AI does the legwork. Judgment stays human.

The situation

Every marketer now says they "use AI." Most mean they paste a prompt into a chatbot. The work these brands actually need asks for something harder: someone who can identify, test, and implement AI tools into the daily strategic and content workflow of an agency, and make that stick across a team, not just at one desk.

I have a decade-long pattern of doing exactly this, well before "AI" was the headline. The throughline is the same every time: find the tool early, prove it works on real work, then scale the discipline so the whole organization runs on it. I have done this three times, at rising stakes: with a SaaS platform, with a personal AI system I built, and with a working artifact aimed straight at a real premium brand active in China.

One thing I want stated plainly up front, because it sets the standard for everything below: I am a strategist who builds AI workflows, not an AI engineer. I do not pretend the machine has judgment. AI should scale judgment, not replace it. Every workflow I build keeps a human decision gate at each step where taste, brand truth, and risk live. That guardrail is the point, not a caveat.

The decision

The decision underneath all three chapters is the same: treat AI and tooling as an operating advantage to be engineered into the workflow, not a buzzword to be name-dropped. That meant being willing to be the person who evaluates the tool, defends the adoption, builds the templates, and stands up in public to say it works. And it meant drawing a hard line between what AI does (the legwork) and what stays human (the judgment).

The strategy

Three moves, escalating in ambition.

1. Adopt and scale a platform (KAWO). Standardize content operations on a single China social-media management tool, so planning, collaboration, and reporting stop living in scattered spreadsheets, then turn my templates into the agency's default.

2. Build my own AI system (Claude). Construct a personal research → verify → plan → produce workflow with a human review gate between every step, so AI does the legwork and judgment leads.

3. Show it on a real brand (Montagut). Rather than claim the capability, use the system on a live premium brand and hand over a finished artifact.

Execution

Chapter 1, KAWO: found it, proved it, got the organization on it. At Virbac I hit a wall I could not solve with effort: explaining China's fragmented social landscape, WeChat and Weibo and RED, to a global team that simply could not see it. A Google search turned up KAWO, a China social-media management platform. I evaluated it, adopted it, and ran on it for roughly two years as the brand pushed into DTC.

What it fixed was concrete. The editorial calendar moved off a single Excel sheet that took two weeks of internal back-and-forth before anything shipped. Team collaboration across local social, marketing, e-commerce, and agency moved into one shared flow instead of email chains. Reporting moved from screenshots to data that actually told us which of our five consumer product lines earned attention and why. The platform became, in the phrase I used publicly, the brand's "monitoring screen and in-ear monitor."

I then carried the same discipline into APR, where I centralized multilingual asset management and approval workflows. That standardization cut production turnaround roughly 25%, and the approval templates I built were adopted agency-wide, which is the real test of an implementation: it outlived me at my own desk.

The capstone is public and usable openly. In October 2022 KAWO invited me to keynote their launch event as a customer, the "Omniscient Narrator for the Social Media Marketer" talk. It is my own authored, public keynote. Being asked to stand on a vendor's stage and make the case for their tool is third-party proof that I do not just adopt tools. I champion them credibly enough that the vendor wants me selling the value.

Chapter 2, the Claude system: an AI workflow with human gates. The second move is the one few candidates can show: I built my own AI strategy system on Claude Code, the knowledge vault this very portfolio was produced in. It is not a chatbot. It is a chain of skills with a human review gate between each step: research → verify → plan → produce.

• Research (/trend-scan): scan for always-on territories and a scored hot-topic radar.
• Verify (/brand-align): a brand-fit, agenda, and conversion gate that ranks what to build and what to kill.
• Plan (/content-plan): a quarterly calendar with the conversion architecture attached.
• Produce (/content-engine): platform-native drafts in native Chinese.

Between each step sits a review, where I, the strategist, edit, kill, or redirect before the next step runs. That human gate is the quality. It is why this is a chain of skills you supervise, not an autonomous workflow that runs unattended. I also encoded my own China-marketing method into the research, strategy, and analysis skills, so the drafts start at senior quality instead of generic. But a draft is where AI's job ends and mine begins.

I am candid about the limits, because honesty is what makes it real rather than theater. There is no live platform API yet: the research runs on web search plus reasoning and is confidence-tagged, with a clear upgrade path to paste in 千瓜 / 微信指数 / 蝉妈妈 / KAWO exports. The analysis is only as good as the brand assets loaded into it. None of that is hidden; it is documented in the playbook.

Chapter 3, the Montagut demo: show, don't tell. The third move turns the system on a real premium brand. Montagut is a French heritage brand active in China, with an established social program across WeChat, 视频号, and Weibo, including the Winter Wonderland H5 that drew around 10,000 new fans. I did not pick a hypothetical brand. I pointed my system at a real one.

In an afternoon, the system produced a diagnosis. Montagut's heritage (1880 France, in China since 1979, the Fil Lumière shimmer thread) is its asset, not its liability, and the RED trend toward 老钱风 / 复古 / 新中式 rewards brands that are authentically old. The content thesis: flip the "my dad's polo" perception into the proof of authenticity that old-money youth crave. I then designed an always-on content workflow that sits underneath a brand's existing campaign work, systematizing the recruitment layer between the big campaign moments, and drafted sample RED and Douyin content built on it. Everything AI-drafted was then strategist-edited.

Two things are true of this artifact at once: it was made with AI, and it designs a workflow that runs on AI. That is AI and workflow innovation made concrete, on a real account. It builds on a brand's existing work; it does not criticize it.

Result

• −25% production turnaround at APR from standardized asset and approval workflows, and the approval templates were adopted agency-wide.
• A public KAWO keynote (Oct 2022): vendor-validated proof of tool championing, usable openly.
• A finished, runnable artifact for a real premium brand: the Montagut diagnosis-plus-workflow, built in an afternoon and strategist-edited, with a working /content-engine skill behind it that reruns on any brand.

I am deliberately not attaching invented AI productivity metrics. This case is about capability demonstrated, not numbers inflated. The −25% is the one hard figure, and it is source-supported.