Urbantz shows how operational AI agents should share control

Introduction

A useful AI agent does more than answer questions. It reads operational context, proposes or performs an action and records what happened. That is also where risk starts. A wrong answer in a chat window is annoying; a wrong action in a delivery plan can affect drivers, customers and costs.

Brussels-based Urbantz launched Milo on 13 October as an AI layer for last-mile delivery operations.[4] The product design offers a practical example of shared control: language models handle interpretation and flexible decisions, while existing permissions, business rules, optimisation software and human approvals constrain execution.[4][5]

What happened

Urbantz says Milo works at three levels. Explore answers operational questions using platform data. Agents translate natural-language requests into changes, such as adding late orders to routes. Workflows let teams describe repeatable processes and insert agent, deterministic and human approval steps.[4]

The company says Milo uses the same APIs, route optimiser and permissions as the underlying Urbantz platform. Customers choose where the system is read-only, where reversible actions may proceed and where a person must approve.[4] Milo and the Explore agent became available to existing customers on 13 October, while customer-built agents and broader workflow capabilities are being introduced progressively.[4][5]

Claims about scale, benefits and product performance come primarily from Urbantz. The company says its platform manages more than 50 million deliveries per year and handles about 52,000 weekly changes to planned routes.[5] Logistiek.be reports that independent figures for time or cost savings from the newly launched Milo functions have not yet been published.[5] Those numbers are therefore context and vendor claims, not proof of realised customer outcomes.

Why it matters to business

Many organisations still frame agent adoption as a choice between full automation and constant human approval. That is too crude. If every action waits for a person, the system creates a new queue. If every action runs freely, one mistaken instruction can propagate into operations.

A better design allocates control according to consequence. Reading performance data can run automatically. Drafting a route change can be automatic while execution waits for confirmation. Reassigning a sensitive delivery, changing a promised time or notifying a large customer may always require an authorised person.

Urbantz also separates probabilistic and deterministic work. A language model can interpret “add the late orders to this afternoon’s routes,” but the route optimiser remains responsible for feasibility. That division is broadly reusable. AI should handle ambiguity where it adds value; validated business logic should continue to enforce capacity, accounting rules, eligibility or safety constraints.

The same pattern applies outside logistics. An accounts-payable agent can extract invoice details, but deterministic checks should enforce totals, supplier status and duplicate rules. A sales agent can classify and draft, but pricing floors and contract approvals should remain controlled.

Practical actions for the next 30 days

1. Map one operational decision. Write down the event, required data, available actions, business constraints and accountable owner.
2. Separate reasoning from validation. Let AI interpret unstructured information. Use rules or specialist engines for hard constraints.
3. Grade actions by consequence. Mark actions as read-only, reversible, customer-visible, financial or irreversible.
4. Place approvals selectively. Require people for high-impact exceptions, not every routine step.
5. Use distinct identities. Give each agent its own permissions and audit trail. Never hide automated work behind a shared human account.
6. Log the full chain. Store the trigger, input, decision, tools used, validation result, approval and final action.
7. Design recovery first. Define how to reverse a change, notify affected people and pause future runs.
8. Measure operations, not demos. Track handling time, correction rate, failed actions, approval burden and business outcome.

A practical pilot pattern

Start with a workflow where staff already make frequent, structured adjustments. Keep execution in suggestion mode for two weeks. Compare the agent’s recommendation with the action a trained operator takes, and categorise disagreements.

Next, allow automatic execution only for the low-impact cases where agreement and validation are consistently strong. Keep the rest behind approval. Review thresholds every week. This produces a graduated autonomy model instead of a one-time leap of faith.

How VDS can help

VDS Consultoria builds AI-powered systems that connect to real operations without discarding the controls already keeping those operations reliable. We can model the workflow, integrate the systems, separate flexible reasoning from deterministic checks and implement approval, logging and recovery paths.

The point is not to add a chatbot to an operational screen. It is to create a dependable division of labour between people, agents and existing software. That is how an AI system earns more autonomy over time.

Sources

[4] https://urbantz.com/nl/press/urbantz-launches-milo — Urbantz lanceert Milo en zet AI-agents centraal in de leveringsoperaties [5] https://www.logistiek.be/brussels-urbantz-laat-ai-assistent-milo-leveringsrondes-bijsturen — Brussels Urbantz laat AI-assistent Milo leveringsrondes bijsturen