AI-Driven Dispatch and Allocation: A Practical Autonomy Model for Deciding What the System Decides

Why the level question is usually asked wrong

Logistics technology leaders ask how much autonomy their operation is ready for. It is a reasonable question and it is unanswerable as posed, because it treats autonomy as a single organisational dial.

In practice, AI-driven dispatch and allocation makes many kinds of decisions, and they carry completely different risk profiles. Resequencing the remaining stops on a route is reversible, internal, and visible within the hour. Purchasing spot capacity above a threshold is financial, external, and not reversible at all. Nobody should run those two at the same autonomy level, and an operation that sets one dial has necessarily set it wrong for one of them.

So the model below applies per decision category. The output of using it is not a level, it is a matrix: this decision at level four, that one at level two, this other one at level one until we have the governance for more.

That framing matters commercially as well as operationally. Gartner predicts that more than 40 percent of agentic AI projects will be cancelled by the end of 2027, attributing this to escalating costs, unclear business value, and inadequate risk controls rather than to model capability. A per-decision approach addresses the third directly, because it never asks the organisation to accept more exposure than its controls support.

Also Read: Logistics Orchestration Autonomy Is a Portfolio, Not a Single Level in 2026

The five levels

Level 1: Advisory

The system analyses and recommends. A person decides and executes. The recommendation may be excellent and nothing happens without human action.

Governance precondition. None beyond normal change control. This is where every decision category starts.

What it is good for. Establishing whether the model’s recommendations are sound, on decisions where being wrong is expensive. Also the correct permanent level for a small number of decisions that should never be automated.

The trap. Staying here across the board, which produces the cost of the platform without the benefit, because throughput remains capped by how many recommendations a person can process.

Level 2: Approval-gated

The system decides and prepares the action. A person approves before it executes. The difference from level one is that the work is done, not just suggested.

Governance precondition. A defined approver, a defined approval window, and a defined behaviour when nobody approves in time. That last one is where this level usually fails, because unapproved decisions queue silently.

What it is good for. Decisions with commercial consequence where the reasoning is auditable and the volume is low enough that approval is not a bottleneck.

Level 3: Bounded autonomous

The system executes within explicit limits and escalates outside them. A reassignment within a cost delta executes; one beyond it escalates.

Governance precondition. Written bounds per decision, a mechanism that enforces them technically rather than by convention, and a decision record for everything executed.

What it is good for. The majority of operational dispatch decisions. This is where most of the value sits, because it removes per-decision human involvement while keeping consequential cases in front of a person.

The design work. Setting the bounds is the whole exercise, and it should be done with the operations lead rather than by the vendor, since the bounds encode business judgement rather than technical capability.

Level 4: Autonomous with review

The system executes and a person reviews after the fact, in aggregate rather than case by case, looking for patterns rather than approving instances.

Governance precondition. An evaluation mechanism that measures decision quality against outcomes, a review cadence with a named owner, and the ability to reconstruct any individual decision months later.

What it is good for. High-volume, reversible, internal decisions where per-case review adds cost and no safety.

What changes for the team. The job becomes reviewing distributions rather than instances, which is a genuinely different skill and needs to be named as such rather than assumed.

Level 5: Self-adjusting within policy

The system adjusts its own parameters within a meta-policy a human set. It might tighten a threshold because the data supports it, without asking.

Governance precondition. Everything from level four, plus explicit constraints on what the system may adjust and by how much, plus a record of parameter changes as first-class events. Without the last, nobody can explain why behaviour changed in March.

What it is good for. Parameters that drift with conditions, such as expected service time per location, where the human judgement is in the objective rather than in the value.

Where it does not belong. Anything with contractual, regulatory, or safety consequence. Self-adjustment is appropriate for calibration, not for policy.

Also Read: Autonomous Doesn’t Mean Ungoverned: Building the Governance Layer for Logistics AI Agents

Where to start, by decision category

This is the table to take into a design session. The starting levels are a defensible default rather than a prescription, and the gate column is what should move a decision up.

Two observations from the table.

The reversible column predicts the starting level almost entirely, which is the useful shortcut when a new decision type arrives and nobody has thought about it yet.

And a mature operation shows a spread across at least three levels. A matrix where everything sits at the same level indicates the levels were set by policy rather than by analysis, which is the failure this model exists to prevent.

Also Read: Agentic-Washing: How to Tell a Real Agentic TMS From a Rebranded Rules Engine in 2026

The evidence gate

Advancement should be gated on data rather than on a quarter passing.

