History Does Not Repeat, It Rhymes #
If you worked in software during the early 2000s, the current AI hype feels familiar. Back then, management discovered IT outsourcing to India. The pitch was simple. Management wanted five cheap offshore coders for the price of one onshore developer.
Today, the same pitch runs on tokens instead of hourly rates. Companies expect AI agents to replace large parts of engineering teams and cut costs fast. Senior engineers who lived through the 2000s see the same pattern again. Companies burn money, hit a quality shock, and return to experienced engineering.
Data from 2024 to 2026 shows this cycle is already running.
What Happened in the 2000s #
In the early 2000s, enterprises chased the lowest hourly rate. They handed vague specs to vendors and expected working software fast.
The real bill arrived quickly. A 2005 Deloitte Consulting study surveyed 25 major global organizations[1]. Exactly 70% reported significant negative experiences with outsourcing. Another 64% brought outsourced services back inside the company, and 44% missed expected cost savings.
Large companies reversed course publicly. Dell stopped routing US corporate support calls to Bangalore in 2003 after customer complaints[2]. JPMorgan Chase canceled its $5 billion IBM contract in 2004 and returned 4,000 workers to internal payroll[3]. Starting in 2012, General Motors flipped its IT from 90% outsourced to mostly internal under CIO Randy Mott, adding 10,000 roles over five years[4].
India did not fail at software engineering. Cheap body shops failed. Once enterprises built real Global Capability Centers with senior ownership, India grew to more than 1,700 engineering hubs[5]. The mistake was thinking cheap code equals good software.
The AI Reversal Is Starting #
The first corporate reversals of the AI wave are now visible.
Klarna cut roughly 700 customer service roles for AI in 2024 and projected $40 million in annual savings. Resolution time looked fast on dashboards, but quality dropped on complex disputes. In May 2025, Klarna admitted it went too far on cost and started rehiring humans for a hybrid model[6]. Gartner expects half of companies that cut staff for AI to rehire them by 2027.
Software delivery now has its own cleanup market. Rushed teams build quick prototypes with AI tools, then hit security leaks and broken architectures. Cleanup agencies like Leanware charge $4,000 to $12,000 just for an initial audit, followed by weeks of hardening[7]. Building fast with AI and paying for rescue later is often not cheaper than building things properly from day one.
Even experienced teams do not get automatic speedups on complex codebases. A 2025 METR study of experienced open source developers on mature repositories found AI tools made them 19% slower, even though they felt 20% faster[8]. A February 2026 NBER survey of nearly 6,000 executives found 89% saw no measurable productivity gain from AI over three years[9].
Non-Functional Requirements Are Where AI Breaks #
Generating code that runs the happy path is easy now. Veracode tested more than 100 models and found syntax pass rates near 100%. Production software rarely fails on syntax. It fails on security, accessibility, maintainability, and stability.
Security benchmarks show a clear gap. In Veracode's October 2026 tests, AI code failed 38% of security tasks overall[10]. Top frontier models still failed more than one in four security tests. Cross-site scripting passed 18% of the time, log injection passed 12.6%, and Java scored 31%. Public scans of AI built apps also found hundreds of leaked secrets and open databases without access rules.
Accessibility shows the same drop. In February 2026, WebAIM scanned one million homepages and recorded the first backward slide in six years, with 95.9% carrying detectable WCAG failures[11]. AudioEye tested five leading AI tools on 15 websites targeting WCAG 2.2 AA[12]. Every single site failed basic Level A, averaging 55 issues per page and 174 on e-commerce pages. Nearly half of those bugs only appear during live runtime checks with keyboards and screen readers.
Maintainability and delivery stability face the same strain. GitClear tracked 623 million code changes through mid 2026[13]. Refactored or moved code dropped from 21% in 2022 to 3.8% in 2026. Copy pasted code rose from 9.4% to 15.7%, and updates to legacy code fell 74%. Google's DORA reports linked higher AI adoption to lower delivery stability in both 2024 and 2025. LinearB's 2026 benchmarks show AI pull requests wait 5.3x longer for review and reach a 32.7% acceptance rate, compared to 84.4% for human pull requests[14].
