KINTO fixed the weak link of an e-commerce brand that sells over WhatsApp with a multi-model audited agent and order-by-order reconciliation; the owner reported a month with channel sales ≈30% above their history.
Problem
A brand with complex variants (model × size × color × stock) sells on the web and mostly over WhatsApp. In the owner’s words, order automation was very good but chat was the weak point: slow replies, wrong sizes, out-of-stock items sold and orders that never reached the store. And nobody knew how much of what the ad platforms claimed was true.
Signals
- Market: Recurring demand for flagship models and unmet demand for some lines (sold-out sizes and colors), visible only by reading conversations.
- Business: Over a third of orders closed in chat, mostly cash on delivery; a “drop” that worried the team was just a return to normal after a one-off spike.
- User experience: Customers ask about sizing, photos by color, financing, payment links and delivery times that change by city and holiday.
- Support: One in eight conversations asked for a human, and the staff corrected the bot by hand.
How we understood it
- Business audit over hundreds of conversations: AI demand extraction, quality judged by a different model, and deterministic computation. Core finding: the bottleneck was not the agent’s conversion but that most conversations left no data behind.
- Multi-model quality audit: the automated judge overstated, so every finding was verified by hand. Specific lost sales from wrong sizes were found.
- Audit of team feedback: months of chat with the client turned into a list of commitments; the costliest debts were our own unfulfilled promises, not bugs.
- Instrumented attribution: real ad API consumption was measured (to operate without exceeding its limits) and each order was matched to its analytics session.
What we built and why
- WhatsApp sales agent that advises by size, sends photos by color, quotes shipping to any municipality, creates the order in the store and notifies the team. Live stock every few minutes with an alarm if it fails, a curated spec sheet (“no spec” rather than invent one), read-only payment checks and a per-chat pause so it coexists with human staff.
- Intelligence tab for the sales team: best sellers, most asked-about items, true stock-outs vs. bot errors, and suggested restocking.
- Ad governor: a quota control that keeps requests within each platform’s limits, with an audit trail and changes only under a plan a person approved; real return per channel (Meta and Google against the same cash) and automatic source tagging of each order.
- Store-wide SEO: unique metadata and descriptions, alt text for hundreds of images, and content for uncovered high-intent searches.
Why this way: The automated judge overstated: only one in five “high” alerts was a real error, so no finding was accepted without human verification. And the budget now moves on cash data, not each platform’s attribution.
Transformation
-
In its first weeks the agent closed about two dozen orders at close to 10% conversion. At the end of the first full month the owner reported channel sales about 30% above their history (the client’s statement, not an independent measurement).
-
Reconciliation matched 100% of the window’s orders to their real source and showed one ad channel returning about twice the other: budget now moves on cash data.
-
Quality moved from “bugs the team reports” to a multi-model audit cycle, with half of the feedback closed with evidence.
-
orders reconciled with their real source: 100% of the window (hundreds of orders) (measured)
-
return per channel measured against the same cash: one channel returned ≈2× the other (measured)
-
agent conversion in its first weeks: ≈9–10% of conversations (measured)
-
channel sales in the first full month: ≈+30% vs. history (client-reported)
What we don’t know yet
- Ad-spend savings in absolute terms, and the agent’s effect isolated from seasonality.
- Still open: unifying product codes between the ERP and the store, and notifying customers when a sold-out item is restocked.
