10 Days to Production: Building AI Operations for a High-Volume TikTok Seller
A fast-growing apparel brand came to us with a straightforward problem that turned out to be anything but.
They sell on TikTok. Six-figure monthly ad spend. 150+ campaigns. 300+ active creatives. A daily production target of 100 new videos. Two non-technical operators running the whole thing: one managing ads and budgets, the other managing product and supply chain.
The business was growing. The operations were not keeping up.
The Problem Nobody Talks About in TikTok Commerce
TikTok Live can sell 3,000 units of a product in 2-3 days. Production lead time for this brand is 6 weeks. That gap is where money goes to die.
If you don't detect a hero SKU early enough to place a bulk order (30,000 to 50,000 units), you stock out. When you stock out, the algorithm stops pushing the product. Competitors copy the design. By the time your restock arrives, the moment has passed.
The brand launches roughly 15 new products per month. Most won't break out. A few will. The entire business depends on identifying which ones will, fast enough to act on it.
Their two operators were doing this manually. Spreadsheets. Gut feel. Daily reviews of hundreds of creatives. It worked when the catalog was small. At their current scale, signals were getting missed.
What We Found
During discovery, three things became clear:
The data was split across two completely separate systems. TikTok has two APIs: the Marketing API (campaigns, ad spend, creative performance) and the Shop API (orders, products, inventory, affiliates). Different authentication flows. Different portals. Different data models. The operators were switching between two dashboards and three spreadsheets to get a single picture.
The biggest spend was invisible from the obvious API. GMV Max campaigns, which accounted for the bulk of that spend, live entirely in the Shop API. The Marketing API only sees Brand Consideration campaigns worth a small fraction of it. Building on the wrong API would mean optimizing 2% of the budget.
Nobody had time to ask the right questions. The ads operator knew which creatives were underperforming. The supply chain operator knew which products were moving. But connecting those two signals, seeing that a product's ad creative was fatiguing while its live sales were accelerating, required manual cross-referencing that nobody had time for.
What We Built in 10 Days
Two AI agent workspaces, one per operator role. Each workspace has its own command set, its own data scope, and its own Git branch. 29 commits, 15 Python scripts, 75 automated tests.
| Component | Details |
|---|---|
| Workspaces | 2 (Ad Optimization + Product Analysis) |
| Production skills | 6 across both workspaces |
| API connections | TikTok Marketing API + Shop API |
| Test coverage | 75 automated tests |
| Codebase | 15 Python scripts, 29 commits |
| Interactive dashboard | "The Hive" with drag-and-drop CSV upload |
The ads operator types /audit and gets a waste report across five hypotheses: placement leaks, creative fatigue, dayparting inefficiency, audience overlap, and budget misallocation. The supply chain operator types /hero and gets a ranked list of SKUs by breakout probability with specific reorder recommendations.
Every command saves output to a timestamped file. Every operator session auto-commits to Git. When someone asks "why did we kill that creative?" or "when did we flag that SKU?", the answer is in the commit history.
The Breakout Probability Engine
The most critical component is the scoring engine that classifies SKUs. The formula:
Breakout Probability = (0.30 × Content) + (0.40 × Live) + (0.30 × Affiliate)
Each factor pulls from different signals:
| Factor | Weight | Inputs |
|---|---|---|
| Content | 0.30 | Views, CTR, completion rate, add-to-cart rate |
| Live | 0.40 | Click rate, conversion rate, units/minute, GMV/minute |
| Affiliate | 0.30 | Active affiliates, converting affiliates, GMV velocity |
A momentum multiplier adjusts the score based on 24-hour GMV trends. Growing 30%+ in 24 hours gets a 1.2x multiplier. Declining gets 0.8x.
The classification:
- HERO (80+): Full scaling. Place bulk order. Expand affiliate coverage.
- WATCH (65-79): Maintain exposure. Monitor for breakout or decline.
- COLD (below 65): Creative retry or route to clearance.
This scoring runs in seconds. The same analysis used to take days of manual review across multiple spreadsheets. More importantly, it runs the same way every time. No signals missed because someone was busy with another task.
Five Architecture Decisions That Shaped Everything
1. Two APIs from day one. We mapped both the Marketing API and Shop API before writing any code. This saved us from building an entire system on the wrong data source. Anyone building TikTok automation: figure out where your money actually lives before you start.
2. One workspace per operator role. The ads operator never sees inventory commands. The supply chain operator never sees ad spend commands. This is not just about UX. It is a safety boundary. Each operator works on their own Git branch and cannot modify shared code or touch the other workspace.
3. Mock-to-live data layer. API approval from TikTok takes days to weeks. We built agents that work on calibrated mock data (built from real business metrics shared during discovery) and auto-switch to live data when credentials validate. The operators started learning the system on day one. When the APIs connected, nothing changed from their perspective.
4. Conversational agents over dashboards. With hundreds of active creatives, the right question matters more than the right chart. "Which creatives should I kill today?" produces a faster decision than a pivot table with hundreds of rows. The agents accept natural language and return ranked recommendations with specific IDs and estimated impact.
5. Git as audit trail. A launcher script handles branch switching, pulling latest changes, starting the agent, and auto-committing outputs on exit. One command to start a session. Full traceability of every recommendation the system ever made.
The Engagement
We structured this as a 12-week engagement: the first production build in 10 days, then deployment, training, and stabilization over three months, with the work (and the fee) declining as the system settles. After handover, the operators run the system themselves, with an optional retainer for continued development.
A fixed-fee build against six-figure monthly ad spend. The math works if the scoring engine catches even one hero SKU early enough to avoid a stockout.
What This Looks Like in Practice
The ads operator starts their day by typing one command. They get a 30-second snapshot: top performers, underperformers flagged for review, estimated waste from the previous day. They ask follow-up questions in plain language. The agent answers with data, not charts.
The supply chain operator does the same. Which SKUs are accelerating? Which ones need reorders placed today? What is the stockout risk on the current top seller?
Both operators work in isolated environments. Both sessions are tracked. Both produce audit trails. Neither needs to touch a spreadsheet, switch between dashboards, or manually cross-reference two data sources.
The system does not make decisions. It surfaces signals fast enough for operators to act on them. The operators are the experts. The agents make them faster.
If This Sounds Like Your Problem
If your team is managing high-volume TikTok operations, manually tracking creative performance across hundreds of ads, or trying to detect product trends before stockouts kill your momentum, the operational challenge is the same regardless of category.
We build these systems. We are happy to have a straight conversation about whether it makes sense for your setup.