AI Summary
AI-generated code quality has not improved: models introduce a known security flaw in 45% of tasks, a rate that has stayed flat across two years of model releases (Veracode, March 2026). The cost lands downstream rather than at authorship: telemetry from 22,000 developers shows median pull request review time up 441.5%, incidents per merged PR up 242.7%, and 31.3% more code reaching production with no review at all (Faros AI, April 2026). Before scaling offshore, engineering leaders should confirm whether they are buying production capacity they no longer need or the validation capacity that has become the real constraint.
That gap is measurable, and it is not a perception problem.
AI coding assistants removed the constraint that shaped engineering org design for two decades. Writing code used to be the slow part. It isn't anymore.
What replaced it is the work of confirming that generated code is correct, secure, licensed, and safe to deploy. Most teams have added none of that capacity.
You are probably reading this because the pressure has become visible. Review queues are backing up. Your senior engineers spend their days reading other people's output.
Someone has proposed scaling offshore to relieve the load.
That can work. It can also make things measurably worse, and the difference comes down to what kind of capacity you buy.
This guide covers what the evidence says about AI-generated code quality, four control layers to require from any partner, and ten questions to send every vendor you evaluate. On rates and engagement models, our 2026 Philippines pricing guide covers the cost math in detail. This article covers the part that comes first.
In This Guide
Why AI made your team faster and your delivery slower
The validation bottleneck is the point where AI-accelerated code production outpaces an organization's capacity to review, test, and verify it. Code enters the pipeline faster than humans can confirm it works. The constraint moves downstream to review, QA, and incident response.
Google's DORA team surveyed nearly 5,000 technology professionals for its 2025 State of AI-assisted Software Development report. It found that 90% now use AI at work and over 80% believe it has raised their productivity. It also found something less comfortable: higher AI adoption is associated with an increase in both software delivery throughput and software delivery instability.
DORA's qualitative analysis explains the mechanism plainly. Time saved during code creation gets reallocated to auditing and verification.
The work didn't disappear. It moved.
Faros AI put numbers to where it moved. Its 2026 AI Engineering Report analyzed two years of telemetry from 22,000 developers across more than 4,000 teams, comparing each organization's periods of lowest and highest AI adoption.
Faros AI Telemetry Impact Comparison
| What improved | Change | What it cost | Change |
|---|---|---|---|
| Epics completed per developer | +66% | Median time in PR review | +441.5% |
| Task throughput per developer | +33.7% | Incidents per merged PR | +242.7% |
| PR merge rate per developer | +16.2% | Bugs per developer | +54% |
| — | — | Code churn (deleted vs added) | +861% |
| — | — | PRs merged with no review | +31.3% |
Both columns are real. Roadmaps genuinely moved. So did the repair bill.
The finding worth pausing on is the disagreement between these two studies. DORA concludes, from survey data, that strong engineering foundations amplify AI's benefits and buffer its downsides. Faros, from telemetry, reports that organizations with mature DevOps practices and high DORA scores experienced the same downstream deterioration as everyone else.
Both can be right. Surveys capture how developers feel, and at the individual level they are more productive. Telemetry captures what happened to review queues, incidents, and rework.
Perception lags reality.
For anyone deciding whether to expand their engineering org, the practical read is this: do not assume existing process maturity will absorb a step change in output volume, or that AI-generated code quality will hold as adoption deepens. The largest available dataset says it didn't for anyone else.
AI-generated code quality: how secure is the output in 2026?
AI-generated code fails security testing in roughly 45% of cases, and AI-generated code quality has not improved in two years. Veracode runs 80 coding tasks across four languages against flagship model releases, and has now evaluated more than 150 models.
Its Spring 2026 update reports syntax correctness above 95% and security pass rates around 55%, essentially flat since 2023.
Models became excellent at writing code that compiles. They did not become better at writing code that is safe.
The failure pattern is specific rather than random, which matters when you design controls around it.
Veracode Security Pass Rates by Language & Flaw Type
| Dimension | Security pass rate | Read |
|---|---|---|
| Python | 62% | Best performer, modest improvement |
| C# | 58% | Small gains over time |
| JavaScript | 57% | Essentially flat |
| Java | 29% | Worst performer by a wide margin |
| SQL injection (CWE-89) | 82% | Models handle this well |
| Insecure cryptography (CWE-327) | 86% | Models handle this well |
| Cross-site scripting (CWE-80) | 15% | Persistent failure |
| Log injection (CWE-117) | 13% | Persistent failure, getting worse |
The split has a clean explanation. Models excel at local pattern recognition, so they reliably avoid a string-concatenated SQL query they have seen flagged thousands of times. They fail at tracking how user input moves across functions and files to reach a rendering context, which is what catching cross-site scripting requires.
