I don't think AI will make your processes go faster
Many engineers report that large language models make individual coding tasks faster, but have far less impact on overall software delivery because real bottlenecks lie in vague requirements, organizational bureaucracy, and slow cross-team coordination. Commenters contrast small, low-overhead teams that can use AI to prototype and ship features quickly with big enterprises where AI-generated code piles up at existing choke points like design, legal review, deployment and human code review. The emerging consensus is that AI can deliver meaningful but limited productivity gains—often 10–50% on well-specified work—while also introducing new risks around code quality, maintainability and overreliance on auto-generated specs and implementations.
Scope of AI Speedups
- Many agree AI can make coding much faster: boilerplate, CRUD, simple services, tests, and small tools often get 2–10x speedups.
- Several report 10–20% net gain on serious projects once debugging, refactoring, and understanding AI-written code are included.
- A recurring theme: development time is a minority slice of the full lifecycle (requirements, coordination, legal, deployment), so overall project speed barely moves unless processes change.
Impact on Teams and Organizations
- Solo devs and small, aligned teams report “lightning fast” progress and the ability to build things previously out of reach (frontends, tools, niche apps).
- In large orgs, benefits are blunted by bureaucracy, slow approvals, “frozen middle management,” and deployment gates; coding was rarely the bottleneck.
- Some big-company engineers claim 3–10x faster delivery in AI-forward orgs; others at similarly large orgs see minimal or even negative net gains.
Code Quality, Review, and Maintenance
- Strong concern that AI encourages “vibe coding”: huge, messy diffs, lots of dead or redundant code, subtle bugs, and security issues.
- Review and comprehension become the new bottlenecks; reading and validating AI output is often harder than writing focused code.
- Good results seem to require: precise prompts, small scoped edits, strict review, strong tests, and architectural guardrails.
Requirements, Product, and “Spec Bottleneck”
- Many argue requirements and understanding the problem are the true bottlenecks; vague asks in → vague or wrong outputs out.
- LLMs can help structure and elaborate specs, but also generate plausible-sounding nonsense that PMs may not validate.
- Faster prototyping shifts pressure onto product and users: more throwaway iterations, potential “Ikea era” disposable software, and unstable UX for users.
Management, Hype, and Process Change
- Reports of top-down AI mandates, token quotas, and leaderboards; skepticism is sometimes penalized.
- Some see AI mainly exposing existing dysfunction: misaligned incentives, cargo-cult innovation, overstaffing, and process theater.
- Several suggest real gains require rethinking workflows (e.g., using agents across ideation, exploration, coordination, and deployment), not just bolting AI onto old processes.
Non-coding Uses and Long-term Views
- High value reported in debugging, log/trace analysis, docs, onboarding, search, UI mockups, inbox summarization, and non-dev staff building small tools.
- Debate over economics: modest productivity gains may not justify massive investment, especially if token prices rise.
- Long-term outlook splits: some see rapid progress toward agentic systems that can own whole features; others point to failed experiments (e.g., AI-written compilers) as evidence we’re still far from replacing human understanding.