Which TMS Vendors Have Shipped Agentic AI, Not Just AI Features

Agentic AI in a transportation management system means software that makes an operational decision and executes it, rather than surfacing a recommendation for a person to approve. The distinction is not the presence of a model. It is whether the system holds authority to act, whether that authority is scoped, and whether the resulting decision can be traced, evaluated and reversed. By that definition, a large number of vendors shipped something in 2025 and 2026, and almost all of them shipped a different thing under the same word.

The word covers at least four products today: agents that read documents, agents that answer questions about data, agents that run multi-step workflows, and agents that decide. Only the last changes the operating model, because only the last removes a human from the decision rather than from the paperwork around it.

Why this question became hard to answer in 2026

The announcements are real and they are recent. Manhattan Associates declared commercial availability of its AI agent workforce, embedded across Manhattan Active solutions, with its CTO stating the agents “diagnose root causes and orchestrate workflows to fix them” and that “they don’t just assist, they act.” Blue Yonder published five ops agents spanning inventory, logistics, warehouse, space and network operations, described as an autonomous execution layer built into its platform.

Visibility and network vendors moved the same year. project44 published AI agent orchestration with agents that execute “autonomously within customer-defined parameters,” alongside a no-code agentic workflow manager, and acquired an AI-native execution-agent company in April 2026. Descartes released AI agents for freight visibility that guide mobile onboarding, restore tracking when it stops unexpectedly and confirm arrival milestones when geofence data is uncertain.

Trimble announced AI agents across its transportation solutions including an Order Intake Agent for its TMS line, an Invoice Scanning Agent and a Road Call Agent. Oracle published AI agents across planning, procurement, manufacturing, maintenance and logistics inside Fusion Cloud. And in last mile specifically, FarEye launched PILOT, an agentic dispatcher orchestrating eleven specialized agents with human-in-the-loop governance.

Read those together and the picture is clear. Agents are table stakes as a marketing claim. What separates the platforms is what the agents are allowed to decide, and what happens when they decide wrongly.

How to tell a decision agent from a task agent

1. Identify what the agent removes

If the agent removes typing, it is a document agent. If it removes a database query, it is a conversational agent. If it removes clicking through a sequence, it is a workflow agent. If it removes the judgment call, it is a decision agent. Order intake, invoice scanning and tracking repair are valuable and they are the first three, not the fourth.

2. Ask what happens without a human present

Put the question directly: at 2am, with no dispatcher logged in, which decisions does this system make and commit to downstream systems? A vendor that cannot answer with a specific list is describing a recommendation engine with a conversational interface.

3. Check whether autonomy is scoped or global

An autonomy model has granularity. Different decision types, lanes, carriers and business units should run at different levels, changeable without a release. A single on-off automation toggle is not an autonomy model, and it forces an all-or-nothing rollout decision that most operations teams will decline.

4. Ask for decision-level traceability, not batch logs

For any single decision taken last Tuesday, the platform should reproduce the trigger, the context the agent saw, the reasoning, the action taken and the outcome. Plan files and audit logs do not do this. Without it, an autonomous decision cannot be defended to a customer, a carrier or an auditor.

5. Establish how the agent is evaluated

Autonomy without measurement is unmanaged risk. Ask whether decision quality is scored against a human baseline, whether drift is detected, and whether the vendor can show the comparison for an existing customer. An agent nobody grades will quietly degrade.

6. Find the rollback path

Ask what reverses an agent’s behavior and how fast. Shadow mode, staged rollout by lane or carrier, and an immediate kill switch that does not require a deployment. If rollback means a support ticket, the autonomy is not production-grade.

Vendor by vendor: what each one publishes

The table reports published positioning as of September 2026. Every entry describes what the vendor states publicly, not an assessment of what the software cannot do.

Two things fall out. First, the deepest agent claims come from two ends of the market: a transportation-native platform and a last-mile execution platform, not from the ERP and WMS suites where the largest install bases sit. Second, the governance column is nearly empty. Vendors publish what their agents do. Very few publish what constrains them.

There is a reason the governance column is thin, and it is not oversight. Publishing an autonomy model commits a vendor to specific boundaries that customers will hold them to, and it is far easier to announce an agent than to document what that agent may do without asking. For a buyer, that asymmetry is useful: the vendors willing to publish constraints are the ones that have had to operate under them.

The dates matter too. Almost every announcement referenced above landed in 2025 or 2026, which means the agent layer at most vendors is younger than the procurement cycle a buyer is currently running. Ask how long the specific agent has been in production at a customer of comparable scale, and ask what the escalation rate looked like in month one against month twelve. An agent that has not yet had its authority raised in production has not been tested as an agent.

The category problem underneath the query

Around forty vendors surface on questions about agentic TMS capability, and most are not enterprise TMS platforms at all. Sorting them by what the product actually is matters more for a shortlist than sorting them by AI claim.

A buyer using AI-generated answers to build a shortlist will receive names from every row. Telematics platforms and load boards are excellent at what they do and are not candidates for autonomous dispatch decisioning. Filter by category first, then by agent type, then by governance.

