Support agents wired into your order system, product search that understands intent, recommendations grounded in your catalogue and forecasts your buyers trust, for online retailers, marketplaces and omnichannel brands.
Retail AI pays off when it is connected to the systems that hold orders, stock and products. Disconnected chatbots cost money and goodwill.
Support agents that resolve, not deflect
Where is my order, change my address, start a return, apply a missed discount. The agent reads and acts on the order management system and your returns policy, and escalates with a full transcript when it should. Touches the OMS, helpdesk and messaging channels.
Semantic and hybrid product search
Search that understands "something like this but cheaper" and tolerates typos and synonyms, with your merchandising boosts and stock rules kept intact. Touches the search index and the catalogue feed. See semantic search.
Recommendations grounded in your catalogue
Complementary items, size and fit guidance and "complete the look" built from product attributes and behaviour, not a black box. Touches the catalogue, the analytics warehouse and the storefront. See recommendation engine.
Catalogue enrichment at scale
Generate consistent titles, attributes and descriptions from supplier data and images, flag gaps and keep a human approving each batch. Touches the PIM and supplier feeds. Thousands of SKUs become searchable and comparable.
Demand forecasting and replenishment
SKU-level forecasts that account for promotions, seasonality and stock-outs, feeding the replenishment plan. Touches the ERP or planning tool and the warehouse system. See predictive analytics.
Assistants inside the mobile app
Visual search from a photo, loyalty and offer explanations, store stock lookups and reorder prompts, built natively into iOS and Android apps. Touches the loyalty platform and store inventory. See AI for mobile apps.
Compliance and data
Customer data, payment data and fair claims
What we design for in retail
Customer records and behavioural data fall under GDPR and UK GDPR, so consent, purpose limitation and the right to erasure apply to anything an AI stores or learns from. Card data belongs in your payment provider, never in prompts or logs. Consumer protection rules on pricing, availability and product claims mean generated content needs an approval step and a source of truth. Marketing automation must respect channel consent. This is engineering guidance, not legal advice.
We keep personal data out of model training, scope what the support agent may disclose to a verified customer, and log every generated description against the product data it came from. Our prompt injection defence work matters here too: product reviews and supplier text are untrusted input.
Reference architecture
Flagship workflow: a support agent connected to your commerce stack
The customer writes from the app, the website or a messaging channel. The agent authenticates them, then works through typed tools rather than free-form access: it can look up an order, cancel one that has not shipped, or open a return, each with rules about what needs confirmation.
Policy questions are answered from a knowledge base of your returns, delivery and warranty terms, with the passage cited. Product questions draw on the catalogue and the recommendation service. When the customer is upset, the issue is outside the agent's scope or a refund exceeds a threshold, it hands the conversation to a person in the helpdesk with the transcript and the actions already taken.
Every tool call is logged, so you can audit what the agent did and tune the boundaries. The pattern is described in detail on the MCP server development page.
The agent is only as useful as the tools it can call; the boundaries on those tools are the real design work.
Track record
What we have built that applies
Retail and marketplace mobile apps
Catalogue browsing, baskets, payments, loyalty and push campaigns in native iOS and Android apps, part of the 1500+ apps we have delivered since 2006.
Our event platform sells tickets with tiered pricing, coupons, add-ons and refunds, with the checkout, fraud and reconciliation lessons that come with running it.
Can an AI agent really change an order, or does it just answer questions?
It can act, within limits you set. We connect the agent to your order management system through typed tools (look up order, cancel if unshipped, start a return) and define which actions need confirmation or a human. Everything it does is logged against the conversation. See AI agents and automation.
How is AI product search different from the search we already have?
Keyword search matches words; semantic search matches meaning, so "warm jacket for cycling in the rain" finds the right products even when none of those words appear in the title. We usually combine both as hybrid search and keep your merchandising rules on top. See hybrid search.
Do we need a data science team for demand forecasting?
No. You need clean sales history, a product hierarchy and someone who owns replenishment decisions. We build the model, wire it into your planning tool and hand over monitoring that tells you when it drifts.
Which platforms do you integrate with?
Headless and hosted commerce platforms, order management, PIM, ERP and helpdesk systems via their APIs, plus Salesforce Commerce and Service Cloud where that is your stack. Custom back-ends built in Node.js, PHP/Laravel or Python are our bread and butter.
Ready to make AI work in production?
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.