E-commerce Multi-Agent System: AOV Up 22%, Abandoned Carts Down 34%
Cart recovery, personalisation and upsells are usually three disconnected tools sending three uncoordinated messages. Running them as agents over one behavioural model is what changed the numbers.
- Client
- High-traffic online retailer
- Industry
- E-commerce
+22%
Average order value
-34%
Abandoned carts
Multi-agent
Personalisation, recovery, upsell
Brand-safe
Guardrailed generation
The problem
The storefront ran the standard stack: one tool for abandoned-cart email, another for on-site recommendations, a third for post-purchase follow-up. Each worked in isolation and none knew what the others had done, so a customer could receive a recovery email for a product they had already bought and an upsell for an accessory that did not fit it.
Average order value was flat and cart abandonment was where it usually is, and the team had run out of ideas that were not just more discounting.
The architecture
We built a multi-agent system over a shared behavioural model rather than three campaign tools. Session behaviour, purchase history and catalogue relationships feed one representation of the customer, and each agent reads from it and writes back to it.
The personalisation agent handles on-site recommendation, the recovery agent handles abandonment, and the upsell agent handles cart-stage and post-purchase suggestions — but because they share state, they coordinate. A recovery message knows what the personalisation agent showed, and the upsell agent knows what the customer just declined.
Predictive scoring decides who is worth contacting and when. A discount is not the first lever; timing and relevance are, and the model reserves margin for the cases where nothing else will convert.
Brand safety
Generated commercial messaging that sounds slightly wrong costs more than it earns. Every outbound message goes through structured prompting with explicit guardrails on tone, claims and product facts, and product statements are constrained to catalogue data rather than to whatever the model believes about the item.
The system is not permitted to invent a benefit, a compatibility or a delivery promise. It composes from the record or it says less.
Results
Average order value rose 22% and abandoned carts fell 34%. The two are related: better cart-stage relevance both recovers more carts and produces larger ones, which a discount-led approach does not do.
The system runs inside the storefront and CRM through API workflows rather than as a bolt-on, and it was built for the site's traffic profile — the peaks are when it matters and the peaks are when naive implementations fall over.
What we'd flag
This depended on a clean catalogue with real product relationships. Where the catalogue data was thin, the agents had nothing intelligent to say and defaulted to generic behaviour, which is exactly the thing we were replacing.
We also held back from full automation on the discount decision. The agents propose and the merchandising team sets the bounds, because a model optimising conversion without a margin constraint will find the discount every time.
