Challenge

Bash, the multi-brand ecommerce platform of South Africa's largest fashion and lifestyle retail group TFG, faced a persistent conversion problem: shoppers were browsing extensively but not buying. Customers were viewing multiple products, showing clear purchase intent, yet dropping off without completing a transaction or signalling why.

Traditional personalisation tools like static banners, algorithmic product blocks, and generic recommendations were not reaching these hesitating shoppers effectively. The team at Bash wanted to create highly personalised experiences that felt helpful rather than intrusive, but existing approaches couldn't interpret the nuanced intent behind queries like "I need an outfit for a garden wedding that isn't too formal."

The scale of the challenge compounded the problem. During peak trading periods like Black Friday, thousands of high-intent shoppers were hesitating simultaneously. Live chat or contact centre agents could not operate at that speed or volume, and rules-based automation lacked the ability to interpret abstract, multi-step queries in real time. Meanwhile, fewer than 2% of site visitors were engaging with traditional chatbots, meaning conventional tools were never going to reach the shoppers who needed guidance most.

Strategy

In August 2024, Bash launched Loomi Conversational Agent, a semi-autonomous conversational shopping agent on bash.com.. Rather than deploying a conventional chatbot, Bash took a behaviour-driven approach: the agent monitors browsing sessions in real time and autonomously decides when and how to engage each shopper.

The agent targets customers who have viewed three or more product detail pages in a single session without progressing toward purchase, a signal of high intent combined with hesitation. When triggered, it initiates a natural language conversation, interprets abstract or multi-step queries, and surfaces relevant product recommendations from Bash's live catalogue.

The agent coordinates across multiple composable systems in production: a conversational engine for dialogue and behavioural triggers, a customer data engine drawing on over 15 years of ecommerce behavioural data, Bash's live product catalogue, and FAQ and content sources used to resolve hesitations around fit, sizing, and availability. No single system powers the agent alone. Its value comes from orchestrating across all of them simultaneously.

The implementation required no engineering resources from Bash beyond specifying test criteria. The agent was configured, not coded, demonstrating that agentic commerce can be deployed without replatforming or deep technical investment. Once a shopper is confidently progressing toward purchase, the agent steps back without further prompting, avoiding unnecessary interruption.

By November 2024, the agent was running autonomously during Bash's highest-traffic trading weekend, Black Friday, engaging a treatment group of eligible users in a rigorous A/B test with no human involvement in individual decisions.

Method

During the Black Friday A/B test, customers who interacted with the conversational shopping agent saw statistically significant lifts versus the control group: a 35.2% increase in online conversion rate, a 39.8% increase in revenue per visit, and a 28.1% reduction in exit rate. These gains were achieved on top of Bash's existing Black Friday performance, confirming the agent delivered incremental value rather than simply re-routing existing demand.

Operationally, the entire implementation required no engineering resources beyond specifying test criteria. This low-overhead model means Bash can expand and iterate the agent across new funnel moments like product listing pages, search pages, and earlier triggers, without engineering dependencies, compressing the time from idea to live experiment.

Building on these results, Bash is now expanding the agent further up the funnel and onto product listing pages, with pre-generated contextual questions surfaced under the "Add to Cart" button to address last-mile conversion friction at scale. Each expansion compounds on the last, with the data and confidence from each phase unlocking the next.

As Bash's Head of Growth Solutions noted, the agent exceeded expectations across the board. It learned quickly, performed better than expected, and was a standout success.

Why composable and agentic matters here

Bash's architecture around the agent is modular and composable. The agent taps into Bash's extensive product catalogue and multiple FAQ and content sources, orchestrated through a customer data engine and AI stack. Behavioural conditions, such as "viewed at least three product pages," are defined in configuration, not code, letting Bash change thresholds and triggers without re-implementing the agent.

The agent itself runs as a front-end experience layer tied into existing ecommerce and analytics systems, enabling fast experimentation and future expansion to product listing pages and potentially other channels. This separation of concerns means Bash can plug agentic capabilities in without replatforming, and evolve them as it learns. This is a direct expression of what composable, API-first infrastructure enables in practice.

The agent makes autonomous decisions in real time: when to engage, what to say, which products to recommend, and when to step back. It coordinates across conversation, commerce, and customer data layers to produce each response. No human reviews or approves individual conversations. Bash's team configures targeting conditions and success criteria; the underlying model and data connections are tuned separately. This vendor-configured, brand-directed model has produced a reusable pattern for agentic commerce deployment that requires no replatforming and no deep engineering investment from the brand side.