The second invoice: why AI technical debt cleanup costs will dwarf the savings
Generating code with AI costs almost nothing. Verifying it still carries a full human price, and the distance between those two numbers is where the next decade of IT budgets is going.
Every invoice for AI-generated code arrives twice. The first is the subscription, and it's trivially small. The second is AI technical debt cleanup costs, and it's the one that sinks the business case. Most organisations haven't priced it.
Follow the mechanism, not the marketing. Writing code was never the expensive part of software. Comprehension is. A line of code gets written once and read many times over: by the engineer patching it at 2am, by the auditor asking why it does what it does, by the new hire working out whether changing it will break something else. When generation becomes instant, that bottleneck doesn't vanish. It moves, from the keyboard to the review queue, from creation to verification.
Verification scales with volume, not with speed. Double a team's output with a code assistant and you've doubled the surface area somebody must read, test and stand behind. The productivity gain and the liability arrive in the same transaction. That asymmetry is the real economics of the AI trade: generation got cheap, judgement didn't.
The data has started to land, and it keeps landing on the same side. Google's DORA research programme, the most rigorous longitudinal study of software delivery, found that a 25% increase in AI adoption was associated with a 7.2% drop in delivery stability. More code, shipped faster, landed worse. Then there's the stranger result. In a 2025 randomised trial run by the research organisation METR, experienced open-source developers working with AI assistance finished tasks 19% slower, while believing they had been roughly 20% faster. A gain everyone feels paired with a loss nobody notices is exactly what a verification tax looks like from the inside.
The repositories themselves are confessing. Analysis by the code-quality firm GitClear across hundreds of millions of committed lines shows copy-pasted blocks and code churn climbing sharply since AI assistants went mainstream: more code being written, then rewritten, because nobody owned it the first time. Churn is the sound of the second invoice being printed.
What do AI technical debt cleanup costs actually look like?
Nobody can honestly quote you a tidy industry average, and you should be suspicious of anyone who does. But the shape of the bill is already visible in the work. There's code that passes review because it's plausible, then fails in production because plausible is not a correctness property. There's the system that works but can't safely be changed, because nobody on the payroll can fully explain why it's built the way it is. There's the security audit that now has to cover thousands of lines nobody remembers specifying.
Ask the people doing the work. In Stack Overflow's 2024 developer survey, fewer than half of respondents said they trust the accuracy of AI tools, and favourable sentiment slipped even as adoption climbed. The industry is using the tools and doubting them simultaneously. That gap gets paid for in review time, because distrust with a deadline attached is just a slower compiler. Each of these is a quiet tax on every future feature, and it compounds.
Where a recurring, expensive problem appears, a market forms around it, and this one is assembling in plain sight in three layers. First, measurement: GitClear's entire business is quantifying the damage, which tells you the damage is billable. Second, automation: firms like Moderne sell large-scale automated code remediation and migration, a product category that barely existed before the AI boom made it necessary. Third, and priciest, human rescue: the consultancies that were selling AI adoption eighteen months ago are repositioning to sell AI recovery.
The demand signal underneath all three is grim. RAND's analysis of AI project failures puts the failure rate above 80%, roughly double that of ordinary IT projects, and every one of those failures is a future remediation engagement. Follow the incentives and the sequence writes itself: the partner who billed you for the migration will happily bill you for the rescue. The rescue costs more for a reason buyers rarely see coming. The fixer has to warrant code they didn't write and can't fully explain, so remediation gets priced like demolition prices asbestos: risk premium first, labour second. An insurance and audit tail will follow close behind, because someone will eventually have to underwrite the question of who is responsible for a line nobody wrote.
Isn't this just the normal hype-cycle hangover?
It's a fair challenge, and it deserves a test rather than a wave of the hand. Every serious technology passes through a trough where the early bills come due: tooling improves, practice matures, the economics settle. If that's all this is, the cleanup trade is a footnote and the sceptics are shaking their fists at a cloud.
