Published
From forecast-driven commerce to intent-driven commerce
Commerce has long operated on one assumption: demand remains stable long enough to plan against. For decades, that assumption underwrote the whole operating model: collect historical sales data, run the statistical forecast, build the sales and operations planning (S&OP) strategy, and allocate inventory against it. Today, that assumption is increasingly under pressure.
Demand increasingly does not arrive in predictable weekly buckets. A product goes viral. An influencer posts a haul. A heatwave hits. A competitor runs out of stock. A marketplace livestream sells through in 20 minutes. A planning system that refreshes once a week may struggle to register these shifts before the opportunity has passed. This can create latency between when demand signals emerge and when supply chain teams can respond.
This is the demand-to-delivery gap. Missed conversions, stranded inventory, broken delivery timelines, and margin lost to markdowns are the four primary symptoms, driven by one root cause: planning cycles measured in days or weeks are increasingly misaligned with customer decisions that happen in seconds.
Addressing it may require more than better ecommerce demand forecasting. In some environments, it can involve a different architecture that uses predictive models and AI-enabled agents to help detect demand shifts and trigger supply-side actions. Passive analytics explains what happened. Closed-loop execution helps influence what happens next.
Why traditional demand planning struggles to address a faster, compressed demand cycle
Organizations have planned in periods based on historical patterns, while customer behavior has evolved continuously. This mismatch results in three major consequences:
Inventory sits where demand was forecasted, not where it is emerging now
Fulfillment is optimized for cost per shipment, not for responding to demand shifts in real time
Delivery promises are quoted from a batch available-to-promise (ATP) snapshot, not the live inventory position
Historical data alone does not fully address this challenge. Recognizing demand sooner – and acting on it before the experience degrades – is the challenge to solve for.
The shift: From prediction to orchestration
Agentic AI can help organizations make operating decisions faster and with greater confidence. A planner may no longer need to identify every pattern manually before action is considered. This is where AI in demand forecasting helps organizations move from passive reporting to decision support and automated recommendations within defined workflows.
Specialized agents can be configured to analyze search and browse behavior, marketplace and social signals, weather and local events, and live inventory and fulfillment capacity, running continuously and feeding a shared coordination layer. This reconciles their conflicting recommendations (a demand spike says expedite; a cost model says consolidate) before committing to action.
We see this shift built on four transformative moves.
1. Sensing the full commerce signal
AI agents can ingest real-time telemetry such as clickstream events, transaction data, review sentiment, permitted marketplace signals, pricing inputs, weather data, and local-event feeds through an event-driven data architecture. The result is a continuously updated view of customer intent across categories and local markets, feeding directly into the orchestration layer.
2. Orchestrating supply the moment intent emerges
Sensing demand is only the first step. Value is created when organizations can act on those signals quickly and effectively – recalculating safety-stock thresholds against the live signal instead of a static buffer, rerunning ATP and capable to promise (CTP) against current inventory and capacity rather than a historical snapshot, and rebalancing stock across nodes.
Case in point: Autonomous supply chain planning at a global ecommerce marketplace
Genpact and the client built an agentic customer hub that senses demand and supply shifts in real time, layering a continuous feedback loop over the existing advanced planning system (APS) by recalibrating agent policy against realized outcomes rather than a static model. This automates selected planning decisions within defined parameters, helping organizations respond faster to changing demand while allowing planners to focus on exceptions that require human judgment. The program is designed to increase automation significantly over time, with the goal of reducing manual intervention by mid-2027. Based on internal analysis, the program is projected to reduce manual planning and exception-handling hours by more than 40%, with an estimated annualized profit-and-loss (P&L) impact of between $75 million and $82 million on a $13 billion revenue base. Projected value drivers include freight and logistics efficiencies, reduced stockout-related lost sales, and lower markdown exposure from better-positioned inventory.
Case in point: Furniture retail supply chain modernization
A large furniture retailer developed a control-tower analytics capability during the pandemic backlog; it has since evolved into a closed-loop, agentic system that continuously senses purchase orders (POs), supplier output, and carrier capacity. On-time-in-full (OTIF) rose from roughly 45% to 55% – still below the 90% or more OTIF that fast-moving consumer goods (FMCG) retailers with simpler, single-box orders can hit – but a meaningful gain for a category where split shipments and made-to-order lead times structurally cap the ceiling. Split deliveries fell from 58% to 42%, and supplier-carrier balancing now occurs continuously rather than as a weekly planning exercise.
3. Activating the seller and merchant ecosystem in real time
In a marketplace, demand only converts into revenue if sellers can respond to it. Slow onboarding, incomplete catalogs, and reactive support suppress a new seller's search visibility and lengthen time to first sale. Our assessment puts the resulting loss ranging between $15,000 and $50,000 in gross merchandise value (GMV) per seller over a typical activation-delay window. The size of this loss varies according to catalog complexity and category.
Case in point: Genpact's NxtSellAI and SellAssistAI
Each agent serves a distinct purpose: NxtSellAI helps sellers get up and running faster by identifying onboarding issues and listing gaps. SellAssistAI helps monitor seller operations after go-live, highlighting stock and pricing exceptions that may require attention. When deployed together across marketplace environments, the solution has been associated with a 40% to 60% reduction in onboarding time, helping sellers become operational sooner. Based on observed deployments, organizations have also seen a 20% to 35% reduction in stockout events, supporting faster replenishment and improved product availability. More complete catalogs and consistent inventory availability have been associated with 45% to 50% higher seller lead conversion rates and increased GMV. In some cases, sellers affected by onboarding delays achieved an uplift of up to 100% after activation barriers were addressed.
4. Enabling experts to be at the center of high-stakes decisions
Autonomy without boundaries is a liability waiting for an audit. The goal is not to remove experts from decision-making. It is to help them focus on the decisions where experience and judgment create the most value. The durable model runs on three tiers of trust.
Tier one: Agents act independently on decisions within predefined operating boundaries of SKU-level replenishment and reallocation within standard safety-stock bands, where thousands of micro-decisions happen faster than any expert could review them individually
Tier two: Agents recommend – and an expert approves – cross-node transfers above a set threshold, promotional pricing changes, and allocation calls that touch brand-critical SKUs. These are routed to an expert's queue with the agent's confidence score and reasoning attached
Tier three stays expert-only: Assortment cuts, vendor relationship changes, and network design decisions like opening or closing a distribution center. Agents surface the signal and the recommendation; the call belongs to an expert
Every autonomous action is logged with what triggered it, what the agent's confidence score was, and what it did, so the override rate becomes a governance metric a category leader can validate – not a black box they have to trust. Done well, this governance model helps organizations scale autonomy responsibly while maintaining operational control.
The next era of commerce will be won by whoever turns live signals into a fulfillment action the fastest. Autonomous supply chain planning powered by agent-based AI makes this possible, shifting the operating model from forecast-driven to intent-driven and closing the demand-to-delivery gap before the customer notices. In a market where expectations move faster than planning cycles ever could, sensing and acting in real time is the price of entry.
