Azure OpenAI for Logistics: Running Dispatch AI Inside Your Microsoft Tenant

Azure OpenAI for logistics is the natural choice for enterprises that already run their business on Microsoft. Many operate ERPs such as Acumatica or Dynamics, identity through Entra ID, and a Microsoft enterprise agreement with committed Azure spend. Running dispatch AI through an Azure OpenAI deployment means the model calls happen inside that same tenant boundary, in a region you picked, billed to a subscription you already manage.

Geofleet supports this through its bring-your-own-LLM setting. This guide explains how to point Geofleet at your Azure OpenAI deployment, why enterprises prefer this arrangement, how to choose model deployments, how to plan quotas and rate limits, and what your IT team needs to check before go-live. We keep Azure claims to well-documented, generally available capabilities.

Why Microsoft-first enterprises pick Azure OpenAI

The appeal is not the models themselves, which are the same OpenAI model families available elsewhere. The appeal is where they run and how they are governed.

• Tenant boundary: the Azure OpenAI resource sits in your subscription and resource group, subject to your Azure policies, role assignments, and audit.
• Regional deployments: you choose the Azure region for each deployment, which settles most data residency questions for delivery addresses and customer details.
• Existing agreement and billing: usage is billed to your Azure subscription under your enterprise agreement, with the cost management, budgets, and alerts you already use.
• Private networking options: Azure OpenAI resources can be restricted to selected networks or reached through private endpoints, in line with your network policy.
• Familiar operations: keys live in your key vault, logs flow to your monitoring, and rotation follows your existing runbooks.

If you are still weighing whether to bring your own provider at all, our launch post on bring your own LLM covers the general case. The rest of this guide assumes you have decided on Azure.

What you need before you connect

Geofleet asks for three values from your Azure OpenAI resource. Gather them from the Azure portal before opening Geofleet settings.

The deployment name matters more than people expect. In Azure OpenAI you deploy a specific model version under a name you control, and the application calls the name, not the model. That indirection is what lets you upgrade the underlying model later without touching Geofleet.

Pointing Geofleet at your Azure OpenAI deployment

1. In Azure, create an Azure OpenAI resource in the region you want data processed. Put it in a resource group dedicated to Geofleet so access and cost are easy to scope.
2. Deploy at least one model under a clear deployment name. If you plan to tier models (recommended below), deploy two.
3. Copy the endpoint and one of the two keys from the Keys and Endpoint page. Store the key in your key vault and record which key slot you used so rotation is straightforward.
4. In Geofleet, open Settings, then Developer, then Integrations, and scroll to LLM model providers.
5. Select Azure OpenAI. Enter the endpoint, the deployment name, and the key. Save.
6. Open the dispatch copilot and run a test question, for example asking which routes are behind schedule. Confirm the response arrives.
7. In Azure, check the resource metrics page to confirm the request was counted against your deployment.

Choosing model deployments

A dispatch platform generates two very different kinds of model calls. There are thousands of short, structured calls a day: cleaning an address, classifying a delivery note, picking a category. And there are fewer, longer calls where a copilot answers a question from several tool results or an agent works through an exception. Serving both from a single frontier deployment wastes money on the first and can starve the second of quota during peak hours.

Because Geofleet calls the deployment name, you can retire an older model version by deploying the newer one under the same name, or by creating a new deployment and updating the name in Geofleet settings. Either way it is a configuration change, not a code change. Our cost guide shows how much of the total bill each tier typically carries.

Quota and rate-limit planning

Azure OpenAI assigns quota per deployment, expressed in tokens per minute, with a corresponding requests-per-minute ceiling. Exceeding it returns rate-limit errors. Geofleet retries with backoff, but an agent that keeps hitting the ceiling during the morning dispatch rush will feel slow, and a copilot answer that takes twenty seconds gets abandoned.

A simple sizing method

1. Count your peak-hour AI calls. For a 500-order operation, imports and address cleanup cluster in a one to two hour window, so assume most of those 500 short calls land there.
2. Multiply by tokens per call. Short calls are a few hundred tokens; copilot and agent calls can be several thousand once tool results are included.
3. Convert to tokens per minute at peak and add headroom of at least 50 percent for bursts and retries.
4. Request quota for each deployment accordingly, and revisit after two weeks of real usage data from Azure metrics.

For sustained, predictable volume, Azure also offers provisioned capacity options that reserve throughput. Whether that is worth it depends on your volume and the pricing available under your agreement, which your Microsoft account team can quote.

"Our first week we put everything on one deployment and the copilot slowed down every morning at nine. Splitting the address cleanup onto a small deployment fixed it the same day and cut the bill."

— Platform engineer, national parcel network

Networking, logging, and monitoring

Geofleet reaches your Azure OpenAI endpoint over HTTPS. If your network policy restricts the resource to selected networks, you will need to allow the Geofleet egress addresses or expose the endpoint in a way Geofleet can reach. Discuss this with your Geofleet contact before locking the resource down, so the test query in step six does not fail for network reasons.

On the monitoring side, Azure exposes per-deployment metrics for requests, tokens, and rate-limit responses. Route diagnostic logs to your log analytics workspace, set a cost budget on the resource group, and add an alert on rate-limit responses so quota problems surface before dispatchers notice them. Geofleet keeps its own audit log of which AI tool calls were made on your data, which complements the Azure view.

IT checklist before go-live

1. Subscription and resource group chosen; Azure OpenAI resource created in the approved region.
2. Two deployments created and named; model versions recorded.
3. Quota per deployment set with headroom; provisioned capacity decision made if volume warrants it.
4. API key stored in key vault; rotation runbook written using the two-key swap.
5. Network access policy applied and verified against a Geofleet test query.
6. Cost budget and alerts configured on the resource group.
7. Diagnostic logs routed to your monitoring workspace; rate-limit alert created.
8. Geofleet LLM model providers configured, tested, and the Geofleet audit log reviewed after the first day.

Teams that already connected their ERP to Geofleet will recognise the pattern: one resource, one set of credentials, one settings page. If you have not done the ERP side yet, our Acumatica last-mile integration guide and the broader guide to connecting an ERP without custom code are the right next reads.