- AI coding tools are accelerating productivity, but they’re also spawning duplicated code, phantom dependencies and hidden technical debt
- Experts warn telcos face especially high stakes if AI is layered onto outdated systems
- Human ownership, centralized observability and steady debt cleanup can help avoid disaster
AI is changing the world, but telcos and enterprises that blindly rely on AI tools to make IT engineering and architectural decisions for them risk being caught in a “wildfire” of technical debt that could quietly cripple forward progress, experts told Fierce.
As Fierce has noted before, technical debt refers to the long-term cost enterprises pay for short-term shortcuts, quick fixes and duplication in their IT infrastructure. Technical debt is something that existed – and plagued – enterprises long before AI, but the advent of coding and other AI-enabled tools is creating a new wave of problems.
“Because the barrier to entry with AI is so low right now, teams are rushing to deploy solutions. And on the engineering side, we’re seeing AI coding tools drive a massive spike in duplicated code and what we call phantom dependencies, where the AI make up connections between systems that don’t actually exist,” Adam Shea, director of AI go-to-market at TEKsystems, told Fierce.
Indeed, a study from the University of Texas last year found the average number of phantom dependencies (or “package hallucinations” as the paper calls them) cited by AI-generated code ranged from 5.2% for commercial AI models to 21.7% for open-source models. Notably, though, the paper also found that three of the models it tested were able to detect their own hallucinations when asked with an accuracy rate above 75%.
But phantom dependencies are only one problem. In addition to dependency issues, Databricks noted that overly complicated prompts to coding tools can create conflicting instructions and a lack of traceability can make it hard to fix errors that arise from the use of LLM-based tools in a blog about AI technical debt.
Shea said this pileup of problems makes it harder to maintain and secure existing infrastructure. As a result, enterprises can end up “crippled” by cleanup efforts, with their software teams spending more time trying to keep a metaphorical house of cards standing than they do innovating.
“You end up with a wildfire of custom tangled integrations instead of clean, centralized repositories,” Shea explained. “That’s a massive drag on organizations’ ROI.”
For telcos, the risk is especially stark. The blast radius of a failure is huge for telcos. And if they try to layer AI on top of outdated systems, “that wildfire might go out of control,” Shea warned.
Getting it right from the start
But there are few things telcos and other enterprises can do to avoid being consumed by AI technical debt.
The first is to begin enforcing human ownership of AI-generated code. That is, the engineers giving AI-written code the green light to go into production, need to deeply understand the underlying architecture of that code and be responsible if something breaks. Leadership also has a role to play here, Shea said. They need to stop pushing just for greater code volume and start rewarding system stability.
Organizations should also work to implement a firm control plane for observability and lifecycle management, with a centralized repository. This will help with not only governance and compliance but with compliance in regulated industries.
“Right now, companies are getting a sugar rush of productivity. But if you have dozens of agents running in production and leadership has no idea what they do, how they make decisions or what enterprise systems they touch, they’re building a house of cards,” Shea said.
Finally, Shea said organizations should take the time to invest a little bit of each sprint project into addressing outstanding technical debt. They should also work to define a fairly flexible and open architecture plan that can evolve as technology changes.