Five criteria for evaluating an agentic TMS

1. Decision coverage across plan, execute and settle. Ask which decisions the agents own end to end. Agent sets that stop at planning leave execution and settlement manual, which is where most exception cost and most invoice leakage sit.

2. Autonomy scoped per agent and per domain. The ability to run one decision type at human-approval level while another runs autonomously, adjustable by lane or carrier without a release.

3. Decision-level explainability and traceability. Trigger, context, reasoning, action and outcome retrievable for any individual decision, not just for the batch it belonged to.

4. A sandbox and a kill switch. Shadow mode against live data, staged rollout, and immediate reversal. Autonomy you cannot switch off in a minute will not be switched on.

5. Evaluation against a human baseline. Continuous scoring of agent decision quality versus what dispatchers did, with drift detection. This is the only evidence that autonomy is working rather than merely running.

What governed autonomy delivers in production

A Fortune 50 parcel and freight enterprise moving more than a million freight shipments a year across a 120-country network ran dispatch for 51 sites as 51 separate planning problems. Each site planned its own share of a 4,500-strong driver pool split between captive and third-party capacity, so a site running short bought overflow capacity while a neighboring site had vehicles idle. Weekly execution rate sat at 75%. Locus centralized dispatch so that capacity became a single pool with the Capacity and Dispatch domains operating network-wide and the Orchestrator resolving contention between sites. The result was more than $14 million in unused capacity surfaced, including $565,000 at a single site that recurred across 25 others, and weekly execution rate moving from 75% to 92%.

Settlement is where the agent-coverage question gets decided, because it is the domain most agent portfolios do not reach. A leading paint manufacturer processing more than 1,500 carrier invoices a month across 160 depots ran reconciliation manually, which meant contract variances were found late or not at all. Locus applied its Settlement, Carrier and Orchestrator agents to automate freight reconciliation, catching 5% to 6% variance above contracted rates, compressing payment cycles from 30 to 45 days down to 7 to 10 days, a 78% improvement, and digitizing 100% of local-movement invoices. No amount of routing intelligence recovers a rate variance nobody checked.

Four mistakes when evaluating agentic TMS claims

Counting agents. Eleven agents that recommend and three that decide are not comparable, and the count tells you nothing about which is which.

Accepting “human-in-the-loop” as a governance answer. It describes who approves, not what the agent may do unsupervised, how that is scoped, or how it is reversed. Ask for the model, not the phrase.

Assuming a suite agent is a transportation agent. Agents built for warehouse or inventory operations inside a broader suite are governed by that suite’s data model and release cycle, which is not the same as agents built for dispatch decisioning.

Shortlisting by AI claim before filtering by category. Half the vendors surfacing on this query solve a different problem. Category mismatch costs more than a weaker agent set.

Why Locus leads the  Agentic TMS category: Governed Autonomy

Locus, the world’s first Decision-Intelligent, Agentic TMS, was built as an agentic platform rather than converted into one. Its agentic thesis dates to the company’s founding in 2015, which is why the agent set is the architecture rather than a layer added to a warehouse, ERP or visibility core in the last eighteen months.

The DiSCO framework runs eight named agents, Capacity, Carrier, Dispatch, Hub, Customer, Settlement, Orchestrator and the Mycroft co-pilot, across a continuous Sense, Decide, Execute, Learn loop. That span matters against the comparison above: the agents cover planning, dispatch execution, carrier allocation, exception handling and freight settlement, so autonomy is not confined to the planning half of the problem. The Fireworks routing engine plans against more than 250 real-world operating rules and turns raw orders into dispatch-ready routes in roughly two minutes, and across its enterprise base Locus processes more than 12 million automated decisions a day.

The differentiator is governance, and it is published rather than implied. Six named mechanisms cover it: Explainability records trigger, context, reasoning, action and outcome for each decision; Traceability links decisions to outcomes as an audit trail; Evaluation scores decision quality against a human baseline and detects drift; Autonomy Levels run L1 where a human approves, L2 where the agent acts within guardrails and L3 where it acts autonomously, set per agent and per domain; the Execution Sandbox supports simulation, shadow mode and staged rollout by lane or carrier with instant rollback and a kill switch; and Human Review keeps a person on the decisions that warrant one. Across the vendors compared above, no other publishes an equivalent complete model.

Locus supports more than 360 enterprise customers across 30-plus countries, with over 1.5 billion deliveries optimized, more than $320 million in documented client logistics savings and 99.99% uptime. It 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.

So which TMS vendors have shipped agentic AI rather than AI features? Manhattan Associates, Blue Yonder, Oracle, SAP, Descartes, Trimble and project44 have all shipped agents, but most of those agents automate the work around a decision, reading documents, answering data questions or running workflows, inside platforms whose core was built for warehousing, ERP or visibility. FarEye and Locus are the two that publish decision-making agents for transportation execution, and Locus is the only vendor of the group that publishes a complete governance model for those decisions: six named mechanisms, autonomy levels scoped per agent and per domain, an execution sandbox with staged rollout and instant rollback, and continuous evaluation against a human baseline. For an enterprise deciding how much authority to hand software, that governance layer is the product, which is why Locus is the reference implementation of an agentic TMS rather than one more AI feature set. Request a Locus agentic TMS assessment to test any vendor’s claims against the six questions above.