The project is ongoing and we have a great team working on it. Obviously, customer work easily overrides internal development, but we have made good progress. I'm still determined that this shoemaker's children will not go barefoot.
One learning experience has been that the traditional solution-based team structures need to be re-evaluated in the agentic world. Building an agentic ERP requires, in addition to traditional ERP solution consulting and development stuff, know-how about data science, Power Platform, security and local finance processes. Our project is a joint effort between Power Platform, Azure, F&O, Internal IT and Internal Finance teams. It's great to see talented people working together for a common goal.
The team started with creating test agents in Copilot Studio connected to the mock data in the Tier2 test environment. This is quite simple and easy but gives very little insight into what challenges lay ahead when we enter our customized D365 Finance -system. Standard MCP Server can connect and understand standard D365 Finance data and business processes, but I learned that it is not advisable to connect directly to the raw data in custom tables. We might have to create custom services within D365 and let AI agents call those services instead exposing them to raw ERP data. In some cases, Azure functions are also needed. We really need to carefully select our use cases and figure out the most reliable and cost-friendly way to fulfil business needs.
I hope that we succeed, because my dream is that we could agentify our, and our customer´s ERP systems without a mandatory ERP -project. The future might be a headless ERP where data and business logic resides in the ERP and transactions are done either via agents or people using fit for purpose applications.
Maybe the most important achievement so far is the comprehensive security study that was done to address what it actually means to utilize MCP ERP server and AI Agents in production environment. Our architects investigated security controls, authentication, networks and end points, audits, logs, encryptions, monitoring, threats and many more important issues that our internal IT and security teams want to ensure.
Creating agents for testing use is ridiculously easy using natural language prompts in Copilot Studio. All you need to do is explain what kind of Agent you want, choose any language model that calls you name, add ERP MCP server as a tool for the agent and voilà, you have a working agent connected to you ERP system. You can then showcase those agents to the astonished leadership team who can then start making plans for how to transfer half of the staff to more productive work. There will be at least one person in the room who will suggest just taking the agents into use in production so that they can annihilate the competition.
Yeah, nah, I do not recommend that because things are a little bit more complicated than that in real life.
Finding the beneficial use cases is just the very beginning of the Agentic journey. There are myriad important technical decisions to be made that have great effect on cost, security, quality, performance and overall usefulness of Agents, including:
What data to use?
Where to store it in each use case?
Datalake, Blob storage, Cosmos DB, Dataverse, Sharepoint..
These are just some of the things to take into consideration, and each decision will have significant financial impact. Not so simple anymore, right?
The good news about all of this is that we will drink our champagne first and tell you how it went.
Until the next time,
With best regards,
Jaakko Heikkinen
ENT ERP Lead