Veracode's own research team notes that this kind of dataflow analysis is difficult even for humans. That is exactly why static analysis tools exist, and exactly why treating a model's output as pre-validated is a category error.
Two caveats worth carrying into vendor conversations. OpenAI's reasoning models reached 70% to 72%, the strongest category measured, which still leaves close to three in ten snippets carrying a flaw. And Java's 29% should give pause to anyone running enterprise Java, since that is where models default hardest to insecure legacy patterns.
Measure your own bottleneck before you add capacity
Before evaluating a single vendor, pull five numbers from your version control and incident tooling. Compare each against your own baseline from twelve to eighteen months ago, before AI tools reached broad use on your team. Direction matters more than absolute value, and your baseline is more reliable than any published benchmark.
Internal Pipeline Bottleneck Checkpoints
| Metric | What it tells you | Warning sign |
|---|---|---|
| Median time in code review | Whether reviewers can absorb current output | Rising while merge volume also rises |
| Share of PRs merged with no review | Whether oversight is quietly being skipped | Any upward drift at all |
| Code churn, lines deleted against lines added | How much shipped code gets rewritten shortly after | Sharp increase, signalling rework rather than refactoring |
| Incidents per merged PR | Whether quality per unit of change is holding | Ratio climbing even as absolute incidents look stable |
| Share of reviews handled by top 3 engineers | Concentration of the senior engineer tax | Growing concentration, which is a retention risk before it is a delivery risk |
Two of these deserve emphasis. Code churn is the one most teams never instrument, and it is the clearest signal that throughput gains are partly illusory. Reviewer concentration is the one that predicts resignations.
If your review time and churn are climbing, adding code producers makes both worse. That is the finding to take into a hiring conversation, and it is worth ten minutes with your data before anyone signs anything.
Will offshore capacity fix your bottleneck or amplify it?
It depends entirely on whether you buy production capacity or validation capacity. If your review queue looks anything like the ones Faros measured, adding developers who generate still more AI-assisted output tightens the constraint rather than loosening it. Adding senior reviewers, test automation engineers, and security specialists relieves it.
This is the single decision that determines whether offshore expansion works in 2026, and it inverts how most teams have historically approached headcount.
The traditional offshore model was headcount-heavy at the junior level. You bought volume, because volume was the constraint. That logic has quietly expired.
Faros found that the acceptance rate of AI-generated code has risen from 20% to 60%, and that 80% of studied teams now pass the mark where more than half of developers use AI tools weekly. Production capacity is abundant. Verification capacity is scarce.
Two costs deserve naming before you build a business case:
- The senior engineer tax. AI-generated code is often superficially convincing: idiomatic, well-named, consistent with surrounding style, with structural failures sitting beneath the surface. Catching them requires reading carefully and reconstructing intent rather than scanning for obvious errors. Faros measured the result: median time to first review up 156.6%, average time spent in review up 199.6%. Your most expensive people absorb this, and adding junior offshore capacity increases their load.
- The coordination tax. When AI resolves a task in ten minutes, waiting twelve hours for an architectural clarification erases the gain entirely. Time zone overlap mattered less when tasks took three days, and it matters far more now that the unit of work has collapsed. This is an argument for real overlap hours and direct developer access, not for account managers relaying questions.
There is also a broader signal on AI cost economics worth registering. Gartner forecast in January 2026 that the cost per resolution for generative AI in customer service will exceed $3 by 2030, higher than many offshore human agents, citing rising data center costs and vendors shifting from subsidized pricing to profitability.
That finding covers customer service rather than software development, so it does not transfer directly. The directional point does: AI is not automatically the cheaper input, and plans that assume permanently subsidized inference are fragile.
The model that works looks like fewer, more senior engineers with automation underneath them. A smaller team of reviewers, test automation specialists, and DevSecOps engineers outperforms a larger pool of code producers on both quality and total cost.
Weighing how to structure your offshore team?
Our guide to AI-augmented offshore engagement models breaks down staff augmentation, EOR, and BOT against compliance exposure and IP ownership.
The AI Delivery Assurance Stack: four layers to require from any partner
The AI Delivery Assurance Stack is a four-layer control model for governing AI-generated code quality across an engagement, covering provenance, validation, licensing, and containment. Each layer answers a question you should be able to put to a vendor before signing, and each has a verifiable artifact behind it.
Provenance: Can they tell you what AI wrote?