The counterargument is structural, and it comes from the history of code nobody reads. Your compiler generates machine code no human ever inspects, and nobody calls that technical debt. The difference is proof. A compiler is deterministic and its correctness is engineered and verified, so trusting its output costs nothing. AI-generated output carries no such guarantee. It's statistically likely rather than proven, which makes human review a permanent feature of the workflow rather than a teething problem the tool will grow out of. Better models will cut the failure rate. They won't remove the need for someone accountable to say: yes, this is right.
Ford has already paid this tuition. The company brought back more than 300 veteran engineers after AI and automated systems fell short on quality, according to TechCrunch, with warranty and recall costs forcing the issue. The people let go as surplus turned out to be the ones holding the quality line. Expect the rehiring wave to be far quieter than the layoffs were loud, and expect the premium on engineers who can read a system, rather than merely prompt one, to keep climbing.
Then there's the price of the tool itself. Today's API pricing is, by any sober reading of the economics, subsidised: the major providers are spending investor money to buy market share, and nobody serious believes these prices are permanent. The only genuine argument is about the date. Expect the flat subscriptions to keep acquiring caps, meters and throttles, because land-grab pricing always ends in metering.
If your business case balances at subsidised prices, you don't have a business case. You have a bet on someone else's runway, and the re-rating will arrive at exactly the moment your remediation bill does.
Perception is moving the same way. Two years ago 'AI-powered' was a selling point; increasingly it reads as an admission that the checking was skipped and nobody looked too hard. When a label flips from innovation to negligence, the premium shifts to whoever can prove the opposite. Expect 'human-reviewed' to become the new 'hand-stitched': a quality claim, priced accordingly.
What should you actually do about it?
The sensible response is arithmetic. Treat generated code the way you'd treat work from a talented stranger: useful, fast, and untrusted until someone who owns the consequences has reviewed it. That means keeping humans in genuine control of AI systems, not as a rubber stamp but as the layer that takes the weight. It means the unglamorous work of readiness before you build anything: data quality, clear ownership, and a way back when something goes wrong. It also means contracting for the mess in advance. If a vendor is generating your systems, negotiate audit rights and exit terms now, while they still need your signature. And know your current exposure. If you can't say which parts of your estate nobody can explain, a hard look at your technical strategy costs less than finding out in production.
The long fuse is the talent pipeline. Senior engineers are made by years of doing junior work badly, then less badly. Automate the junior work away and the supply of seniors a decade from now shrinks, precisely when the remediation market needs them most. Ford's returning veterans are a preview of the scarcest resource in software's next ten years.
The demo was free. The maintenance is the product. Budget for the second invoice, or pay someone to survive it.
Questions people ask
What is AI technical debt?
It's the future cost that accumulates when code and systems are generated faster than anyone understands them. Ordinary technical debt at least has an author you can ask; AI technical debt doesn't, because the author was a model and the reasoning behind its choices exists nowhere. Every later change, audit or incident pays interest on that gap.
How much does it cost to fix AI-generated code?
No honest industry average exists yet, so any vendor quoting a precise figure is selling something. What follows from the mechanics: remediation costs more than the original generation, because untangling a system requires understanding it first, and understanding is the part AI made neither cheaper nor faster. The realistic budget line is review, testing and audit time that scales with how much code was generated, not how quickly.
Are companies really rehiring engineers they replaced with AI?
At least one major manufacturer has, publicly: Ford brought back more than 300 veteran engineers after AI and automated systems failed to deliver the quality the business needed, as TechCrunch reported. Whether that becomes a wave is informed expectation, not settled fact. The first signal will show in the market for experienced engineers who can maintain systems rather than merely generate them.
Related
- The Memory Oligopoly Behind AI's Cost Inflation
- The AI Subsidy Cliff: Your Vendor's Investors Have Been Paying Your Bill
- AI Infrastructure's Circular Financing Risk: a Hall of Mirrors With a Mortgage
- AI & Automation
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.