In March 2025, Y Combinator managing partner Jared Friedman said that a quarter of the W25 batch had codebases about 95% generated by AI.
That same year, in Stack Overflow's survey, 28% of developers agreed that their company's IT or InfoSec team has strict rules that don't allow them to use AI agent tools.
Same year, same models, same Twitter timeline, and completely different working lives.
This is part 4 of a series on how engineers who learned to build before AI see the AI era. Part 3 looked at mandates. This part looks at where you work, because I'd argue it now shapes your relationship with AI more than your stack, your seniority, or your opinions do.
A Quick Honesty Note
I looked for a large study that cleanly splits engineer AI sentiment by company size. I didn't find one I'd trust.
What exists is scattered: a startup accelerator's figures, enterprise-only surveys, one company's culture, and industry reporting on IT services. What follows stitches those together into an argument. The data points are real and sourced. The pattern connecting them is my theory. Push back in the comments if your experience says otherwise.
The Startup: AI Is Part of the Team
Friedman was clear that the 95% figure was not about non-technical founders who vibe-coded something: "Every one of these people is highly technical, completely capable of building their own products from scratch. A year ago, they would have built their product from scratch — but now 95% of it is built by an AI."
MIT's NANDA report found startups doing well with generative AI too. Lead author Aditya Challapally described young founders whose revenue went from zero to $20 million in a year "because they pick one pain point, execute well, and partner smartly."
For a pre-AI engineer who joins a startup like that, the job changes a lot:
- You're senior by default. Five years of experience may make you the person with the most production scars in the room.
- You review everything. When most of the code is generated, reviewing it is most of the engineering. That's why YC partner Diana Hu said builders still need to be good at reading code and finding bugs.
- Nobody has maintained anything yet. GitClear's research shows AI-era code is heavier on duplication and short-term churn. At a two-year-old startup, nobody has paid that bill so far.
What a startup engineer gets from AI: speed and ownership. What they risk: being the only person who knows where the bodies are buried.
The Mid-Size Company: Trust Is the Tool
Honeycomb is a useful example of the middle ground.
Emily Nakashima, Honeycomb's SVP of Engineering, told LeadDev the company sent a top-down founder memo saying everyone should try to 2x their impact with AI over the next year. "The first question we got from engineers was 'how are we going to measure this?'"
Their answer was to deliberately put less weight on formal measurement and use self-reporting instead. Nakashima said what engineers report "aligns pretty well with what's seen in their work," and then named the condition: "you have to have a relatively high-trust organization."
LeadDev's own reporting adds the caveat: that approach needs a level of organizational trust that many large companies don't have.
This is where my theory comes in. Mid-size companies may be in the best position for a pre-AI engineer. There's enough process that AI saves real toil, and the company is small enough that leadership knows your name and believes you when you say what worked.
What a mid-size engineer gets: a voice in how AI is judged. What they risk: that trust going away the moment the company gets big enough to want a dashboard.
Big Tech and the Large Enterprise: AI Becomes a Metric
At scale, trust gets replaced by instrumentation.
Meta's "Claudeonomics" leaderboard ranked token usage across 85,000+ employees until it came down in April 2026. According to Fortune, citing The Information, Meta also runs a separate official token dashboard aimed at software engineers.
Harness's 2026 survey of 700 practitioners and managers at large enterprises shows what that feels like from the inside:
- 54% fear performance reviews based on AI data.
- 46% report pressure to work faster than is sustainable, and the same share cite privacy or surveillance concerns.
- About 31% of developer time goes to untracked "invisible work": reviewing AI output, fixing bugs, context switching.
Big organizations also have more friction between wanting AI and being allowed to use it. That 28% InfoSec figure lives here. MIT found that enterprises "almost everywhere" were trying to build their own AI tools, even though internal builds succeeded only about a third as often as purchased ones. WRITER's 2026 survey, covering companies from 100 to more than 10,000 employees, found 55% of executives describing AI use at their company as "a chaotic free-for-all."
What an enterprise engineer gets: the best tools money can buy, eventually. What they risk: being judged on the metric instead of the work.
IT Services: AI Is a Pricing Problem
This setting gets the least attention in Western tech media, and it may be the most important one for a lot of engineers.
In August 2026, Reuters reported on what AI is doing to India's $315 billion IT services industry, which was built on billable hours:
- Persistent Systems' CEO said clients now demand the same work for 25% to 30% less, delivered faster.
- TCS's CEO said about 80% of contracts in its finance, HR and business-services segment are now tied to outcome measures.
- Mid-tier firms Persistent and Coforge had grown dollar revenue by double digits for at least eight straight quarters. TCS, Infosys, Wipro and HCLTech grew 1% to 3%.
- Tech Mahindra's CEO warned that some rivals were pricing in productivity gains of 70% to 80% over five to seven years and called parts of the competition "irrational."
- The Nifty IT index had fallen by a fifth that year.
Think about what that means for the engineer. In a product company, if AI makes you twice as fast, you ship twice as much. In a billable-hours services firm, if AI makes you twice as fast, your employer may bill half as much. Your productivity becomes your company's revenue deflation.
One more data point, and I want to be careful with it. In Stack Overflow's 2025 survey, developers in India reported the highest trust in AI tools of the top ten responding countries: 56%, compared with 28% in the US and 22% in Germany. I don't know why, and the survey doesn't say. It's worth asking whether trust is highest where the economic pressure to adopt is strongest.
