AI could prove CIOs’ worst tech debt yet

Some AI advocates promote the technology as a solution for rooting out expensive technical debt, but IT leaders are finding that too many pilot projects or too much reliance on AI-generated code can create their own problems.

AI tools can be used to clean up old code and trim down bloated software, thus reducing one major form of tech debt. In September, for example, Microsoft announced a new suite of autonomous AI agents designed to automatically modernize legacy Java and .NET applications.

At the same time, IT leaders see the potential for AI to add to their tech debt, with too many AI projects relying on models or agents that can be expensive to deploy and maintain and AI coding assistants generating more lines of software than may be necessary.

Without AI roadmaps, organizations risk willy-nilly deployments that can lead to long-term quality and maintenance issues, says Simon Wallace, CTO and cofounder at online civic action platform Suffrago.

“Without a clear AI implementation strategy, or a basic idea of why you are using AI, you’re going to end up with a whole bunch of tools gathering digital dust,” he says. “We saw this with microservices and cloud computing — tools were used for pitches and never saw the light of day again.”

Wallace recalls building tools for machine-learning pilot projects, then seeing them deployed to the cloud and never used. “Now you have an outdated tool that may be three or four model versions behind and using legacy connectors,” he says.

Moreover, whether it’s experimentation or production, many organizations are emphasizing expediency for their AI projects — a classic recipe for accumulating maintenance and cleanup issues that will one day come due, even for successful implementations.

Pilot paralysis

Endless AI pilot projects create their own form of tech debt as well, says Ryan Achterberg, CTO at IT consulting firm Resultant. This “pilot paralysis,” in which organizations launch dozens of proofs of concepts that never scale, can drain IT resources, he says.

“Every experiment carries an ongoing cost,” Achterberg says. “Even if a model is never scaled, it leaves behind artifacts that require upkeep and security oversight.”

Part of the problem is that AI data foundations are still shaky, even as AI ambition remains high, he adds. “That means most pilots already start on brittle ground, making them more likely to stall or fail,” he says. “Yet once they exist, CIOs are on the hook to manage the risk.”

Achterberg advises organizations experimenting with AI to conduct rigorous code reviews, use continuous deployment pipeline, and document relentlessly.

“Dozens of small, uncoordinated AI pilots can inflate tech debt just as much as decades-old legacy systems,” he says. “Without governance and ruthless prioritization, organizations end up paying the price to maintain expensive experiments that didn’t deliver business value.”

At many organizations, AI development leads to tech debt when many projects sprout up in different business units, adds Kurt Muehmel, head of AI strategy at AI development platform Dataiku. At some companies, workers empowered to spin up AI projects don’t report directly to a CIO or other senior IT leader, he notes.

“We can imagine a scenario where you probably have some of your data scientists who start setting up MCP servers,” he says. “And then they start using those tools to rapidly build out agents, but you don’t necessarily have monitoring built into those types of systems natively.”

Too much code

In addition to tech debt from too many AI pilot projects, coding assistants can create their own problems without proper oversight, adds Jaideep Vijay Dhok, COO for technology at digital engineering provider Persistent Systems.

In some cases, AI coding assistants will generate more lines of software than a developer asked for, he says.

“Unless I as a developer am diligent about what I’m accepting, it is going to generate superfluous code,” he says. “There is an inherent tendency of the models to throw additional things at you.”

In addition, if all members of a software development team, including QA testers, product managers, and release engineers, aren’t using the same AI coding tools, outputs can become misaligned, Dhok says.

“When you have these digital agents working with each other in a lot more collaborative manner, in terms of core test cases and requirements, they all get elevated together,” he adds. “When you elevate the generative AI usage from the individual to a team, collectively together, the risk of application technology debt goes down.”

AI project oversight

Muehmel recommends that organizations use governance tools that not only track data security and compliance issues but also monitor AI project spending and metrics. The CIO or another senior IT leader should have control over that oversight function, he suggests.

Governance tools are vital to ensuring that AI projects are meeting realistic metrics, he says.

“This comes up in every discussion that I have with a CIO or somebody who’s thinking at a strategic level about agents, is governance: ‘How do I make sure that I have my hands around everything that’s being built?’” Muehmel says. “There’s a real fear of shadow agents that have been built up by some smart people in the organization who are doing things that we don’t have a full grasp on.”

IT and business leaders overseeing AI projects also need to ensure that agents, in particular, need to be tightly coupled to business processes, he adds. Rogue agents that don’t map to business processes can create distrust among users, he says.

“What’s unique about agents is that, in a way, they’re supposed to be almost new colleagues or collaborators for the businesspeople in the marketing department, in the finance department, in operations,” Muehmel says. “If those people aren’t able to have confidence in the way that the agent is performing, then there’s a real trust deficit where you could end up building a bunch of things that people are hesitant to use.”

IT leaders should constantly evaluate their organizations’ AI projects and not be afraid to fail fast, whether at the early development stages or the deployment stage, he adds.

“In that fail-fast approach, there needs to be a really deep understanding for everyone who’s involved with building and deploying agents that the process cannot end at building it once,” Muehmel says. “External situations are going to change, because the business objectives might shift slightly, and so this process of ongoing adaptation of the way that agent is functioning needs to be anticipated as well.”