The setup: 125 stores, 183 POS devices, a four-person team
mobilezone has 125 stores in Switzerland. They rely on 183 point-of-sale (POS) devices, and a back office supports hundreds of B2B customers. Microsoft says the company combined Copilot Studio with Dynamics 365 Commerce and Dynamics 365 Finance. The agents address two different problems.
- Retail: Field Support Engineer Marino Kneringer and three first-level supporters oversee the whole POS estate. They also handle broader hardware and office support.
- B2B back office: Staff re-typed customer order documents into the ERP system by hand. Each order took several minutes.
Chief Digital and Information Officer Mark Schwarz calls agentic AI a strategic capability. He says building expertise in-house with Copilot Studio lets the company scale faster and reduce complexity. That is an executive talking point, not a measured outcome.
Agent one: Store Monitoring
Before
Store staff opened a ticket after something broke, and the support team then hunted for the root cause. Application errors and hardware-station faults were hard to diagnose and often needed escalation across teams. Kneringer said problems surfaced only after they had hit transactions and customers. In his words, that is a business risk in retail.
How it works now
Kneringer adapted a prebuilt Store Monitoring Agent despite limited AI experience, and Microsoft says he deployed it in three weeks. The data flow runs as follows:
- Azure Monitor agents collect selected logs from every POS device and from Azure resources. The POS devices run Dynamics 365 Commerce.
- The Copilot Studio agent processes those signals.
- At 7:00 a.m. each day, the agent posts a Microsoft Teams update. It lists the devices with the most errors, the error categories and the latest occurrence of each issue.
- Support staff ask natural-language questions about a specific device and get AI-generated summaries of the technical detail.
Kneringer can narrow the 24-hour report to a three- to six-hour window and chase unusual patterns. Stores open at 9:00 a.m., so he starts at 7:00 and fixes overnight problems first.
Reported results
- He resolves an average of one to two issues each morning before they hurt the business.
- Support tickets now get an initial response within 30 minutes. Earlier backlogs sometimes delayed a response by three days to a week.
- Microsoft frames the ticket-response change as part of broader support-process improvements. It does not credit the agent alone.
What the template documentation says
Microsoft Learn describes the Store Monitoring Agent as a lightweight, chat-based monitoring agent built with Copilot Studio that reasons over POS telemetry collected via Azure Arc and Azure Monitor. The agent is intentionally read-only, focusing on monitoring, assessment, and triage to minimize operational risk. It surfaces insights through natural language conversations in Microsoft Teams or Microsoft 365 Copilot.
The Learn page lists these prerequisites: Dynamics 365 Commerce Scale Units or POS devices, Azure Arc, an Azure Log Analytics workspace, Copilot Studio, and Copilot Studio credits capacity. It also notes that the agent templates and assets are hosted in Viva Engage Communities. Access is through a request form.
Treat that as template documentation. Microsoft's customer story does not provide mobilezone's exact configuration. It gives no Azure query design, retention policy, hardware models or licensing costs.
Agent two: Sales Order Creation
Fabian Kistler, the AI and Automation Engineer, rebuilt a sales order agent in Copilot Studio. It connects to the Dynamics 365 ERP environment through the Microsoft ERP connector and a Model Context Protocol (MCP) connection. The workflow is:
- An employee drops a customer PDF into the agent in Teams.
- The agent extracts the order details and maps them to the right fields.
- It creates a draft sales order and posts status or warning messages for the back-office group.
- A person reviews the draft and confirms it in Dynamics 365.
The agent works for several minutes while the employee handles email or another task. The human sign-off stays in place, and the early production period pays particular attention to quantity and price. The agent flags missing or unclear information. It can also leave fields empty for review instead of presenting uncertain data as complete.
Kistler reports saving five to ten minutes per contract or order. The first automated customer generates roughly 30 to 40 orders a month. That works out to somewhere around 2.5 to 6.5 hours of data entry saved monthly. This is our arithmetic from the stated figures, not a number Microsoft reports. The company says it will extend the agent to more B2B customers once it validates their document formats.
Agent three: Product Creation
The third agent targets product creation and catalog administration, which Microsoft frames as costly work. The story gives no workflow, integrations or results for it. It says only that the agent is expected to move into production once remaining requirements are done. It is not live, and none of the outcomes above apply to it.
What comes next, and the caveats
Kneringer is testing a next version of the Store Monitoring Agent. Possible enhancements include:
- alert-based monitoring
- connections to the company knowledge base
- automatic ticket creation
- guided remediation
- plain-language summaries for store managers
The company is cautious here. More frequent Azure queries and autonomous actions could change the cost profile. Higher-impact fixes will still need human oversight. That also fits the Learn documentation, which calls the current template read-only, so remediation would be a new capability.
mobilezone is also exploring a way for business groups to submit agent ideas. Each idea would carry an estimate of time savings, cost savings and business value, so leaders can prioritize.
Analysis: what IT admins can take from this
- Start with read-only. The monitoring agent reports and triages and does not change anything. That limits risk, and it is a sensible first step for teams wary of autonomous agents.
- Keep a human at the commit point. The sales agent only drafts. Because the draft is reviewed before confirmation, a misread price or quantity is caught before it posts.
- Budget for telemetry. mobilezone itself flags Azure query frequency as a cost variable. Anyone copying the pattern should model Log Analytics and Copilot Studio credit consumption.
- Read the numbers carefully. Savings of one to two issues a morning and five to ten minutes per order come from the company and Microsoft's marketing story. They are not audited. A 183-device Dynamics 365 Commerce fleet is also a narrow case, so results elsewhere will vary.
- Mind the dependencies. The pattern relies on Azure Arc, Azure Monitor, Teams, Dynamics 365 and MCP connectivity. Organizations outside that stack would need to rebuild much of it.
Overall, the story is a modest, believable example of agentic AI doing triage and data entry instead of replacing staff. Two of the three agents are in production, and the third is still waiting on requirements.