Software Improvement Group publishes a benchmark most people in this industry only half-read: over 30,000 systems, 400 billion lines of code, updated yearly. This year's edition, State of Software 2026, measured what changes when AI-assisted coding goes fully mainstream and landed on a conclusion that's blunter than the usual vendor-report hedging: AI does not fix or break software discipline on its own. It amplifies whatever discipline was already there. Where quality is measured and managed, AI accelerates delivery. Where it isn't, AI accelerates debt, cost, and security exposure, at the same rate.
The numbers behind that conclusion are worth sitting with. 72% of AI systems currently in production score below SIG's own recommended build-quality rating. 71% of code across their sample has a low degree of security controls. And the relationship between the two isn't a coincidence: systems rated 4-star on maintainability carry a 72% higher security rating than 2-star systems. 3-star systems sit at a 36% uplift. Better structure and better security aren't separate initiatives, they're the same initiative measured two ways.
Using an average loaded developer cost of €150,000 a year, SIG puts the direct labor cost of poor maintainability at roughly €870,000 per system per year for a 2-star system, close to €9 million a year across a 10-system portfolio held at that level. That's not a hypothetical, it's what a real portfolio of unmanaged debt actually costs in salary alone, before counting the incidents, the failed AI rollouts, or the deals that stall in due diligence.
Why we built the cost calculator at function-level granularity
Most technical debt frameworks, SIG's stars, CAST's per-line-of-code estimate, SEI's SQALE ratio, price debt at the file or the whole-application level. That's useful for a board slide, less useful for an engineer trying to decide what to actually fix Tuesday morning. TrueMend's cost calculator works at function-level granularity instead: every anti-pattern instance, in every function, priced individually.
fix_cost = estimated_fix_minutes * (hourly_rate / 60)
* overhead_multiplier * complexity_multiplier
complexity_multiplier by tier:
deterministic = 0.0 (free, AST transform, no LLM)
cataloged_refactor = 1.0 (bounded, named procedure)
structural_rewrite = 2.0 (human decision, high variance)
The complexity multiplier is doing the same job as SIG's star system, translating structural quality into a number a non-engineer can act on, but it's tied to our own three-tier taxonomy instead of an abstract rating. A deterministic fix costs nothing because it's mechanical and verified. A cataloged refactor has a bounded, predictable cost because it's a named procedure with an estimated fix time. A structural rewrite carries a 2x multiplier specifically because that's an investment decision with real variance, not a mechanical task with a knowable price.
That split produces three different sentences for three different audiences in the same room: "we can eliminate this much debt for free right now," for the free tier; "this investment pays back in this many months," for the bounded tier; "these items need an actual architectural conversation," for the tier that shouldn't be priced like the other two. SIG's report makes the case for why that conversation needs to happen at all. Function-level pricing is what makes it possible to have without a six-week manual audit first.
The part that connects back to AI-generated code specifically
SIG's finding that AI amplifies existing discipline rather than creating new discipline is the whole reason a pricing model needs to exist before the AI-generated code lands, not after. A codebase with unmeasured debt that starts accepting AI-generated pull requests at volume isn't getting a fresh start, it's compounding whatever was already unmeasured, faster than before. The 72% of AI systems scoring below SIG's recommended build quality aren't failing because the AI is bad at writing code. They're failing because nobody priced the foundation before building faster on top of it.
Where TrueMend fits
A dollar figure a CFO can read, tied to a specific function, split by whether the fix is free, bounded, or a real investment decision. That's the answer to "how much is our technical debt actually costing us" that doesn't require a six-week consulting engagement to produce.
- For engineering leads building the business case for a cleanup sprint: function-level pricing tells you which function to fix first, not just which file is worst.
- For anyone about to scale AI-generated code across a codebase: price the foundation before the volume goes up. SIG's data says the amplification effect runs both directions.