Isn't "System of Action" just a rebrand of workflow automation?

No. Workflow automation executes pre-scripted rules a human wrote. A System of Action like AgentFleet takes bounded, context-aware decisions using models trained on 20+ years of operational data — including decisions the operator never explicitly encoded. The difference is the decision space: rules cover the known, agents cover the ambiguous.

Where do humans still belong in this stack?

In three places: defining the bounds (what an agent is allowed to do), reviewing true exceptions (the cases the agent flags as outside its confidence), and auditing outcomes. The ratio we see at mature customers is roughly 1 human for every 300–500 agent decisions, versus 1 for every 5–10 in a control-tower world.

How does this work alongside my existing TMS and control tower?

AgentFleet sits above them. Your TMS stays the System of Record. Your control tower stays the System of Visibility. AgentFleet reads from both, takes actions through APIs into carrier systems, settlement platforms, and CX channels, and writes outcomes back. It is additive, not a rip-and-replace.

What's the fastest way to see if this is real?

Pick the most painful decision domain you have — usually settlement, allocation, or CX — and run a 6-week bounded pilot. If it doesn't produce measurable action automation within that window, the theory is wrong for your context. We've run this pilot shape with CEP, FMCG, pharma, and postal operators.

The System of Action: why visibility was never going to be enough

By Soham Chokshi, CEO

Logistics spent the 2010s building dashboards. The dashboards were correct and the outcomes didn’t change. The next decade will be won by software that acts, not software that watches.

What most CXOs believe

Ask a logistics CXO what they bought in the last decade and you’ll hear two answers. First, a System of Record — an ERP, a TMS, a WMS — to get data out of spreadsheets and into a schema. Second, a System of Visibility — a control tower, a dashboard, a real-time tracking layer — to see what was actually happening across the network.

The belief is that these two systems, stitched together well, produce operational excellence. If you can record every shipment and see every exception, the argument goes, the right humans will make the right decisions fast enough. Add an analytics layer, maybe some ML-driven alerts, and you have a modern stack.

This is the story behind every six-figure control tower investment of the last five years. It is also the story behind a universally shared feeling among operators: the tech stack got smarter, the headcount didn’t shrink, and the margin didn’t widen. Something is broken in the theory.

What’s actually happening

The honest truth is that Systems of Record and Systems of Visibility share a fatal design assumption: that a human will take the next action. They were built to inform a dispatcher, a planner, a CX lead, a finance controller. When the number of decisions per shipment exploded — multi-carrier allocation, dynamic routing, exception handling, dispute resolution, first-mile windowing, customer comms — the human layer became the bottleneck the dashboards couldn’t see.

At Shipsy we ran the numbers across our 250+ enterprise deployments. In a typical CEP operation, a single parcel generates 40+ decision points between order capture and POD. In FMCG primary distribution, a single load triggers 60+ settlement and reconciliation events. Control towers render these as rows. A human resolves maybe 15% of them before SLA. The rest age into cost.

The unlock isn’t better visibility. It’s action. Specifically, it’s software that takes defensible, bounded decisions autonomously and escalates only true exceptions. That is why we built AgentFleet — Clara for CX resolution, Nexa for settlement automation, Vera for dispute resolution, Astra for planning and routing — as a System of Action layered on top of the record and visibility stacks, not a replacement for them. Vera alone has autonomously resolved $25M+ in carrier/vendor disputes at Heineken. DPD Poland recovered $37M in unit economics through AI-native routing. Aramex unlocked $27M in cross-border throughput. These aren’t dashboard wins. They are decisions that were never going to be made in time by a human staring at a screen.

What to do in the next 90 days

If you run logistics for an enterprise, I would do three concrete things in the next quarter.

Audit your decision latency, not your data latency. Most CXOs know the freshness of their tracking data. Very few know how long the median exception sits unresolved after it surfaces in the control tower. Pull the numbers. If your P50 exception-to-resolution is over 30 minutes and your P90 is over 4 hours, you do not have a visibility problem. You have an action problem, and no amount of dashboard investment will close it.

Pick one bounded decision domain and automate it end-to-end. Not “AI everywhere.” One domain. Carrier allocation for a single lane. First-attempt-failure rescheduling for one city. Freight invoice reconciliation for one 3PL. Automate the decision, not just the alert. Measure: what percentage of cases now resolve without a human? What is the false-positive rate on the escalations that do reach a human? If the answer to the first number is below 70%, you are still in copilot mode, not action mode.

Rewrite your vendor evaluation criteria. When you next evaluate a logistics platform, stop asking “can you give me visibility into X?” Start asking “what decisions does your system take without a human in the loop, and what is your published accuracy on those decisions?” Any vendor who pivots back to dashboards, alerts, or “recommendations” is selling you last decade’s architecture in 2026 clothing.

The organizations that will widen margin in 2026–2027 are the ones that treat their dispatcher, their settlement clerk, and their CX rep as the scarce human layer they are — and build an agent layer that absorbs the decisions those humans were never going to get to.

Why this matters now

The cost side of logistics is not getting easier. Fuel volatility, driver scarcity, e-commerce fragmentation, and SLA pressure are all compounding. The margin recovery available from “better visibility” is close to zero. The margin recovery available from shifting decisions from human to agent is, based on what we see across Heineken, Aramex, DPD Poland, and Catalent, between 2 and 8 points of operating margin depending on vertical. This is the gap that will separate the operators who compound and the ones who get acquired over the next 24 months.