AI Automation of Order Intake and Customer Support for a Kenyan FMCG Distributor

Context

Our client distributes household and personal care products to several hundred small retail outlets, kiosks, and mini supermarkets across Nairobi and its surrounding satellite towns, operating a model built on daily and weekly reorders rather than large infrequent purchases. Retailers typically placed orders by calling a sales representative directly or sending a WhatsApp message with a rough list of items, which a small internal team then manually re entered into the company's order system before dispatch could be scheduled from the Nairobi warehouse.

The business had grown steadily, but its order intake process had not kept pace, still relying on the same handful of staff fielding calls and WhatsApp messages that had worked when the retailer base was a fraction of its current size.

The Problem

The manual order intake process was the clearest bottleneck in the business. During peak ordering windows, typically early mornings before shops opened and again in the late afternoon, the order desk was overwhelmed, with retailers waiting on hold or getting delayed WhatsApp replies. Orders taken over the phone were frequently miscaptured, a wrong quantity or a missed item, which then surfaced as a delivery dispute later that day, consuming further staff time to resolve and occasionally damaging trust with retailers who depended on accurate stock for their own daily sales.

There was also no structured way to handle common, repetitive customer queries, questions about current pricing, whether a particular SKU was in stock, or where a delivery currently was in the Nairobi traffic. These all funnelled to the same small team already stretched thin on order capture, meaning genuinely urgent issues sometimes waited behind routine questions that could have been answered automatically.

Staff turnover on the order desk compounded the problem further, since every new hire needed weeks to learn the full product catalogue, pricing tiers, and common retailer shorthand for products before they could reliably take orders without errors.

What We Did

1. Mapped the full order journey and catalogued common query types. We reviewed weeks of WhatsApp and call logs to understand how retailers actually phrased orders, including local shorthand and abbreviations for specific products, and grouped support queries into categories by frequency and complexity.

2. Built an AI powered WhatsApp ordering assistant. Retailers could send their order in natural language, including the informal phrasing they already used, and the assistant parsed it against the live product catalogue, confirmed quantities and current pricing, flagged any ambiguous items back to the retailer for quick clarification, and only escalated to a human when confidence was low.

3. Connected the assistant directly to inventory and dispatch systems. Confirmed orders were pushed automatically into the order management system without manual re entry, which removed the double handling that had been the main source of quantity and item errors.

4. Automated responses to routine customer queries. Pricing, stock availability, and standard delivery timeframes were handled instantly by the AI system, with the sales team only stepping in for disputes, credit terms discussions, and genuinely new or unusual requests.

5. Built a delivery status layer tied to the dispatch schedule. Retailers could ask the WhatsApp assistant where their order stood and receive an accurate update based on the actual dispatch queue and estimated Nairobi route timing, reducing the volume of "where is my order" calls reaching staff directly.

6. Trained the internal team to supervise rather than manually process. We shifted the order desk team's role toward reviewing flagged exceptions and handling escalations, and ran a two week parallel run where both the old and new process operated together so staff could build confidence in the system before it took over full order intake.

Results

Within three months, 68 percent of total order volume was captured automatically through the AI assistant without any human re entry, up from a baseline where effectively all orders required manual handling. Average time from order placement to confirmed dispatch scheduling improved 3.1x, since orders no longer waited in a queue behind staff availability during peak hours. Order accuracy improved noticeably as automatic parsing against the live catalogue removed a large share of the misheard or mistyped quantity errors common under the phone based process. Support cost per order fell 39 percent as the team's time was redirected from routine data entry and repetitive queries toward exception handling and relationship management with larger retail accounts.

What We Would Do Differently

We would have spent more time upfront cataloguing local shorthand and abbreviations retailers used for specific products, since early on the assistant misread a meaningful share of orders that used informal names rather than official product titles, and expanding this vocabulary earlier would have raised the automatic capture rate faster. We also initially set the escalation confidence threshold too high, meaning too many orders were routed to a human unnecessarily in the first few weeks, and we had to recalibrate it downward once we saw how often the assistant's own confirmed reading matched what retailers actually meant. Finally, we would introduce the delivery status feature earlier in the rollout rather than in month two, since it turned out to be one of the highest volume query types and its absence in the early weeks meant staff were still fielding a large share of routine status calls that the system could have handled from day one.