Shashank .

Domain: Supply Chain – Can AI Be Creative or Does It Just Remix Human Ideas? In logistics and fulfilment, the AI can generate ideas on its own, but those ideas tend to be generic and disconnected from how the real operations actually behave. If you ask an AI to design a peak-season fulfilment strategy or an improved picking flow(within a warehouse) without giving it any process context, it will suggest things like dynamic batching or congestion-aware routing. While these look creative, they are just pattern recombination from human suggested ideas, underlying textbook principles or best-practice libraries. They don’t account for operating constraints such as aisle widths, dock congestion, carrier cut-offs times, SKU velocity, WMS limitations, or resource roster. Without these constraints, AI’s creativity becomes theoretical — impressive to read, but impractical for implementation. Suggestions from AI become useful only when a human provides context around operating parameters, historical performance, certain business rules and the expected outcome. Suppose, if you tell it something specific — ‘the cross dock vehicle will dispatch at 5 PM regardless of the fill rate’ or ‘Aisles G–J are one way aisles and become a congestion point every afternoon because two reach trucks can’t pass there’ or “the pick modules slow down 20% whenever fast movers are out of slot’ — the AI’s output changes completely. It stops producing generic playbooks and starts generating site-specific, operationally creative and implementable options that would never emerge from pattern remixing alone. You can see the difference in the ideas generated by the AI in both absence and presence of operating context: Example 1: Picking Strategy Without detailed context: AI proposes “switch to wave-less continuous release” — a trendy but risky generic textbook tactic. With detailed context: AI designs a hybrid model with pre-waves for bulky items, micro-waves for SKUs in congested aisles, and continuous release only for fast movers. It also introduces a replenishment buffer for Aisles G–J during peak to avoid reach-truck blockages. Example 2: Transport Planning Without detailed context: AI recommends consolidating all northbound loads during off-peak hours. With detailed context: AI learns the 3PL’s cross-dock refuses late arrives after 5 PM and hence designs a split-transport strategy where high-velocity SKUs ride on an early “feeder” truck, and low-priority loads move later. It also adjusts dock sequencing to keep shippers synchronized with carrier cut-offs. In each case, the AI’s response is initially curated based on patterns learnt from human provided data but then it provides suggestions that are operations friendly only when a human provides the real operational constraints around warehouse and network (which the AI would never infer alone). In short: AI without user process knowledge generates: Generic optimization ideas Strategies that ignore local bottlenecks Solutions that may be unrealistic for execution AI with user process knowledge generates: Realistic fulfilment flows Adjustments tied to carrier cut-offs, workforce capacity & layout Strategies that are realistic and will work during peak operations.

Domain: Regional Distribution Network – AI-to-AI Coordination Between Warehouses, Carriers, and Retailer Replenishment Systems One of the plausible examples is a regional FMCG distribution network which requires AI coordination between the warehouse orchestration AI and transport routing AI with only the exceptional decisions being routed through the Logistics and Fulfilment coordinator. Imagine a situation where the warehouse’s process orchestration AI notices that an outbound pick wave for a major retailer is slipping by 2 hours because a high-velocity SKU is being relocated between aisles. Before any supervisor even looks at the dashboard, the warehouse AI pushes a signal to the contracted carrier’s routing AI, which in turn recalculates the mid mile and realigns a trailer assignment to book a slightly later cross-dock slot at the regional hub. At the same time, the retailer’s replenishment AI gets an update and automatically adjusts shelf replenishment cycles and store allocation to avoid stock-outs. When these AIs sync up without human lag, the whole supply chain network becomes smoother, i.e. - Fewer missed delivery windows - More stable on-shelf availability, and - Efficient supply chain operations. Though this close coordination brings efficiency to operations, there is a good chance of over information sharing, An example of this being that a warehouse operator might accidentally reveal its labor-cost structure or inventory turns —things a carrier or retailer should never see. On the other hand, a transport provider can’t risk exposing its margin model or or fleet utilization. Also, we would not want the AI models to have conflicting objectives which would mean seeking impossible service levels that only get discovered at a later stage when a forklift operator, or store manager highlight an inconsistency and then the issue gets routed to the Logistics and Fulfilment coordinator. The solution could be a set of guardrails that everyone understands such as: Share short-term operational risk signals: Pick-wave or loading delay that affects departure in the next 2–4 hours Updated ETA windows with confidence levels and top-line reasons Keep strategic and sensitive data completely sealed: Labor models, productivity metrics, wage structures etc Transport margins, subcontractor dependencies, fleet utilization Personal or employee-level information With this simple structure in place, AI-to-AI coordination becomes a stabilizer rather than a liability. Warehouses run fewer overtime spikes, carriers avoid costly missed slots, and retailers get far more predictable inventory flow — all without exposing the data no partner should ever be asked to share. In short, the model will only work if we stick to one simple principle: share what is necessary and impacts the next few hours of movement while keeping all the strategic, personal, and cost-sensitive information locked down.