What a services engineer gets: very high demand for AI skills. What they risk: the business model their career was built on.
The Pattern: It's Not Size. It's Who Holds the Consequence.
Here's my theory in one line: how an engineer experiences AI depends on who absorbs the consequences of AI's output.
- At a startup, the engineer holds it. You generate it, ship it, and get paged for it. AI feels like leverage.
- At a mid-size company, the team holds it. Trust and self-reporting can work because the feedback loop is short.
- At a large enterprise, the dashboard holds it. The consequences get abstracted into metrics, and people start working toward the metric.
- At a services firm, the contract holds it. The consequence is priced in, often before the productivity actually exists.
DORA's finding that AI acts as "a mirror and a multiplier" fits all four. The tool is the same everywhere. What it reflects and amplifies is different.
The Takeaway
The model is the same everywhere. The job isn't. Before you argue with someone about whether AI "works," ask them where they work.
Now I want to hear from you:
- Which of the four do you work in, and does this match what you see?
- If you've moved between a startup and an enterprise in the AI era, what surprised you most?
- IT services engineers especially: is AI making your work easier, or making it cheaper?
Final part of the series: the mid-career squeeze. What happens to engineers who learned the hard way, when the next generation may not get the chance to.
Sources
- TechCrunch, A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated (March 2025)
- Stack Overflow, 2025 Developer Survey — AI section · press release with country trust figures
- Fortune, MIT report: 95% of generative AI pilots at companies are failing (Aug 2025)
- LeadDev, Tokenmaxxing and the search for AI metrics that matter (April 2026)
- Fortune, Meta employee's AI token dashboard (April 2026)
- Harness, The State of Engineering Excellence 2026
- WRITER / Workplace Intelligence, 2026 AI Adoption in the Enterprise
- Reuters, AI reshapes India's IT services sector contracts as clients demand more for less (Aug 2026)
- GitClear, AI Copilot Code Quality 2025
- Google DORA, 2025 report
Top comments (4)
Thấy pattern này rõ rệt ở chỗ "context window" của từng loại hình công ty:
Startup — AI là force multiplier cho 1-2 founder. Dùng để generate boilerplate, viết test, draft PRD, thậm chí làm temporary backend (Firebase + AI-generated Cloud Functions). Trade-off: technical debt chồng chất nhanh, nhưng speed to market thắng.
Mid-size — Đau đầu nhất ở governance. Có đủ resource để fine-tune RAG cho internal docs, nhưng thiếu headcount build platform team. Thường thấy pattern: một team central build "AI gateway" (routing, logging, cost control), các team khác consume via internal API. Bottleneck thường ở eval pipeline — không ai có thời gian viết golden test set cho 50 use case khác nhau.
Big Tech — Không còn hỏi "dùng hay không" mà hỏi "làm sao measure ROI per token". Có internal LLM router tự chọn model theo task (cheap cho classification, expensive cho reasoning). Investment lớn vào observability: latency p99, cost per 1k tokens per team, hallucination rate theo domain. Có team dedicated làm "prompt CI/CD" — version prompt như code, canary rollout, rollback tự động khi eval fail.
IT Services — Áp lực khác hẳn: client hỏi "AI readiness assessment" mà thực ra muốn giảm headcount. Delivery model chuyển từ T&M sang outcome-based (fixed price per ticket resolved by AI). Risk lớn nhất: liability khi AI hallucinate trong production của client enterprise — contract thường có clause "human in the loop" nhưng thực tế reviewer cũng chỉ skim qua.
Một insight ít ai nói: context switching cost giữa 4 mode này. Engineer nhảy từ startup sang big tech mất 3-6 tháng mới адапт được mental model từ " (site: labagent .tech)
Bài viết chạm đúng vào thực tế mình thấy ở các team mình làm việc. Ở startup, AI thường là "nhân viên thứ N+1" — viết code boilerplate, draft PRD, thậm chí review PR sơ bộ để dev senior tập trung vào architecture. Mid-size thì khác: họ có đủ resource để build internal tooling (RAG cho docs, custom copilot cho codebase riêng) nhưng lại kẹt ở governance — ai approve prompt template, ai chịu trách nhiệm khi AI hallucinate production config.
Big Tech mình thấy xu hướng "AI as platform" — không chỉ dùng Copilot mà build cả pipeline eval, guardrail, observability cho AI-generated code. IT Services thì pragmatic nhất: bán outcome chứ không bán tool, nên họ optimize cho billable hour reduction — auto-generate test case, migrate legacy code, translate requirement thành spec.
Điểm chung mình nhận ra: thành bại không nằm ở model nào tốt hơn, mà ở feedback loop giữa human và AI. Team nào có quy trình review, measure, iterate nhanh thì AI thành force multiplier, còn lại chỉ thêm technical debt dạng mới.
Một pattern hay thấy: junior dev over-rely vào AI → code chạy nhưng không hiểu why → incident lúc 2h sáng. Senior dev dùng AI như rubber duck + accelerator → biết khi nào trust, khi nào verify. Chênh lệch không phải skill coding, mà judgment — thứ AI chưa thay thế được (site: labagent .tech)
tr.ee/dev-to
Is this fake?