The problem
The brand was running 5 TikTok Lives a week, each pulling 18,000–32,000 concurrent viewers. On-stream conversion was solid — about 290 orders per stream — but TikTok Shop's analytics showed another 4,200+ users adding items to cart and never completing checkout. Most never came back. The post-Live email sequence was the same generic 10%-off blast for everyone, and only 8% of recipients opened it.
ROAS on TikTok Ads dropped from 4.1 in Q4 2025 to 2.6 by March 2026 as creative fatigue set in and competing brands flooded the FYP. The growth team was hired to scale, not to fix conversion — but the leak was so big that any new ad euro mostly poured into the bucket without filling it.
Customer service was drowning in DMs from TikTok ('e disponibil în M?', 'când ajunge la Cluj?', 'se potrivește pentru 1.65m?'). Three part-time agents tried to keep up across TikTok DMs, Instagram DMs, WhatsApp, and email — most messages got an answer the next day, well after the customer had already bought from H&M or Zara instead.
What we built
TikTok Shop in CEE finally hit critical mass in early 2026 after Romania, Hungary, and Czechia got native checkout. The unlock for D2C brands wasn't another ad — it was capturing the demand that TikTok was already creating for free but not monetizing. We built an AI agent that sits on top of TikTok Shop, the brand's Shopify store, and WhatsApp Business, and behaves like a senior sales associate who knows the entire catalog cold and never logs off.
The agent owns three jobs the brand could never staff for: live Q&A during streams, sub-10-minute cart recovery on WhatsApp, and a continuous post-purchase relationship that turns one-time TikTok shoppers into repeat customers.
- Unified TikTok Shop ↔ Shopify ↔ WhatsApp inventory and order layer — no more 'sorry, sold out' apologies after the order is placed
- AI styling agent on WhatsApp: when a customer abandons cart on TikTok, an automatic message with the exact product, a size suggestion based on past orders or a 4-question fit quiz, and 2 styling alternatives — within 8 minutes
- Live-aware product Q&A: during TikTok Lives, the agent answers comments in real-time about size, fabric, fit, and shipping ETA — and pushes a personal 'save this look' WhatsApp link to anyone asking a sizing question
- Smart sizing predictor combining the customer's past orders + the fit quiz — return rate dropped 37% on size-related returns alone
- Influencer-aware checkout: orders coming from a specific creator's Live get attributed automatically and trigger a thank-you DM 24h later that drives repeat purchase
- Romanian-first language layer trained on the brand's voice — no robotic 'dear customer', warm casual copy that matches the founder's TikTok persona
- Predictive replenishment: customers who bought basics 90 days ago get a soft WhatsApp nudge with a one-tap reorder of the same SKU plus a new-arrival recommendation
- Live dashboard with cart-recovery rate per Live, per product, per creator — so the merchandising team finally knows which Reels and Lives to put real ad spend behind
The results after 8 weeks
Recovered cart revenue hit €480k in Q1 alone — a category that was effectively zero before. TikTok Live conversion jumped from 1.4% to 8.7% as the agent absorbed Q&A and pushed personalized WhatsApp links during the stream. Customer-service ticket volume dropped 64% even though monthly orders grew 38%.
The biggest surprise was the repeat-purchase loop. The smart replenishment WhatsApp flow brought back 1,890 customers who hadn't ordered in 90+ days — at a CAC of essentially zero. Average order value climbed €18 because the agent suggests a styling pair every time a customer checks out.
Returns dropped from 22% to 14% on apparel, a number the operations team had been chasing for two years. Better sizing recommendations alone unlocked enough margin to pay for the entire stack inside the first month.
What we'd do differently
The first deployment treated WhatsApp messaging as transactional. After a week of A/B-testing voice and tone, we shifted to a 'stylist-friend' persona — slightly informal, emoji-light, occasionally humorous. Open rates went from 41% to 78% and the same campaign suddenly produced 2.3x more recovered revenue.
We also learned to throttle. The agent originally pinged any cart abandoned for 30+ minutes; customers told us it felt pushy. We moved to a 4-touch sequence (8 min, 2h, 24h, 72h) with each message progressively more personalized — recovery actually improved, complaints dropped to zero.
"We thought we needed to spend more on ads. Turns out we just needed to actually answer the people who already wanted to buy."
— Co-founder & creative director