The single most useful signal is override rate by decision category, tracked with reasons. Three readings matter.

Falling override rate with clustering reasons means the constraint model is catching up with reality and the decision is a candidate for advancement.

Flat override rate means either the model is missing something nobody has diagnosed, or the team has not accepted the change. Both need addressing before advancing, and they need different responses.

Zero overrides is not a success signal. It usually means dispatchers have stopped engaging, which removes the corrective mechanism you were relying on.

The outcome evidence supports the direction of travel. Gartner found that organisations using autonomous planning technologies reported improvements in decision speed at 78 percent and decision quality at 75 percent while reducing manual intervention, and predicts 60 percent of supply chain disruptions will be resolved without human intervention by 2031. McKinsey has found that with advanced system support, 80 to 90 percent of planning tasks can be automated while still delivering better quality than the same tasks performed manually.

The constraint is not capability. It is governance readiness, and that is measurable. Deloitte found that only 21 percent of organisations report having a mature governance model in place for agentic AI, based on a survey of 3,235 IT and business leaders across 24 countries.

Also Read: Which Dispatch Decisions Should Your AI Make? A Decision-by-Decision Autonomy Map for 2026

What has to exist before each level

Governance is not a document, it is a set of capabilities the platform and the organisation have to hold. Five, and the level you can operate at is the level your weakest one supports.

A decision record. What was decided, on what inputs, under what policy. Required from level three, and the thing most often absent when a decision is questioned six months later.

Enforced bounds. Limits implemented technically rather than as guidance. A bound that relies on nobody exceeding it is not a control.

An evaluation mechanism. Something measuring decision quality against outcomes rather than measuring uptime. Required from level four, and rarely built without deliberate design because outcome labels arrive after the decision.

An override path that is used. Not reserved for emergencies. An override mechanism nobody exercises is untested, and it is the control everyone points to in a review.

A named owner per decision category. Someone accountable for the bounds and the outcomes those bounds produce, with the authority to change them.

Ask which of the five you have, honestly, before setting any level above three. Operating above your governance is the specific failure Gartner’s cancellation data describes.

Where Locus fits

Locus, the world’s first Decision-Intelligent, Agentic TMS, implements autonomy as a per-decision property rather than a system-wide mode, which is what makes the matrix above configurable rather than aspirational.

Six governance mechanisms bound autonomous action: explainability, traceability, evaluation, autonomy levels, an execution sandbox, and human-in-the-loop override. Mapped to this article, autonomy levels are the dial, traceability is the decision record, evaluation is the quality mechanism, the sandbox is where a level change is tested before production, and human-in-the-loop is the override path. Within DiSCO, eight named agents make different decisions, which is what allows the Dispatch agent to run at one level while the Carrier agent runs at another.

Locus has been recognized by Gartner for seven consecutive years, featured in the 2026 Hype Cycle for Supply Chain Execution and Logistics Technologies, named a Leader in TMS by QKS Group (SPARK Matrix), and ranked #1 in Route Planning on G2’s 2026 Best Software Awards. In October 2025, Ingka Investments, the investment arm of Ingka Group, the world’s largest IKEA retailer, acquired Locus. Locus continues to operate independently.

Two deployments show the model operating. A Fortune 50 parcel and logistics provider governs 4,500+ drivers, 1,500+ captive and 3,000+ third-party, across 51 service-centre locations, with zone-based, tendering, dynamic, and on-demand assignment logic running inside one engine and every autonomous decision logged for explainability, traceability, and human override. Weekly execution moved from 75 percent to 92 percent. Different assignment types running under different logic in one operation is the matrix in production.

Siam Makro shows the organisational half. Across 160+ stores and 10,900+ active riders, continuous wave-based dispatch replaced two hours of human planning per store per day with under 30 minutes of agentic execution, and planners moved from clicking dispatch to setting policy. That shift, from executing decisions to governing them, is what advancing a level actually means for the people doing the work.

Also Read: AI-Driven Dispatch and Allocation Software: A Practical Evaluation Guide for Enterprise Logistics Leaders in 2026

The exercise to run

List every decision your dispatch operation makes in a day. Mark each one reversible or not, internal or customer-facing, and financial or operational.

Then assign a starting level using the table above, and note which of the five governance capabilities you would need for the next level up. That document is your autonomy roadmap, and it is more useful than any maturity assessment, because it names decisions rather than describing a stage.

Most operations find two things when they do this. Several decisions are already effectively autonomous without anyone having decided that deliberately. And one or two are being reviewed case by case at considerable cost with no evidence that review is catching anything.