In a logistics and fulfillment setting, specifically the warehousing domain, the senior management often take decisions that are strategic or to tactical in nature. One such example is AI enabled workforce planning and task sequencing during peak hours. Here, the AI can add real value by projecting how different staffing and automated route selection (asset-based task assignment) will affect throughput, travel time or dock congestion within minutes. But it can also mislead if decisions are taken at face value. For example, the model might recommend redeploying pickers based on past spikes, while supervisors know that several aisles are running low on replenishment and such a move would actually create downstream delays the AI cannot foresee. For AI to be leveraged in an advisory role, the warehouse manager must actively blend the model’s logic with on-floor reality. For instance, after the AI proposes a staffing shift, a manager can walk through the simulation with the floor supervisor, check whether the assumptions match current stock levels, and adjust the plan to avoid new process bottlenecks and congestion points. They may even use the AI to run a second scenario that incorporates frontline feedback, comparing outcomes to further fine tune the model. This practice of reviewing assumptions, validating with human insights and refining the scenario rather than executing it blindly ensures that AI supports decision-making without overriding the contextual judgment and operational experience that the senior leadership depends on.

The AI system is built on assumptions, architecture and data patterns which might eventually become outdated as the industry or business grows. In certain cases, the fundamental changes (process, business goals, hardware or regulations) in the working environment create a void as the AI’s logic does not match with the reality or intended use. Hence, even with continuous learning capability, the system would still retire or be replaced as it cannot overcome the obsolete model design, structural changes in business goals or incompatibility with new technology. An example of this will be the retiring of legacy slotting optimization AI when the warehouse transitions from manual picking to autonomous mobile robot (AMR) operations. The legacy system used to recommend storage locations for SKUs in a warehouse, based on order frequency, item velocity and was designed for human pickers operations in the warehouse. Now the AMRs introduce entirely new constraints around robot turning radii, safety zones, charging oaths and robot specific aisle rules which need to be incorporated with the current asset entitlements to churn out optimal task sequence involving AMRs and human labor. But the legacy AI system cannot incorporate the robot specific physical parameters even with continuous learning capability. To operate the AMRs efficiently a new model built on robotic kinematics and real-time fleet coordination must replace the old one. To manage the transition, the warehouse and IT team will have to: Capture lessons from the legacy system around what worked well and what instances operators frequently override during daily operations Test the new AI model in shadow, alongside the old one, comparing their recommendations without changing daily warehouse operations Gradually roll out the new system by introducing the new AI model in a single zone (say returns area) with a smaller set of SKUs (during the low volume window in the day) Monitor core KPIs (pick rate, travel distance etc) along with operator override instances to ensure the effectiveness of the new system After 3 – 4 months of close monitoring by the warehouse team, confirm that the new AI model successfully outperforms the legacy system while working seamlessly with the real time fleet coordination. Then the team can retire the legacy system and fully migrate to the new system.