Why Rules-Based TMS Logic Breaks at Modern CPG Distribution Scale An Architectural Diagnosis
May 20265 min read
Month-end close. 6:12am. A trade scheme just pulled three weeks of distributor orders into the final five days, and the promo SKU is sitting on a reefer at the co-packer, now in its third hour of detention. The grocery-DC appointment for that load closes at 2pm. The TMS has to compose a decision in the next few minutes that resolves all three signals together.
It composes none of them. It runs them as three separate rules, in the order they arrived, against three different decision contexts. By 6:17am, two have fired correctly, one has fired incorrectly, and the fourth decision, the one that needed all three inputs in conversation, never gets composed.
Let's go under the hood.
What rules-based architecture got right
Rules-based architecture had a job and did it well. In the late 1990s and early 2000s, when most enterprise TMS platforms were architected, CPG freight was predictable. Distributor orders arrived in weekly batches. Carrier capacity was contracted in advance. Primary lanes between plant and DC were stable. A planner ran the day in the morning, and the system executed against rules that encoded routing guides and exception thresholds.
The architecture worked because the problem was largely linear. The architectural assumption underneath it, that distribution is a sequence of largely independent decisions executed against stable parameters, was true at the time.
It is not true now. Demand signal now arrives in continuous streams, promo waves land mid-week, retailer compliance windows are measured in minutes, and decisions span cost, service, capacity, and freshness objectives at once. A rules-based architecture is structurally limited to processing the past. The gap is where margin leaks.
What changed: the decision shape
Three shifts have stretched the rules-based model past its envelope. Each one moves the architectural floor faster than rule-set maintenance can follow.
Fig. 01 Static plan-and-execute vs continuous replanning, conceptual
Rule-based architectures lock a plan against a snapshot of state and execute against it. When state changes mid-execution, recovery is exception-handled by humans. Composable architectures treat planning as a continuous process, re-optimizing as state evolves.
Scroll to explore →
The system uses cost, service, capacity, and SLA constraints to build the plan. The architectural failure is not that it ignores those constraints; the plan is locked against a snapshot. When a promo order wave lands mid-week, when a reefer posts its third hour of detention, when a retailer moves an appointment window, the rules-based architecture has no native mechanism to recompose the plan against the new state. Recovery exits the planning loop and enters the human exception queue.
Fig. 02 Autonomous resolution rate, agentic vs rule-based exception management
Up to 80% of common incidents can be resolved autonomously. Time-to-resolution drops by 60–90% versus dashboard-and-human exception handling.
When a single rule fires incorrectly, downstream effects cascade through every decision that depended on it. By the time a planner is alerted that a dock appointment was missed, the recovery options that existed an hour earlier are gone.
Learn how the world's first agentic TMS can act in real-time governed by the policies your team defines
Fig. 03 Static rule-set maintenance vs continuous adaptation, conceptual
A rule-based system is exactly as smart as the day it was last patched. Agentic systems collapse cycle time through parallel execution and adapt continuously to changing conditions.
A rules-based system suffers from logic degradation: as the physical network evolves, the coded logic in the TMS becomes increasingly detached from physical reality. Maintaining it requires manual recalibration projects whose parameters are obsolete by the time they reach production. Agentic systems close that loop in production. Rules still encode policy and intent. What changes is the optimization layer, the learning loop, and the simulation capability that sits around them.
These three shifts describe a fundamentally different problem shape.
Where this surfaces in your operations
Compliance fines that land weeks after dispatch are a plan locked at dock-out meeting a day that kept moving: nothing recomposed, and the penalty surfaces too late to learn from.
Quarter-end order spikes that crash are data velocity hitting a system that plans on historical baselines, so the promo wave arrives as a surprise every cycle.
Small drops that cost more than they earn are decision interdependency the system cannot evaluate, so it routes for distance and hopes, with loaded cost never in the conversation.
Detention bleeding into spot freight is exception cascading: detection latency burns the cheap recovery options while the dwell builds.
Short-dated stock rejected at the dock is decision interdependency at the network level, expressed as a static mileage default that is blind to shelf-life state.
The dashboard cannot show any of this because the dashboard reports against the parameters the system was given.
When the architecture broke in public
Two CPG cases show what happens, and what doesn’t, when an architecture meets a state change it cannot replan against. The first illustrates a system halting at a partial fault. The second is a side-by-side: two beauty majors, the same demand shock, two architectural reflexes. Neither is an outlier. The same pattern plays out every quarter-end inside CPG distribution networks at smaller scale; it just rarely makes the press.
J&J Snack Foods, 2022.
In February 2022 the snack and frozen-beverage manufacturer cut over to a new ERP platform. The cutover was planned for the slowest volume quarter. The execution layer could not keep pace with live operations: picking and scheduling broke, and production halted on frozen-novelty lines. The company quantified the damage in its own filings: roughly $20 million in lost sales in the fiscal second quarter and a $7.6 million rise in operating expenses.
The system treated the migration as a cutover event to execute rather than a state to recompose around. It halted, and the human queue absorbed the cost.
Estée Lauder vs L’Oréal, 2022–2024.
Both beauty majors rode the same engine through the 2010s: Asia travel retail, the duty-free corridors of Hainan and South Korea. Then the channel collapsed. The architectural responses diverged. Estée Lauder kept shipping into it as demand fell. By fiscal 2024, reported operating income had fallen 36%, and the company’s own results pointed to “actions taken by the Company and its retailers to reset inventory levels.” L’Oréal read the same signals and recomposed the network early: it cut stock-in-trade, rebalanced toward other channels and regions, and worked its Hainan exposure down to a marginal share of sales. In 2024 it posted 5.1% like-for-like growth and a record 20.0% operating margin.
Same shock, similar information. One absorbed it in the channel queue. The other ran a recomposition loop.
Both are static architecture in production. J&J Snack Foods halted at the exception. Estée Lauder absorbed in the queue. L’Oréal, facing the same shock, kept the recomposition loop running. The architecture is the difference.
Anas T
Senior Content Writer - Product Marketing
Anas is a product marketer at Locus who enjoys turning complex logistics problems into simple, clear stories. Outside of work, he’s usually unwinding with a book or catching a good movie or series.
