Retail's AI Rebuild: Why Point Solutions Are Losing to Orchestration

Talk to enough retail and e-commerce operators and the same pattern surfaces. First, customer experience breaks under its own growth: the more successful a brand gets, the longer its queues get, and the more customers quietly churn rather than wait for an answer. Second, personalisation isn't actually personal: most 'AI-powered recommendation' engines are still running rules tuned years ago, and customers can tell the difference between a system that knows them and one that's guessing. Third, operations run on duct tape: inventory forecasting sits in one tool, support in another, marketing automation in a third, none of them talking to each other, all of them creating blind spots at the handoffs. Fourth, the data to fix this already exists but isn't usable: retailers sit on enormous transaction and behavioural datasets that live in silos no single team can act on in real time. Fifth, pilot fatigue has set in: nearly every retailer has tried AI somewhere; far fewer have gotten it past a demo and into production workflows that survive Black Friday-level load.

The shift underway is not 'more chatbots,' it's multi-agent systems replacing single-task automation end to end. Instead of one bot answering one question, an orchestrator now coordinates a support agent, an inventory agent, a fulfilment agent, and a logistics-notification agent against a single customer event. A return request doesn't just get acknowledged; it triggers a refund, updates on-hand inventory, and notifies logistics, without a human touching any step unless the workflow crosses a defined risk threshold. This is the same architectural pattern reshaping enterprise BPM more broadly: a DAG-based orchestrator with specialist agents, shared context, and human-in-the-loop gates on the decisions that actually need one. In retail, the highest-leverage places to apply it are returns and refunds, order-exception handling, and post-purchase service: the workflows that today burn the most headcount on repetitive coordination rather than judgment.

Generative AI in retail started as a content tool: product descriptions, ad copy, email subject lines, and that's still real, but it's no longer where the growth is. The faster-moving use cases now sit inside operations: merchandising copy generated and localised at catalogue scale, demand forecasting models that generate synthetic scenarios to stress-test inventory plans against demand spikes they haven't seen in historical data, and generative design support for private-label and seasonal assortments. Retailers report generative AI spend within the broader AI-in-retail category scaling faster than the category average, roughly doubling over a couple of years, precisely because it's being redeployed from marketing into forecasting, merchandising, and product content generation at a pace point-solution vendors weren't built to support.

Batch personalisation, the nightly job that recomputes 'customers who bought X also bought Y,' assumes intent doesn't change between runs. It does. A shopper who searched for a competitor's product an hour ago, abandoned a cart at checkout, or just switched from browsing formalwear to browsing luggage has told you something a batch job won't see until tomorrow. Real-time personalisation systems act on that signal at the moment of intent: session-level context feeds a recommendation and pricing layer that updates within the same browsing session, not overnight. This requires infrastructure most retailers don't yet have: a low-latency feature store, streaming event ingestion from web, app, and POS simultaneously, and a serving layer that can make a ranking decision in under 100ms without falling back to a stale cache. It's also where the gap between 'AI-powered' marketing and something a customer actually notices is widest, and where the ROI is easiest to measure, because the counterfactual (a static recommendation) is right there to compare against.

Per Fortune Business Insights' AI-in-retail research, the solutions segment (smart-store technology, e-commerce personalisation, and intelligent supply chain) now accounts for roughly two-thirds of all AI-in-retail spend, ahead of services. That's a meaningful signal: retailers are no longer paying primarily for consulting and integration work, they're buying software that runs the operation directly. Adoption has also crossed a tipping point: the large majority of retail and CPG companies report actively using or piloting AI, and the share in active deployment, as opposed to testing, has grown sharply year over year. The trajectory doesn't read as a market approaching a plateau; the CAGR holds through the 2034 projection, which means the fragmentation problem retailers have today is being built on top of, not solved by, most current spend.

This is the uncomfortable part. A rising AI-in-retail market doesn't mean any given retailer is capturing value from it: most of that spend is still going toward point solutions that solve one problem while leaving the underlying operational fragmentation untouched. A best-in-class forecasting tool bolted onto a support stack that can't see inventory in real time still produces stockouts the forecasting tool predicted correctly. A generative merchandising tool that can't hand its output to a workflow that updates the PIM and the storefront simultaneously just creates another manual copy-paste step. The retailers actually compounding value from this market are treating AI as infrastructure, a layer that every other system reads from and writes to, rather than a growing list of standalone tools that each solve their slice and leave the seams for someone else to reconcile.

In practice, closing that gap requires five components working together rather than in isolation. Intelligent customer experience: chat, voice, and async channels that resolve issues end-to-end instead of routing to a human for anything non-trivial. Agentic workflow orchestration: the coordination layer connecting support, inventory, fulfilment, and marketing so one customer action triggers the correct downstream chain automatically, with audit logging on every step. Enterprise automation with governance: replacing manual reconciliation between systems with workflows that are auditable and scale with order volume instead of degrading under it. Persistent memory and knowledge: systems that retain context about a customer across sessions and channels, so personalisation means an actual memory of prior interactions rather than a first-name merge field. And a deployment discipline built for production, not demos, because the retailers stuck in pilot purgatory usually aren't blocked on model quality; they're blocked on the unglamorous work of integration, monitoring, and rollback that never gets funded as its own line item. Retailers that build these five as one connected layer, rather than five separate vendor relationships, are the ones positioned to compound with the market instead of just contributing to its topline number.

Found this useful?

Share the key finding on X.