Over the past two years I built two production admin backends — 好易美’s hym-manager and the cat-litter one, mzmeso-manager.
The two projects look like twins: Vue3 + Vite + Cloudflare Pages + Supabase, 8000+ lines of business code between them, with orders, products, and marketing rules changing every single day.
The thing I did in them that looks least like engineering is, in fact, the most underrated: “AI infrastructure.”
But the requirements-translation step is where both the business side and the purely technical side struggle to explain things clearly — that’s the root of the problem.
Why What AI Writes Doesn’t Work for the Business Side
Many people’s expectation of AI is: write an internal SaaS for the business team to use, and hope development speed doubles.
What usually happens instead: AI writes a demo, and the business side can’t use it.
Business says “this campaign has to go live right now”; engineering says “that’s 8 files to change plus the full test suite” — both are right, but the requirement has already mutated along the way.
The real leverage of AI acceleration isn’t that AI writes fast enough. It’s that what AI writes has to land in the hands of people who understand the business.
Step One: Pull the Business Rules Out of the Code
Marketing rules, coupon rules, pricing rules — the things that change constantly shouldn’t be buried deep in the code.
Buried in code means every change requires finding a developer, and every change re-runs the whole “requirements translation” step.
Pull them out and turn them into config the business side can actually read: YAML, JSON, or a Feishu spreadsheet.
Change the config and it takes effect immediately. The business side doesn’t need to understand code, only the business.
Step Two: Give the Business Side an Agent That Understands Their Business
Not a general-purpose chat like ChatGPT — an Agent that can read their business code, call the APIs, and run the flows.
The AI assistant in hym-manager is “two-phase”: first it produces a preview (zero writes), and only after confirmation does it execute the full flow.
This is to prevent accidental operations — what’s being changed is real business data, and you can’t just let the Agent go all-in in one shot.
The business side doesn’t need to wait for a developer to execute. They ask the Agent a question, and the Agent reads the code, calls the APIs, and gives them an answer.
Step Three: Business-Side Changes Skip the CI/CD Queue
Business rules aren’t core transaction logic. If it’s wrong, just roll it back; there’s no need to run the full test suite.
The customer manager on mzmeso can open custom entitlements for a specific customer directly, without finding the boss to sign off every time.
In these scenarios “fast” beats “correct,” and “iteration” beats “perfection.”
The Numbers This Playbook Actually Produced
Adding up all three steps:
hym’s ops team can finish changing a campaign rule during lunch break, and it goes live at 2 p.m. sharp. Previously, waiting for a developer to be scheduled took 3 days.mzmeso’s customer managers can directly open custom entitlements for a specific customer, without finding the boss to sign off every time.- The business side’s fondness for the system went from “being dragged along by engineering” to “I can set the pace.”
The real leverage of AI acceleration isn’t replacing developers. It’s removing the “requirements translation” step, the one most likely to distort.
That Repeatedly-Confirmed Line
Finally, back to that line that keeps getting confirmed —
Specialists do the specialist work.
Nobody understands the business better than the business side. Nobody understands AI better than the engineers. But when it comes to requirements, both the business side and the purely technical side struggle to spell them out.
So let’s stop making them spell it out.
Let the business side touch the code directly, and let the engineers build the infrastructure and the guardrails.
