The Seven Tenets of an Agentic TMS

The legacy promise was that visibility would make planners faster. A decade in, the data says otherwise. When the control tower flags a carrier rejection spike or a regional disruption, the dashboard demands a planner who can model alternatives, assess downstream impact across inventory echelons, and execute a rebooking in the TMS. The window for effective mitigation usually closes before the planner finishes interpreting the alert.

An agentic system reverses the flow. The carrier rejection is not an alert; it is the trigger. The agent evaluates SLA risk, queries spot capacity, calculates cost-to-serve, tenders to the new carrier, updates the ledger, and notifies the customer. McKinsey's March 2025 work with COOs frames the shift as one from AI-as-recommendation-engine to AI-as-execution-engine.

Every vendor claims continuous learning. Configuration looks like learning when a planner only sees the output. The difference shows up when conditions change. A carrier starts rejecting tenders on a specific lane because of an unannounced capacity reallocation. A configured system keeps offering that carrier until someone codifies a new rule. A learning system observes the rejection pattern, deprioritises the carrier on that lane, and shifts volume, all without an admin altering the routing guide.

Crucially, the system is not changing itself on its own. It is absorbing signal from your operators' overrides and from the outcomes your network produces, then refining its decision policies within the boundaries you set. Deloitte's Tech Trends 2026 makes the point directly: organisations that still improve sequentially cannot keep pace with those running continuous learning loops.

The fake-proof test: ask the vendor to walk you through three specific behaviour changes the system made in the last 90 days, each with the data signal that triggered it and the audit-trail entry that recorded it.

The risk in batch is not slowness. It is silent algorithmic failure: a mathematically optimal decision against a reality that no longer exists. A 30-minute-old picture of inventory is enough to make a confident, wrong allocation. A 24-hour-old picture is enough to do it at scale.

Event-driven architecture treats every state change in the physical world (a geofence crossing, a reefer temperature deviation, a carrier rejection, a customer reschedule) as a discrete signal the decisioning layer consumes in milliseconds. McKinsey's February 2025 supply-chain work notes that real-time, connected ecosystems remove silos and produce "lower costs, smaller inventories, and fewer lost sales." SCMR's June 2024 coverage of the CSCMP State of Logistics frames the broader context: logisticians are adapting to "permanent volatility." A system that integrates that volatility as it arrives produces different decisions than one that absorbs it overnight.

Components can be replaceable through APIs, but they can also be replaceable through config layers or through natural-language policy interfaces that operators write into the system directly. What matters is whether you can pull the carrier-management module without re-implementing the pricing engine or the optimisation layer. The seam between components and modules.

Gartner's January 2024 "Predicts: Composable Modularity Shapes the New Digital Foundation" calls this the foundational architecture for "continuous access to adaptive change." In practice this looks like Packaged Business Capabilities, MACH-style decoupling, and agent-builder interfaces that let an enterprise hot-swap a routing optimiser or a settlement workflow without triggering a regression test across the entire suite.

The agentic shift demands a layered governance regime built into the architecture, not bolted on after a deployment review. Forrester's March 2025 analysis frames the requirement plainly: scaling agent workflows depends on "effective data and communications governance," not assumed system benevolence.

In practice this means every autonomous decision must carry its own receipt: why it was made, what data triggered it, what the outcome was, and whether a human needs to intervene. The system should let operators dial autonomy up or down by decision type, run new policies in shadow mode before they touch live freight, and score its own choices against what a planner would have done.

Most legacy TMS were designed for a happy-path operation: EDI arrives clean, dock times hold, inventory updates synchronise. Real logistics is none of those things. Carrier APIs fail. GPS pings drop. POD scans come in late. Weather closes a corridor at 03:00 and the system finds out at 09:00. Naive systems treat these as edge cases. Agentic systems treat them as the default operating condition, and they architect for it.

The cost of getting this wrong is silent. A traditional application that hits bad data throws an error. A naive AI agent that hits stale data confidently executes a wrong decision and nobody knows for hours. McKinsey's April 2025 supply-chain research frames the operating reality: companies are "striving for end-to-end visibility, efficiency, and agility... yet outdated infrastructure, fragmented data, and supply chain disruptions make these goals difficult to achieve." Ground-truth orchestration is what closes the gap. Exception handling is downstream of it; without ground-truth, exception handling is just reconciliation labour.

Local or node optima are routinely network-pessimal. Saving $200 on a cheaper truckload can delay a delivery, trigger stockouts at the destination, and force $10,000 in expedited primary freight. The TMS that picked the cheap leg "won" on its own metric; the enterprise lost. Network-aware decisioning is what the industry calls Multi-Echelon Inventory and Transportation Optimization (MEIO). The system weighs every transport decision against inventory holding, real-time demand, facility throughput, and working-capital constraints in real time.

The empirical case is increasingly hard to argue with. BCG's October 2025 work on AI-first companies documents a medtech firm that embedded AI forecasting across its end-to-end planning process and optimised multi-echelon inventory: 60% reduction in backorders, more than $125M in inventory reduction in under 12 months. Those numbers don't come from local optimisation. They come from a system whose decisions are conditioned on the whole network.