AI-Assisted Business Automation

The problem

There's a wide gap between "AI wrote some code" and "AI-assisted development that's actually trustworthy in production." The first is a demo. The second requires a workflow — one where AI tools handle the parts they're genuinely good at (drafting, boilerplate, exploring an unfamiliar API, first-pass tests) while a human still owns system design, correctness review, and the decision to ship.

The workflow

Business problem first. Every task starts with the actual operational problem — a catalogue mismatch, a slow report, a manual process eating hours every week — not a prompt in search of a use case.

System design stays mine. Architecture decisions, data models, and failure-mode thinking happen before AI tools touch the problem. This is the part that determines whether the resulting system is maintainable six months later, and it's the part I don't delegate.

AI-assisted implementation. Different tools for different jobs: Claude Code and Codex for larger, multi-file implementation work and refactors; Cursor for fast in-editor iteration; Gemini for research and second opinions on approach. I treat model output as a strong first draft, not a finished product.

Tests and human review. Nothing generated ships without tests and a real read-through — the same bar I'd hold a human collaborator's pull request to. AI-generated code gets the same scrutiny as anyone else's, arguably more, because it doesn't yet have a track record with this specific codebase.

Deployment and monitoring. Once something ships, it's watched the same way any production change would be — logging, alerting, and a rollback plan, not "the AI said it works."

Why this matters

The honest pitch isn't "AI writes my code." It's that AI tools meaningfully speed up the implementation phase of systems I'd have designed and validated either way — the catalogue pipelines, marketplace sync jobs, and internal tools described in my other case studies get built faster because of this workflow, not because judgment got outsourced to a model.

What I'm still refining

The weakest link right now is turning implicit operational knowledge — the stuff I know about how a specific supplier feed misbehaves, or how a marketplace's edge cases work — into context the AI tools can actually use, instead of re-explaining it every session. That's the direction most of my current tooling investment is going.