A single tariff policy can determine a company’s survival. A policy announced today can reshape your business tomorrow, and six months later, it can reorganize your entire supply chain. The problem is, if you don’t discover that policy in time, you don’t even have a chance to respond.

Read also: A Global Trader’s Guide to Building a Digital Presence That Converts

I experienced the same fear. As an overseas futures trader for 50 months, I traded in the S&P 500, crude oil, and U.S. Treasury markets. Every morning before the market opened, I manually tracked policies that could move markets and spent five hours daily on risk management. A single tariff policy, a single central bank announcement could shake the market, and I had to adjust positions accordingly. But one question haunted me: “What if I missed something?”

Eight months ago, As a complete non-programmer, I had never touched Python or written a single line of code—not even “Hello World”. Today, I built a system that reduced approximately $60,000 worth of work to $18.42 and generates 125,332 lines of raw screening logs. The secret? I didn’t learn to code. I asked AI to translate my worries into requirements.

Why Policy Tracking Feels Impossible

It’s not carelessness that makes it hard to keep up with policy changes. It’s because the task itself is structurally impossible for most organizations.

The flood of documents from Federal Register, USTR, CBP, and OFAC is overwhelming. Even with 20-25 senior experts working full-time for a week, designing 959 searches, reading 5,755 documents, and determining future implementation feasibility would cost $60,000-75,000 per execution.

But the real problem isn’t cost. It’s speed. Policies are announced today, but assembling an expert team and starting the work takes a month. By then, your competitors have already adjusted their supply chains. In trade, a slow response equals no response.

Four barriers make traditional approaches nearly impossible. First, recruiting 20-25 experts who understand both Federal Register’s complex structure and semiconductor tariff context is unrealistic. These specialists are already senior partners at major law firms. Second, even if you could assemble such a team, onboarding them takes weeks while policies are published daily. Third, human fatigue means judgment quality degrades after reviewing hundreds of documents. Fourth, sustaining weekly updates would require full-time employment of this team, costing $3-3.75 million annually.

What if there was another way?

From Worries to Requirements

The breakthrough came not from learning to code, but from clearly expressing what kept me up at night.

During 26 weeks of AI prompt engineering, I knew exactly what would go wrong. When an AI session hit the token limit and work was interrupted mid-task, I didn’t just restart. I told the AI: “I want to understand the progress and continue from where I left off. How can I do that?”

When API calls randomly failed, I explained my fears: “What if it fails during 900 queries at 3 AM?” The AI translated each worry into code—auto-save functions, resume features, error recovery systems.

When I needed to verify AI judgments but had thousands of results, I said: “I want to check the AI’s filtering rationale and the actual documents directly.” The AI wrote code that logs verification rationale along with URLs and document snippets for direct confirmation.

This was the secret. I didn’t need to learn Python syntax or debugging techniques. I just needed to clearly explain my problems to the AI. When errors occurred—and they did—I would copy the error message, attach the code, and ask: “Why doesn’t this part work?” The AI would review, explain the issue, and provide a corrected version. I’d run it immediately in Google Colab to check if it worked. If problems remained, I’d ask again. Gradually, through dozens of iterations, the system took shape.

The key point is this: I didn’t become a developer. My 26 weeks of wrestling with prompt engineering taught me what could go wrong—and that knowledge became the foundation for building verifiable automation.

 

How Three Tools Became One System

When I explained my worries to AI, I discovered I needed three distinct tools, each playing a unique role. I call this approach “The Trinity.”

Claude handled the planning. Over seven months, I used Claude to design 15,000 lines of prompts that mapped out a complex information structure: Strategy, Product, HTS codes, Tariff Rates, and Timelines. Claude helped me think through the search strategy—which government sources to query, which keywords to use, how to filter for future implementation potential. This was a process of learning domain knowledge and translating it into natural language instructions.

Gemini API handled the execution. Once the prompts were ready, Gemini API conducted 959 searches across Federal Register, USTR, CBP, Commerce, and OFAC. It screened 5,733 documents, using natural language understanding to judge: “Does this document indicate future implementation potential?” Gemini’s strength is context—it reads between the lines, understands regulatory language, and makes nuanced judgments that pure keyword matching cannot achieve.

Python handled control and verification. While Gemini judged in natural language, Python cross-verified using precise keyword filtering. Every 10 searches, Python auto-saved progress. If interrupted, it resumed from the last checkpoint. Python logged every single judgment—125,332 lines documenting which documents were flagged and why. This created an audit trail I could verify by clicking URLs to read original Federal Register snippets.

Gemini provided flexibility—understanding context and making judgment calls. Python provided accuracy—checking semiconductor-related keywords and filtering non-semiconductor terms. AI’s flexibility met code’s precision, and each compensated for the other’s weaknesses.

Every result included a URL and text snippet. One click took me to the original Federal Register document so I could verify whether the AI’s judgment was correct. This wasn’t blind trust in AI. This was AI-assisted collection with human verification.

What this process produced: $18.42 in Google API costs, 7 hours of execution time, 125,332 lines of raw screening logs flagging 245 candidate documents for human review. The Trinity worked because each tool did what it does best—and I orchestrated them without writing a single line of code myself from start to finish.

