How a large retail organization brought cloud self-service into Slack

A support team at a large retail organization we work with was spending its days on the same handful of requests, most of them coming in through a Slack channel or a ticket. These were basic tasks the team didn’t want to spend its time on, such as someone locked out of an account, a password that had expired, or a login that wouldn’t go through. The fixes lived in CloudBolt CMP, but even when a user knew exactly what they needed, the fix could sit a few clicks deep in the UI: find the server, navigate to it, locate the right day-2 action, submit.

This is a large, complex environment. Thousands of team members are spread across the globe, they carry multiple logins with different scoped access depending on the environment, and the team supporting them all is limited to its own working hours.

What they built

Their answer was an assistant in Slack, running on CMP’s MCP server, that lets users ask for exactly what they need. It brings together two things. The first is a catalog of CMP’s own action tools, the operations users could already run through the platform. The second is a knowledge base drawn from the team’s documented procedures, so the assistant can answer how-to questions as well as take the right action. A user describes a problem in the chat, and the assistant either answers it or carries out the fix. Because the bot runs on CMP’s MCP, OAuth2 handles authentication, so every request is federated to the actual user.

How it works, and why it fits the environment

A few choices make it work in an enterprise. The assistant recommends what to do, but the agent, not the language model, executes the action through CMP.

Because the work runs through CMP, the audit trail lands in both the chat and the platform, and anything that would change a production resource asks the user to confirm first.

The assistant is also careful about where it answers. It can be added to any channel or group DM, but when a request involves account details, it escalates to a direct message rather than spilling IDs or access information into a shared channel, then picks the conversation back up privately.

The diagnosis behind “I can’t log in” shows how much reasoning is running under the hood. A login can fail for many reasons:

• The server could be retired
• The network could be down
• The account could be locked
• The password could have expired
• The user could be in the wrong access group

The assistant checks the live state through CMP and answers with the actual cause—telling a user they lack the required group membership, flagging that their password recently changed, or pointing out a secondary account they forgot they had. Just as useful, it rules things out: the server is active, RDP is responding, the network interface is connected. And it omits checks that don’t apply, so users see a diagnosis instead of a checklist.

Quotas and role-based access apply exactly as they do in the platform. A resize request that exceeds quota fails as a normal order, with the attempt and the failure recorded in the audit trail. And when any order fails, the platform opens a ticket for human triage automatically, so there is always a human in the loop when one is needed.

What it changed

Users triage and resolve common issues themselves instead of waiting on the queue. The support team gets back the hours it was losing to those basic tasks. Help is there beyond staffed hours now, when no one used to be. And because the assistant can see the state of a resource, it can suggest actions a user might not have known to ask for, offering to reboot a slow server, clear its temp files, or add memory.

Any CMP customer can stand up the same assistant — try it in the CMP sandbox.

Where this is going

It is a preview of where infrastructure management is heading, people describing what they need in plain language, with a governed platform doing the work underneath. For this organization, CMP is that layer, and the assistant is simply the front door their users prefer. AI makes the interface and the action easy, but the need for governance never goes away, especially in the enterprise.

The team also runs the assistant on a small, inexpensive model. The reasoning that stitches tool results together doesn’t need a frontier model, and they’re still tuning what to run it on to keep costs down.

Try it yourself

Build the same thing in your environment

The MCP interface this team used ships with CMP. Spin up the sandbox and point it at your own catalog and knowledge base.

Open the CMP sandbox