things your organisation knows into near-impenetrable silos. This further weakens your value chain as your workers, human and especially AI, need that data to function.
For example, if you were to create a product with 10 different phases in the chain, you’d need 10 tools, which means 10 different interfaces, but also 10 different databases. With the data scattered around like this, you have no idea what’s going on in any phase. Maybe one phase has a security issue, while another struggles with compliance. The overview is missing, and the machines in question can’t communicate any of it to each other.
The result is a massive amount of technical debt. While you’re deploying systems quickly, they will soon start slowly grinding down into a quagmire as less and less gets done. After all, your main focus has become making sure every system is talking to each other instead of pushing out code.
Data issues
The worst part is that your problems don’t end there: Data is the lifeblood of AI and its main source of sustenance, so to speak, but because of this approach, there just isn’t enough of it to feed most AIs properly. Many of the AI and SaaS services you’re using give you only curated, processed data, while keeping the raw data for themselves.
This speeds things up in the short term, but robs you of the best “food” for whatever AI you’re trying to build or operate long term. You’ll find it difficult to train anything yourself, meaning you’ll remain reliant on others. These, in turn, will provide you with worse AI products as the AI slowly eats itself; after all, while they have more data than you, it’s still a very limited amount.
At the same time, the data you do have is likely not doing much. You don’t have an awful lot of it, and your third-party programs are handling all that for you. Chances are, since there’s so little, you have no pipeline or warehouse. Since your data still lives in Excel, your AI “strategy” is really just a spreadsheet with hallucinations.
Now, you have a system reliant on AI – one you can’t easily fix because of all the technical debt you’re facing – and the components of the system aren’t working very well, and degrading, too. You’re in real trouble, so what’s the fix?
This is the core difference between bolted-on AI and platformed AI: one adds tools to an already fragmented system, the other reshapes how work, data, and responsibility flow through the organisation.
Starting from the ground up
Let’s discredit the Luddite solution first: there’s no way around AI. If it’s used right, it can do amazing things and improve productivity in every phase and sector. However, it needs to be platformed correctly if it’s going to be of any use. You can’t just throw AI at problems in the hopes it fixes them through sheer luck.
Bolting AI onto existing systems won’t work. AIs aren’t human, and you can’t replace them one for one, or even five for one. It doesn’t work. Instead, I’d advocate a system where your human staff are using AI as a way to speed up their own work, rather than replacing them outright.
You’d need to platform this system correctly, of course. Users would all have to have the same access, more or less, and both humans and bots would need to be able to communicate directly with each other as well as all dip from the same data lake – not to mention contribute to it.
That platform would not sit on top of your data pipeline as another abstraction, but run through it as an integral part of how work actually gets done.
You’d not only own the data, but you’d also decide where it flows and when, meaning you can fully harness the power of AI. The only way forward out of the AI quagmire is to take ownership of your data and let both people and machines work with it in a shared system. Anything else is just patching a problem that won’t go away.
More from our Opinion section
About The Author