 Why AI Still Needs You

Despite all these mechanisms, AI is still not perfect. The 245 documents flagged weren’t information—they were candidates requiring human judgment. Some covered multiple industries in a single Federal Register entry, a structural reality I learned to accept rather than eliminate.

In Week 29 of my project, I learned that complete filtering was impossible. Federal Register’s document structure creates unavoidable edge cases. When a document announces both semiconductor tariffs and steel tariffs simultaneously, how should AI classify it? If I added it to a blacklist to filter it out, I might miss legitimate semiconductor policies announced in similar multi-industry documents. If I kept it, I’d have false positives.

This trade-off between precision and recall is fundamental. Adding more blacklist rules increases precision but decreases recall—you filter out noise but might miss important signals. The problem is unpredictable. New document structures emerge constantly, and maintaining blacklists for every edge case becomes impossible when you’re running weekly updates.

That’s why human verification remains essential. When AI identifies 245 potentially relevant documents, I don’t blindly trust they’re all semiconductor-only announcements. I click each URL, read the snippet, and judge: “Is this actually about semiconductors and future implementation?” For complex cases where a document covers both semiconductors and other industries, I make the call on whether to include it.

The collaboration model is clear. AI collects and filters vast amounts of raw documents at inhuman speed and consistency. You verify based on domain expertise and make final judgments on ambiguous cases. AI’s speed combined with your judgment creates trust.

I spent eight months on this project and it’s still not perfect. Room for improvement remains. Research continues. But that’s precisely the point—this is not about AI replacing expertise. It’s about AI amplifying what experts can accomplish when they understand both the tool’s capabilities and its limitations.

The Real Value Beyond ROI

This is not simply about cost savings. AI has transformed what was previously in the “impossible domain” into something executable at the push of a button.

Consider four barriers that made comprehensive policy tracking impossible before. First, expert recruitment. Specialists who understand Federal Register’s complex structure and semiconductor tariff nuances are extremely rare globally. They’re already senior partners earning top-tier salaries. Gathering 20-25 of them for a week is unrealistic. AI is ready instantly, with no recruitment delay.

Second, the time barrier. Policies are announced today. Assembling and onboarding an expert team takes a minimum of one month. By then, the policy window has closed. AI delivers results in six hours, making raw findings actionable rather than stale.

Third, quality consistency. Humans tire. Even expert teams struggle to maintain consistent judgment criteria across thousands of document reviews. Gemini API becomes a tireless brain, maintaining unwavering standards from the first search to the 959th. The 125,332 lines of logs document not just results, but also the AI’s 7-dimensional verification rationale, enabling users to strengthen and verify the judgment basis.

Fourth, annual sustainability. Policies change weekly, requiring weekly updates. Twenty-five experts working year-round costs approximately $3-3.75 million annually. An AI system costs about $2,600 per year for 52 weekly executions. ROI is roughly 1,200-1,500 times.

But here’s the deeper value: AI didn’t replace expertise—it made previously unattainable tasks accessible. It made “25 expert brains for one week” available for $50, anytime, reproducibly. It converted tasks that organizations couldn’t even attempt into workflows they can run on-demand.

For those who know how to use these tools properly, AI isn’t a bubble. For those who plan well, it’s amplification. As proof, I continue paying for Claude’s Max Plan every month since May 2025.

What This Means for You

You already have what’s most important: domain expertise.

As a trade professional, supply chain manager, or logistics leader, you know which policies matter to your business. You understand the impact of tariff changes. You recognize critical information when you see it. This knowledge is irreplaceable.

AI’s role is straightforward: it’s a tool that translates your domain knowledge into large-scale data collection and organization. It handles the scale humans cannot—reading thousands of documents, maintaining consistency across hundreds of judgments, logging every decision for verification.

The collaboration model is this: AI collects and organizes at machine speed and scale. You verify and decide based on expertise and business context. Your expertise multiplied by AI’s scale makes the previously impossible suddenly possible.

This extends far beyond semiconductors. Supply chain managers can track USTR and CBP regulatory changes. Compliance teams can filter for future-effective policies only. Logistics leaders can monitor port and customs regulation updates. The same Trinity structure applies—only the industry focus changes.

What this learning process produced is transparent: 959 searches, 5,755 documents screened, 245 candidates identified for verification, 125,332 lines of logs generated, $18.42 cost, 7 hours execution time. These aren’t claims. These are measurements anyone can reproduce.

The Humble Truth

Eight months of work, and there’s still room for growth. Research on better filtering methods continues. This isn’t completion—it’s progress.

But for those who use tools properly, this represents revolution. For those who plan carefully, it’s expertise amplification. For those who verify rigorously, verification becomes possible.

The core message is this: Your planning determines AI’s output quality. Your verification ensures trustworthiness. Your domain expertise makes AI useful rather than dangerous.

AI is not magic. It’s a powerful tool that amplifies what you already know. It doesn’t replace your judgment. It extends your reach. It doesn’t eliminate the need for expertise. It multiplies what expertise can accomplish.

This article captures the essence of an eight-month learning journey in building AI usage capabilities. Information mapping strategies, prompt execution techniques, verification frameworks—contains lessons learned that might help others avoid the mistakes I made and develop your own AI usage capabilities.

There is much more to share, and I hope it can help the trade community.