You cannot govern what you cannot identify. This layer covers tagging AI-generated code at commit, tracking tool usage and acceptance rates, and maintaining a software bill of materials (SBOM). A partner who cannot distinguish AI-generated from hand-written code in your repository cannot make any credible claim about the rest of the stack.
Validation: Is review capacity sized to output?
Static application security testing (SAST) gates on every push, not sampled. Mandatory human review with a stated reviewer-to-developer ratio. Automated test coverage targets you can verify in a report. The question to ask is not whether they review code. It is how their review capacity scales when AI-assisted output doubles.
Licensing: Does compliance scanning run pre-merge?
Software composition analysis (SCA) in the CI/CD pipeline, triggered at code push. A quarantine directory for output from third-party or unapproved tools. Documented handling for flagged code. The next section covers why this layer carries more legal weight than most buyers realize.
Containment: What does AI never touch?
A written boundary around production credentials, customer PII, and core business logic. Approved tools only, with shadow AI use treated as a policy violation. Content exclusions configured and documented in repository guidelines.
At Hireplicity, this is how we already operate. Every AI tool is formally approved before it touches client code, and no student PII passes through an AI system without written client approval.
Our engineers use GitHub Copilot, Cursor, and Claude as standard tooling while maintaining 80% to 85% automated test coverage, with compliance checkpoints at every phase gate. We have passed SOC 2 and FERPA audits cleanly for multiple clients, and we walk through our data handling documentation before an engagement starts.
Ask any vendor to show you the equivalent. The answer tells you a lot.
The licensing exposure you inherit when you outsource
License laundering is what happens when an AI assistant reproduces code derived from a restrictive open-source license while stripping the attribution and license metadata that would have flagged it. The output looks clean. The obligation travels with it anyway.
Researchers at the University of Alabama documented this risk in a 2025 framework called DevLicOps (Sharma et al., August 2025). Their paper opens with a scenario that should be uncomfortable reading for anyone about to sign an offshore contract: a manager reviewing deliverables learns the offshore team used AI assistants with no license compliance checks, and the contract contains no clauses covering AI-generated code.
That is a pre-signature problem, not a post-delivery one.
Most buyers assume vendor indemnity covers them. The DevLicOps authors read the major providers' clauses and reached a blunter conclusion.
Nearly all indemnity terms exclude coverage where output is modified, transformed, or combined with products from other providers. Fine-tuning a model can also void protection. Since developers routinely iterate on generated code, the authors advise that most firms should assume the protections do not apply at all.
The consequences are not hypothetical. In 2024 the Paris Court of Appeal ordered Orange S.A. to pay more than €900,000 in damages for GPL violations.
Software Freedom Conservancy's case against Vizio, which seeks release of source code for a smart TV platform, cleared summary judgment in February 2026 and is headed for trial (Baker Botts, May 2026). A ruling there could clarify how far copyleft obligations reach into adjacent proprietary modules.
Awareness inside engineering teams is thin. In a survey of 574 developers cited by the DevLicOps authors, only 11.9% reported any formal training in copyright law or the legal implications of code generation tools.
Four contract terms to negotiate before you sign:
- Disclosure. The vendor names every AI coding tool in use and notifies you of changes.
- Scanning evidence. Composition analysis runs pre-merge, with reports available on request.
- License warranty. The vendor warrants delivered code is free of copyleft obligations you have not accepted.
- Audit right. You can inspect compliance records and commission an independent scan.
None of these are unreasonable asks. A vendor who resists them is telling you something useful.
Ten questions to ask before you sign
A strong vendor answers these with documents and configuration screenshots. A weak one answers with reassurance. The distinction is the whole point of asking.
Context first. Zscaler's ThreatLabz 2026 AI Security Report analyzed 989.3 billion AI and machine learning transactions across roughly 9,000 organizations.
Enterprise data flowing to AI applications rose 93% year over year to 18,033 terabytes. ChatGPT alone was tied to 410 million data loss prevention policy violations, including attempts to share source code.
Most security questionnaires ask whether a vendor trains models on customer data. That is the right question and an incomplete one.
A contractual no-training term does nothing to stop a developer from pasting a production key into a prompt. Context-window leakage is the higher-frequency risk, and it needs its own controls.
