A story keeps circulating in engineering communities, and it is worth your time even if you have never written a line of code. A junior developer at an "AI-first" company was reportedly fired after AI-generated code reached production and broke it. The version that went viral on Reddit, and got picked up across developer blogs, carries a detail that makes people wince. The company had pushed everyone onto the AI coding tool Cursor and tracked who used it most. When the code failed, the person let go was the most junior one in the room. The manager had reviewed the code with AI and merged it.
I want to be honest about that story up front. It is an anonymous, second-hand Reddit account, and neither I nor the blogs that amplified it can independently verify the specifics: the person's tenure, the company, even the country. Treat it as a parable that went viral, not a confirmed case. What makes it worth writing about is that the pattern it describes is documented elsewhere, on the record, and it is exactly the trap a small business can walk into this year.
The part that is actually on the record
You do not have to rely on an anonymous post to see the trend. In August 2025, Coinbase CEO Brian Armstrong publicly told engineers to onboard onto AI coding tools like Cursor and GitHub Copilot by the end of the week, and said he let go some who did not do it without a good reason (TechCrunch, Fortune). He has since said roughly 40 percent of Coinbase's daily code is AI-generated, with a goal above 50 percent (The Block).
Coinbase is not the company in the Reddit story. They are two separate things. But together they sketch the same management instinct: mandate the tool, raise the speed expectation, and measure adoption. The piece that quietly goes missing is the one that actually protects you.
Why this bites a small business harder
Here is the thing that makes AI-written work dangerous in a way a sloppy human draft is not. It looks finished. The code is clean and well-formatted, the copy is fluent, the spreadsheet formula looks right. The defects hide in the edge cases and the assumptions, the places only experience catches. A six-months-in junior cannot catch them yet, and neither can a busy owner skimming output between four other jobs.
A big tech company has senior engineers to absorb that risk. A small business usually does not. So the failure mode is sharper. The bookkeeper who lets AI categorize invoices ships a wrong number to a client. The one-person shop that ships AI-written code hands a customer a bug nobody read. The marketer who auto-generates a campaign sends a confident, plausible, wrong claim to the whole list. The work looked done, so nobody slowed down to check.
Build the guardrail, not the dashboard
The mistake in every version of this story is the same. Leadership measured the wrong thing. Tracking who uses the AI tool the most is a dashboard. It tells you adoption and nothing about quality. What you actually need is a guardrail, the step that catches the confident mistake before a customer sees it.
If you run a team of one to twenty and you are bringing in AI tools, pair every tool with these instead of a usage chart:
- Name a human approver for anything that reaches a customer. Not a glance. A named person who signs off, and who is allowed the time to actually read it. If that person is you, block the time.
- Write down a permission to slow down. Say out loud that hitting the deadline with broken work is worse than missing it. People ship plausible-but-wrong output when they believe speed is the only thing being scored.
- Keep a no-blame habit when something breaks. When AI output causes a problem, fix the review step that let it through, not the most junior person who touched it. Fear makes people hide mistakes, which is how the next one reaches production.
- Never make AI usage a performance metric. The moment "Cursor usage" or "prompts per week" shows up in a review, you have told people to generate volume, not to think. You will get volume.
- Let the tool draft, never decide. AI is a fast first draft on tap. The judgment about whether the draft is right stays with a person who understands the stakes.
None of this slows you down much. A review step on customer-facing work costs minutes. Shipping a confident error to a paying customer costs the customer.
If you want the wider playbook for running a lean operation on cheap AI tools without these traps, that is the whole argument of my book The $20 Dollar Agency (search the title on Amazon Kindle). The short version is on this blog, in the posts below.
Related reading
- What AI Coding Agents Leave Out: the specific things AI ships without, and why review is still your job.
- Stop Prompting, Start Designing Loops: building a reliable process around AI instead of hoping for a good prompt.
- How a Small Business Runs AI Agents Without a Surprise Bill: the cost guardrails that belong next to the quality ones.
- Build a Small-Business AI Front Desk: two AI tools to adopt, with the handoff rules built in.
- Learn to Secure AI, Then Build With It: the training side of adopting AI responsibly.
Fact-check notes and sources
- Coinbase AI mandate: Brian Armstrong told engineers to onboard onto AI coding tools by the end of the week and let go some who did not, August 2025 (TechCrunch, Fortune). Roughly 40 percent of daily code AI-generated, goal above 50 percent (The Block).
- The viral firing story: an anonymous Reddit account, covered by DEV Community, Storyboard18, and Stack Junkie. The community reaction overwhelmingly blamed leadership, not the engineer. The specific personal details are not independently verified, so I have treated the story as an illustrative, unconfirmed account rather than established fact.
- Cursor is a real AI code editor from Anysphere; Coinbase purchased enterprise licenses for it and GitHub Copilot per the reporting above.
This post is informational, not legal or management-consulting advice. The Reddit account described here is unverified and presented as illustration only. Mentions of Coinbase, Cursor, GitHub, and other third parties are nominative fair use. No affiliation is implied.
