Back to blog

How an MCP server let me talk to my own company

How an MCP server let me talk to my own company

Getting information from my company usually requires knowing where each piece of data lives. If I wanted to find out what had happened over the past week, I had to go through calendars, tasks, sales activities and projects, or ask several people directly. The information was there, but finding it, connecting it and turning it into an answer took time and no small amount of patience. In reality, it was another manifestation of information silos: each application knew part of the story, but none of them could tell the whole thing. This becomes particularly obvious when you come back from vacation and someone asks what happened while you were away. The answer is there, scattered across half the company, but there is no single place where you can simply go and look for it.

RSM Assistant is the interface to RSM AI, the artificial intelligence that governs Redsauce's ERP data.

The situation changed when we connected our systems to artificial intelligence through an MCP server. Now I can ask who did what during the past week, which customers the sales team visited, what problems arose, which ones were resolved, or what actions were taken to prevent an order from arriving late. I can even start with a fairly generic question and let the AI search for the pieces of information it needs to build the answer. The questions are the same ones I used to ask myself in the past. What has changed is how I obtain the answers. I have gone from digging through data to simply asking my company what is going on.

For example, I can ask which orders have encountered problems this week and, from there, ask who was involved in each one, what tasks were carried out, and what was done to prevent the problem from happening again. Answering a question like this requires more than querying a single table or application. Orders, customers, incidents, tasks, projects and people all have to be related to one another. In the past, I was the one who had to understand those relationships and navigate them. Now I can simply ask the question and let the AI use the tools available to it to find the information.

MCP, short for Model Context Protocol, makes precisely this possible: it allows an AI system to discover and use tools provided by external systems. In our case, the server can access projects, tasks, calendars, customers, orders and other entities in our systems in a controlled manner. The AI interprets the question, determines what information it needs and uses those tools to retrieve it.

MCP does not replace our API, nor is it responsible for the intelligence behind the answer. The API remains the structured and controlled way of accessing our systems. MCP provides an interface that allows certain capabilities of that API to be exposed in a way that an AI model can discover and use. Put very simply, the question reaches the model, the model decides what it needs to look up, it uses the MCP tools, and those tools ultimately access our systems through the APIs that were already there.

At Redsauce, we already had a well-designed API, the result of the discipline and stable contract required by an API-first strategy. The data had a known structure and a stable way of being queried. We had to create a token, define permissions and design the tools the AI would use. Without that API, the work would have started much earlier, organising access to the data and explaining how the different pieces of information relate to one another. And this is probably the most important part of the whole thing. Connecting an AI system to a company is not simply a matter of giving it access to a database. You have to decide what it can query, what it can modify, what information must remain off-limits and what operations it is allowed to perform. Allowing it to check the status of an order is not the same as allowing it to modify that order. Nor is showing it a list of customers the same as giving it access to all the information we have about them. The conversation may look simple, but behind it there must be the same level of permissions, security and traceability that we would require from any other application.

I can also ask it to indicate which records it used and how it obtained each result. This is not a minor issue, because a convincing answer is not necessarily a correct one. If the AI tells me that an order was delayed for a particular reason, I want to know where that conclusion came from and what information it used to reach it. It is the same issue that arises when asking who tests AI: it is not enough for the system to provide an answer; we need to be able to judge whether that answer is accurate and well-founded.

Does this mean that MCP servers are the APIs of tomorrow? Not exactly. An MCP server does not replace a good API. In our case, it works precisely because that API already exists. The API organises access to the systems, while MCP exposes their capabilities in a way that an AI system can discover and use. It is quite possible that many APIs will have an MCP server in front of them, just as today they serve websites, mobile applications and internal processes. What does change is how we use those systems. For many years, we have built applications on the assumption that a person would have to learn where everything was: which menu to open, which form to use, which filters to apply and which report contained the information they were looking for. That made sense because the interface was how we interacted with the data. Now we can add another, much more flexible interface, in which people do not need to know in advance where the information they need is stored.

This does not eliminate the need to design systems properly either. For an AI to answer questions about a company, it needs to understand what a customer is, what an order is, how an order relates to a project, who is responsible for a task, and what permissions each person or process has. The form of interaction is new, but the craft behind it remains the same: integrating systems, organising information and building reliable software.

Ultimately, after so many years of creating applications that enable people to navigate data, we have reached a point where the data itself can answer our questions. It is not necessarily about replacing the systems we already have or building another system to take their place. It is about finally being able to ask our own company what is happening without first having to remember where the answer is.

If you want to find out what your own company could answer, tell us where your data lives today and we’ll point you in the right direction.