Quick does not mean bad code if governance and tests are in the pipeline. The real risk is shipping generated code without checking runtime behavior.
Why Senior Experience Rises in Value #
Payroll data confirms who gains value in this shift. Stanford researchers analyzing ADP records found employment for software developers aged 22 to 25 fell roughly 20% from its late 2022 peak[15]. Over the same period, employment for developers aged 35 to 49 grew 6 to 9%. US Bureau of Labor Statistics data from May 2024 to May 2025 shows computer programmers writing code to spec dropped 16%, while systems analysts grew 4.4% and broad software developer roles grew 2%[16].
Fred Brooks explained this in 1986. Typing syntax is accidental complexity. Understanding the business domain, managing scope, and keeping a system stable over years is essential complexity.
AI also creates comprehension debt. In an Anthropic study of 52 engineers, AI users scored 50% on a comprehension quiz compared to 67% for hand coders, with the biggest drop in debugging[17]. When nobody understands the codebase, every outage turns into a crisis.
Where the 2000s Comparison Breaks #
The 2000s parallel is helpful, but three structural facts are different today.
First, speed. Offshore teams scaled with human hours. Previously, a senior engineer could audit code faster than a junior could write it. AI flips that ratio. One developer or agent can flood reviewers with large pull requests in minutes.
Second, accountability. A 1979 IBM training manual stated that a computer can never be held accountable[18]. Offshore engineers can learn, challenge bad specs, and take ownership. An AI model never pushes back on a bad premise.
Third, the junior talent pipeline. Offshoring moved junior roles to India, where those developers still grew into senior architects. AI automates entry level tasks everywhere at once. If companies stop hiring and mentoring juniors now, the industry will lack senior engineers in five to seven years.
How We Should Work With AI #
AI is a strong tool when clear engineering rules sit around it.
Use AI inside your delivery pipeline for linting, static analysis, security scanning, PR summaries, and boilerplate. Keep automated pipeline gates strict.
Keep merges and UI checks strictly human. Every pull request needs a human owner who understands the change. 100% coverage across every part is not needed. Test components and interfaces where it makes sense. Pair automated browser checks with manual testing using keyboards, screen readers, and braille tools. Refresh code continuously with each release as part of daily delivery instead of saving cleanup for a separate rewrite.
Protect your team and talk directly with people. Make clear to stakeholders that fast code generation does not erase review time, testing, or NFRs. If they want more scope on the same timeline, you push back other features or sell additional manpower.
Software engineering was never about typing faster. It is about owning what runs in production.
References
Deloitte Consulting (2005) via ZDNet, Study: Outsourcing falls from favor: zdnet.com
CNET (2003), Dell drops some tech calls to India: cnet.com
InformationWeek (2004), Insourcing? JPMorgan Reverses Outsourcing Deal With IBM: informationweek.com
Metis Strategy, GM CIO Randy Mott IT Transformation: metisstrategy.com
GCC Pulse (2025), The Pioneers of India's GCC Movement: gcc-pulse.com
Future Skills Academy (2025), Klarna reinvests in human support: futureskills.academy
Leanware (2026), Fix Your Vibe-Coded App: leanware.co
METR (2025/2026), AI Developer Productivity RCT & Update: metr.org
Startup Fortune (2026), Stanford Payroll Data & NBER Executive Survey: startupfortune.com
Veracode (Oct 2026), Fall 2026 GenAI Code Security Report: veracode.com
WebAIM (2026), The WebAIM Million: webaim.org
AudioEye (2026), AI Coding Tools Do Not Write Accessible Code: audioeye.com
GitClear (2026), The AI Code Quality Maintainability Gap: gitclear.com
LinearB (2026), The AI Code Review Bottleneck: linearb.io
SoftwareSeni (2026), Stanford/ADP Data on Junior Developer Employment: softwareseni.com
Seldo (2026), US BLS Data on Junior Programmers: seldo.com
Augment Code (2026), Comprehension Debt in AI Code Review (Anthropic RCT): augmentcode.com
IBM, AI Decision-Making (1979 Accountability Rule): ibm.com
Comments
Comments are hosted by EchoThread. Loading them connects your browser to their servers in the US. Details in the privacy notice.