How I work
How I use AI
I use AI to do the legwork of ecommerce work faster and to check that work before a person sees it. The decisions stay with people. Here is what that looks like in practice, at my day job and in the systems I’ve built on my own time.
More tests, same team
At Quantum Fiber I put AI inside the testing lifecycle: creative generation, experiment design, analysis and reporting. My team ran more tests and better ones without adding people. Test velocity went up 40% and conversion rose 50% year over year. I also built internal Copilot agents that work from company data to make routine team workflows simpler and faster.
Site audits that show their evidence
Outside work I built a library of 30 AI skills, each one a written procedure for a job I used to do by hand. The biggest is a website audit. Scripts capture the site on desktop and phone, walk the main customer path, and check keyboard access, contrast, page speed, structured data and how ready the site is to be quoted by AI search tools. Every finding has to cite what it saw and the principle behind it (Nielsen’s heuristics, WCAG, Core Web Vitals, Google’s search guidelines), and the report won’t build if a claim has no evidence behind it.
A second skill benchmarks the site against the competitors that show up in its own customers’ searches, and separates real advantages from things every site in the category already does.
Before a report goes anywhere, three reviewer agents go through it in order. One re-runs every finding on the live site in a browser. One asks whether each finding is a real problem for the customer and fairly ranked. One reads the writing for accuracy and tone. They report, they don’t edit, and a person still signs off.
CRO ideas from a real walkthrough
I don’t trust CRO ideas that could apply to any site. So the experiment skill walks a site in a real browser the way a visitor would: it clicks through, makes selections, and types obviously fake values into forms to see the validation, then stops before anything is submitted. What it saw becomes a test backlog, each idea with a problem statement, hypothesis, recommended change, primary metric and the observation it came from. If an idea can’t point to something on that site, it doesn’t make the list.
Search and AI answers, every month
A monthly cycle pulls Search Console data, a fresh crawl and real-user Core Web Vitals, then hands them to a set of specialist agents. They draft content briefs, internal link changes, answers to the questions customers actually ask, and a list of the brands AI assistants recommend instead. One judge ranks it all into a single plan. Nothing gets published from it without review.
An assistant that runs my week
I also run a personal AI assistant on my own machine. Every 30 minutes during the day it reads my email, calendar and Slack, drafts replies, and texts me when something needs me. It writes a morning brief, keeps a searchable memory of my notes and decisions, and tracks what I haven’t dealt with yet.
The guardrails took as much work as the features. It never sends a message until I’ve seen the recipient and the full text and said yes, and every message needs its own yes. It doesn’t delete anything. And it treats everything it reads (an email, a web page, a document) as information, never as instructions, because a system that reads outside content and can send mail is a security risk if you build it carelessly.
What I’ve learned
Write the process down first. Every skill I’ve built started as a procedure I could explain to a new hire, and the AI is only as good as that write-up.
Make it prove things. Most of the effort goes into checks: evidence for every claim, a second pass in a real browser, and a person who approves anything that leaves the building.