Stairway to Dataverse Heaven – Governance & Business Skills

In the last blog post, I covered how Copilot Studio really accesses Dataverse through various lenses with the Dataverse Knowledge Source, Dataverse Connector and Dataverse MCP Server. From the way the core model works, to the way inputs and outputs are managed, schema awareness, CRUD support and bound/ unbound actions.

Here is what the summary looked like:

Now, I could not help myself but expand on how you can supercharge your Dynamics 365/ Power Apps solution even further by using tools like the MCP Server (not just Dataverse obviously but also with the Power Apps MCP Server to solidify your app’s logic and context provision).

Since the announcement of business skills, I have been pondering a lot over the best starter use cases for the enterprise companies and solutions. The concept alone is fantastic; it is now possible to “teach” the MCP Server to discover which business data is actually used into which processes, under what context and operational scenarios. Sure, someone may be using the sales order table like many others, but the agent may not know Dynamics 365 is also integrated with SAP to manage invoices and they need a specific format to always be in place. Or that the Finance team uses American accounting standards which impact data flow.

All this context is gold and now the Dataverse MCP Server can:

• discover and leverage business skills

• help the right users update business skills through direct interactions which is a LOT more accessible than going the back-end way

It also helps that business skills are solution aware and you can manage their deployment appropriately through your environment strategy.

Ultimately, this provides the agents using the Dataverse MCP Server more grounding and predictability.

As you start using business skills, you will realize that the agent is not only much more insightful when they are available, but also is cognizant of their absence and will prompt you to create them.

See below an example when I tried to query my oldie-but-goodie Avenge 365 app for extra budget for armour spend for the Avengers.

This was a classic scenario of an original quote being exceeded by actual spend.

In this case I did not have the right skill for my query, and the agent knew! It searched through the database, found another skill and it was not the right match.

In fact, it did not stop there. It suggested to help me create a new skill. This is a prime example of how agents can become more proactive AI teammates. Business skills are just one of the many facets of it.

And remember, to set this up you first need to enable it for your environment under Settings > Features and ticking Dataverse Intelligence. Ironically, this one is not a preview feature on by default!

Then, you also have to enable the Preview version of the Dataverse MCP Server which allow you to use the business skills.

This was a stairway test, and you passed! It is actually a brilliant question as you see more and more tools available.

Secure, safe and conscious access to the tools in MCP Server, like any other, is key.

In theory, we are all lines of defense. From end users, to makers and admins. For the last two groups, there are a few options to ensure only the right people perform sensitive operations (delete being an obvious suspect but others too:

1. As an admin, you have a clear RBAC process and your access packages limit CRUD operations appropriately in terms of which tables can be accessed and what way for each. These security roles will ALWAYS be honored with agent end-user authentication even if the tools are in theory available from an admin or maker perspective. It is your omnipresent frontier of defense, but also the last frontier even if tools are seemingly available. Of course, you also have to think of Advanced Connector Policies and how you can block ex ante the permissive tools you do not even want the agents to access. This could include blocking the whole of an MCP Server, or just some of its tools.

2. As a maker, you toggle the Dataverse MCP Server’s tools on and off according to who the end user and their role is e.g. table-specific tools might be turned off as end users should not be performing schema operations etc. Of course, this is on top of using end-user authentication so the agents only gives access to the end-users actual permitted data to access. (There are always exceptions but this is the wider approach)

I don’t know if Led Zeppelin would build agents. But if they did, they would write songs about how these architecture patterns work when used intentionally!

If you wanted a brief summary of how these tools work to explain to your colleague, friend or family, here is how you could approach it:

• Very predictable scenarios without a need for an LLM to reason or orchestrate a series of events.

• The main focus is repeatability and consistency in the outcome, hence pre-configuring all the operational logic paths and configuring the related parameters (entities/columns used) makes sense!

• Trust in the processes, service and brand reputation is directly connected to these outcomes, so for lack of a more poetic metaphor you cannot leave things to chance! For example, imagine if the agent decided to give x3 the approved loan amount to a low credit score customer because they asked nicely. Not a good outcome for the bank!

• This is very different territory because the path and outcomes aren’t always repeatable or consistent.

• A more flexible orchestration layer will look into fulfilling an objective by “self-orchestrating” the path to get there (tool selection, what data to include etc.).

• Ironically, this also connects to how unpredictable user requests are. The richness of language expressions, people’s unique tone of voice and preferences are all factors lending themselves well to conversational agents with generative orchestration. For example, an agent helping couples prioritize the top destinations for a honeymoon can use generative orchestration to decide the right deals based on the unique request put through; from the top priorities for leisure activities to budget involved, risk appetite for adventurous destinations etc. The output is as unique as the users themselves!