AI Vendor Evaluation Questionnaire
| # | Question | What a weak answer sounds like |
|---|---|---|
| 1 | Which AI tools are approved, at what licence tier? | "Our developers use whatever works best." |
| 2 | Do your terms include zero data retention? | "We're confident the vendor handles that." |
| 3 | Are you contractually exempt from model training? | "We assume so." |
| 4 | Which sub-processors and foundation models sit in the request path? | "That's on the vendor's side." |
| 5 | How do you prevent secrets entering prompts? | "We tell developers not to." |
| 6 | Is SCA license scanning running pre-merge? | "We scan periodically." |
| 7 | Can you identify AI-generated code in a commit? | "Not specifically." |
| 8 | What is your reviewer-to-developer ratio? | "Everything gets reviewed." |
| 9 | What is your policy on unapproved tools? | "We trust our team." |
| 10 | Who is accountable when AI-generated code causes an incident? | Silence, or a discussion of limitation of liability. |
Send these to every vendor on your list. Send them to us too.
Frequently asked questions
Yes, if you hire for the current constraint. AI increased code production, which shifted the bottleneck to review, testing, and incident response. Offshore senior reviewers, QA automation engineers, and DevSecOps specialists relieve that pressure, while junior developers who produce more output usually deepen it.
Veracode's testing across more than 150 models found AI introduces a known security flaw in roughly 45% of coding tasks, with cross-site scripting and log injection failing most often. The rate has stayed flat for two years while syntax correctness climbed past 95%. Treat generated code as unreviewed input requiring the same scanning as any external contribution.
Copyleft obligations attach to whoever distributes the software, so absent explicit contract terms the exposure tends to follow you rather than your vendor. Vendor AI indemnity often will not close the gap, since most policies exclude modified or recombined output. Allocation varies by jurisdiction and contract, so treat this as a question for counsel, not a settled answer.
Four terms cover most exposure: disclosure of every AI tool in use, evidence that composition analysis runs before merge, a warranty that delivered code carries no unaccepted copyleft obligations, and an audit right letting you commission an independent scan. Each is standard and verifiable.
Contractual no-training terms are necessary but insufficient. Require enterprise tiers with zero data retention, enforce secret scanning before prompts leave the IDE, configure content exclusions for sensitive directories, and treat unapproved tool use as a policy violation. Most real exposure comes from developers pasting credentials, not from model training.
There is no published benchmark, so treat any specific ratio with suspicion. The workable approach is measuring your own pipeline: track median time in review and the share of pull requests merging without review. If either climbs after adding capacity, you bought the wrong kind.
The decision before the decision
Scaling offshore is a sound response to engineering pressure. It is the wrong response to the pressure most teams are actually feeling right now, unless you deliberately choose what you are buying.
AI-generated code quality is not a tooling problem you solve by picking a better model. Every major release since 2023 has left the security pass rate roughly where it started, even as the same models got dramatically better at everything else. It is a governance problem, and governance is something you build into an engagement from the first sprint or retrofit expensively after an incident.
Before you sign anything, get answers on provenance, validation, licensing, and containment. Put the ten questions to every vendor you are considering. The ones who answer with documentation rather than reassurance are the ones worth your shortlist.
Ready to scale with a partner who can show you the controls?
Hireplicity is a US-led engineering team based in Cebu, with 18 years building secure software for EdTech and SaaS companies. Book a 30-minute call and we will walk you through our AI governance documentation before you commit to anything.
Sources & References
- Veracode — Spring 2026 GenAI Code Security Update (March 24, 2026) — https://www.veracode.com/blog/spring-2026-genai-code-security/
- Faros AI — AI Engineering Report 2026: The Acceleration Whiplash (April 12, 2026) — https://www.faros.ai/blog/ai-acceleration-whiplash-takeaways
- DORA (Google Cloud) — 2025 State of AI-assisted Software Development & Balancing AI Tensions — https://dora.dev/insights/balancing-ai-tensions/
- Sharma, P. N. et al. (Univ. of Alabama) — DevLicOps: A Framework for Mitigating Licensing Risks in AI-Generated Code (arXiv:2508.16853, August 2025) — https://arxiv.org/abs/2508.16853
- Zscaler — ThreatLabz 2026 AI Security Report (January 27, 2026) — https://www.zscaler.com/press/zscaler-2026-ai-threat-report-83-year-over-year-surge-ai-activity-creates-growing-oversight
- Baker Botts — When Consumers Enforce Open Source: The SFC v. Vizio Case (May 2026) — https://www.bakerbotts.com/thought-leadership/publications/2026/may/when-consumers-enforce-open-source
- Gartner — Gartner Predicts GenAI Cost Per Resolution for Customer Service Will Exceed Offshore Human Agent Costs by 2030 (January 26, 2026) — https://www.gartner.com/en/newsroom/press-releases/2026-01-26-gartner-predicts-genai-cost-per-resolution-for-customer-service-will-exceed-offshore-human-agent-costs-by-2030
