AI coding tools were supposed to make software delivery faster. They did. They also created a new kind of mess that most teams are only now starting to notice.
AI code is not the issue. Volume is. It ships fast, in bulk, and a lot of it reaches production with nobody able to explain what it does or why it holds together. Gartner gave this a name: AI debt. Their warning lands hard. The kind of mess that used to take fifteen years to build now shows up in fifteen months.
Why AI Debt Compounds Faster Than Anything Before It
Old debt piled up slowly. A shortcut taken on a Friday. A test skipped because the release was already late. Teams had years to notice the mess and chip away at it.
That timeline is gone. IBM asked 1,300 senior AI decision-makers about what happens when debt gets ignored. Returns dropped between 18% and 29% for that group, and schedules stretched by up to 22%. Forrester’s numbers point in the same direction. Three out of four technology leaders expect debt to hit severe levels this year.
It’s easy to pinpoint the cause. Artificial intelligence creates code faster than anyone can actually read the code. Sonar conducted a survey among developers and concluded that 53% said that while the code looks correct, it performs poorly when challenged. An additional 40% believed that AI creates debt by creating code that is already somewhere in the repository.
1. Avenga
Avenga provides digital transformation consulting services and treats AI debt as something you design against, not something you clean up later. Most firms wait until the mess shows up and then sell you a cleanup. Avenga builds the guardrails into the delivery work itself.
Here is why that matters. AI helpers write code that passes the tests and then leave nobody able to say why it works. GitLab asked organizations about this in 2026. 43% admitted they cannot reliably tell AI-written code from human-written code. 82% said the AI code in their systems is creating a kind of debt they are not ready to handle.
So Avenga handles it where it starts. Quality gates run on AI output before it merges. Acceptance rates for agent-written code get tracked. Delivery gets measured across milestones, code quality, and whether the system can scale, not just how fast it ships. That last one catches the debt that hides behind quick releases.
Where debt prevention actually lives:
- AI-generated outputs get evaluated against defined quality gates before they merge, not after they cause problems
- Delivery gets measured across scalability, not just speed, so debt surfaces while it is still small
- Agent-produced code is tracked by acceptance rate, which exposes code that looks correct but is not understood
2. Luxoft
Luxoft provides digital transformation consulting and published a warning about the tools it sells.
Many organizations that provide delivery services based on AI emphasize speed over all other factors. Luxoft suggests a different approach to the problem. At the beginning of 2024, Luxoft specialists noted that AI copilots encourage risky coding practices. The blog post contained the following direct question: when the code needs to be rewritten or rolled back, how was speed an advantage?
The question constitutes the basis for further discussion. With the increase in system complexity, code churn is increasing, code reuse is decreasing, and maintenance is becoming more costly. Luxoft links these tendencies to digital transformation consulting efforts that include integration of AI tools into the delivery process without the advantages of the software development lifecycle.
What makes the position unique is that Luxoft is not opposed to AI. The same team performed an experiment that included feeding COBOL into ChatGPT in order to find out what value language models have. The result was straightforward. Large-scale code conversion became too expensive to beat the current tools. The areas in which AI showed its benefit were vague bug report interpretation, generation of test cases, and writing unit tests. In addition, they developed a tool that generates inline code documentation, which they consider to be valuable since it does not change the code itself.
For those who manage AI-influenced code, this combination of research results is crucial. Luxoft knows what parts of the process AI helps and what parts it hinders.
Where the debt conversation actually happens:
- Luxoft’s own research links AI copilots to rising code churn and falling reuse, which drives up maintenance costs over time
- Their COBOL experiment found AI conversion too expensive at scale, but useful for testing, error parsing, and documentation
- The firm treats documentation generation as a net win because it leaves the codebase untouched
3. Intellias
Intellias provides consultation services related to digital transformation and sees the AI-generated content through the lens of validity with respect to the existing system and not trust by default.
While many companies using AI during modernization rely on speed to demonstrate success, Intellias subjected its process to a more profound analysis on a ProAg insurance core system project. Intentionally, one of the most complex processes in a regulated area, as opposed to some easy-to-manage process, was chosen for modernization. At stake was the ability of agentic AI to manage production-ready code, and not generate something that would simply compile.
The validation technique became a part of deliverables in this project. The migrated build was run in parallel with the old process, with the comparison of outputs taking place in each step, and not only at specific boundaries as it would happen otherwise. What used to take several months in a small team for similar modernization took only five weeks in this case.
It is significant for the problem of AI debt since, in reality, the matter is not in some unexecuted code but in some executed code that is not trusted. Intellias incorporates this process of validation in its workflow. AI works within the client’s AWS in a manner provided by Amazon Bedrock, thus keeping the regulated code within the compliance boundary.
Those who deal with some AI-affected code, which can’t be validated, follow this practice of comparing the result of migration to the original one.
Where the debt conversation actually happens:
- Validation gets scoped as a deliverable, not treated as an afterthought
- The migrated build runs in parallel with the legacy system, compared at every stage
- AI inference stays inside the client’s own cloud environment, so regulated code never leaves
Where Each Firm Starts the Cleanup
The three firms approach AI debt from different entry points. One prevents it during delivery. One takes over systems that already collapsed. One rebuilds the architecture so the debt cannot compound. The table below maps where each one begins and what the work looks like in practice.
| Firm | Entry Point | Primary Fix | Evidence |
| Avenga | During delivery, before debt accumulates | Embeds AI into the workflow with debt prevention as a design constraint | Measures engagement across milestones, quality, and scalability; up to 50% faster delivery |
| Luxoft | After other modernization attempts have failed | Rebuilds missing knowledge and builds custom tooling when nothing exists | 100% success rate on mainframe modernization; 77% industry failure rate cited |
| Intellias | On platforms where AI created tangled dependencies | Decouples components and migrates to managed cloud services | Planstin case: cleaner architecture, faster delivery after AWS migration; Travis Perkins: 30% lead time reduction |
Each firm starts where the problem is most visible. Avenga prevents it from forming. Luxoft fixes what already broke. Intellias untangles what AI made messy.
Why AI Debt Is Different From What Came Before
Old debt remained centralized. A problem yet to be addressed. A test that was skipped since the release was already delayed. It could be detected, identified, and fixed.
However, the new debt is moving. As per VentureBeat, it is hidden within prompts, dependency models, retrieval pipelines, and generated code that has not been reviewed adequately. The last part is what makes this kind of debt so hard to find. The code passed. The tests have passed too. Next comes the model update at another place, and the problem surfaces, which cannot be replicated.
Here lies the problem in a nutshell. Conventional bugs can always be found in a particular piece of code. The bug that results from using AI technology can result from prompts changed months ago, model changes without documentation, and data sources changed in the retrieval store. One cannot discover such problems by analyzing the code. They become visible only when the system is operated.
The firms mentioned earlier are well aware of it. Their approach to fixing this problem starts from the identification of existing components rather than recommending alternatives. This is the only way to locate the debt.
Conclusions
AI was a catalyst for advancement as well as chaos; the latter, especially, has become harder to decode than the code it has replaced.
Avenga provides digital transformation consulting geared towards solving the former issue. The idea is to avoid technical debt accumulation so as not to pay for it in the future. Performance delivery will be judged not only by how fast the product is launched but also by how well it scales. AI is embedded into the process and not added to it as an afterthought. The result is code meant to last beyond the next rebuild.
If your codebase already exhibits signs of having been written with the help of AI, which you can no longer track, the three companies mentioned above have a solution for you. They all start with the same thing: identify what you already have first. Everything else comes later.
