
AI IN PRACTICE
Every application you commission today needs an MCP
If you're having an application built today, whether fully custom or on the Power Platform, there's one question many clients still don't ask: does it have an MCP? Without one, an AI agent can't work for you inside that application. If you want to build for the future, it's non-negotiable.
On this page
▾
Want an application that's ready for AI?
Tell us what you need. Together we'll look at what the application has to do, including for AI agents, and build that into the project from the start.
Discuss your projectIn short
Every new application should have an MCP, even if it's fully custom-built. It's non-negotiable.
An MCP is more than a way to read data for reports. It lets an AI agent do real work in your application.
Put that requirement in your project brief from the start. Adding it later is more expensive and slower.
An agent can't rely on the screen. The checks have to happen on the server side, so they always apply.
Building an application with AI is lightning fast these days. But the people who use your application will soon be joined by AI agents that look up data, create cases, and set up tasks on their own. An application that doesn't allow for that will be outdated by tomorrow. That's why there's one requirement we put on the table for every new application: it needs an MCP.
What is an MCP?
MCP stands for Model Context Protocol. It's a standard way for an AI assistant or agent to carry out a task for you inside other applications. The application tells the AI what it contains and what it's allowed to do with it: look up data, create a quote, move a case forward.
When people think of "AI and data," many of them think of reporting: pulling data and summarizing it. That's possible, but it's the smallest part. What really matters is that the AI can work for you. That it can prepare a task in your application, fill in missing data, or kick off a process, within the limits you set. Without an MCP, that's impossible, or only possible with an expensive, fragile integration that has to be rebuilt for every assistant.
Why it's non-negotiable
Whether your application is fully custom-built or runs on the Power Platform, the question is the same. If you have an application built today, you'll still be using it five years from now. In those five years, AI agents will become ordinary coworkers. An application without an MCP will be an island: your employees still work in it, but your agents can't get to it.
On the Microsoft platform, that starting point is already there. Dataverse, the data source underneath Power Apps and Dynamics 365, has its own MCP server that's generally available. With a fully custom application, you build that layer yourself, and you can and should do it from the start. It costs far less when you include it in the design than when you have to add it later.
What to ask for as a client
Is there an application on your roadmap? Then state in your request or tender, right from the start, that the vendor must be able to work with AI. Specifically:
The application offers an MCP with clear actions like "create a quote" or "mark this case as completed," not just raw data.
An agent works with the permissions of the user it's working for, not with an account that can access everything.
You can see afterward what an agent did, when, and for whom.
Sensitive steps, like changing a price or sending something to a customer, can be approved by an employee first.
A vendor who has no answer to this is building an application for yesterday.
Server-side checks: the guard at the back door
This is where quickly built applications fail most often, and it's the most important thing to understand, even if you're not technical.
Say your office has a sign on the front door: "by appointment only." That works as long as everyone comes in through the front door. An AI agent often uses a different entrance. It doesn't read the sign, because it doesn't use your screen. If your rules only live on the screen, for example that a quote must always have a customer or that a discount can never exceed 15 percent, those rules don't exist for the agent.
The solution is a guard at the back door: server-side validation. That means the checks happen in the application itself, where the data is stored, and not in the form you use to enter it. Whoever or whatever comes in, an employee, an integration, or an agent, the rules always apply. With a new application, this has to be built in from the start, not added after something goes wrong.
An app for people isn't automatically an app for agents
A person sees a screen, reads the labels, and understands from context what's meant. An agent sees your data, the names of your tables and fields, and the actions your application offers. That's why three things matter:
A data model that makes sense. Clear names, relationships that are correct, and fields with a description. If a field is called field7, the agent is guessing.
Actions that carry out an entire process correctly. Not individual fields, but steps with clear input and an understandable error message when something isn't allowed.
The user's permissions. On the Microsoft platform, an agent can sign in through Entra ID with the permissions of the person it's working for. That only works if those permissions are actually set up.
Five questions to test your own application
Does it have an MCP, or an equivalent layer for AI agents?
Do your most important rules still apply when someone creates a record directly, without using the form?
Does it offer actions for a complete process, or just individual fields?
Does every integration work with the user's permissions?
Can you see afterward who changed what, whether a person or an agent?
If the answer to any of these questions is no, that's the work that needs to happen first.
Keep an eye on the costs, too
An agent that uses your data isn't always free. Microsoft charges for the use of the Dataverse MCP tools when agents built outside Copilot Studio call them, with exceptions for certain licenses. Factor that in when you decide where the agent runs.
Kai for Dataverse, our own product that uses our MCP, takes care of that. You pay a low, fixed price, and if you have a support contract with us, it's included, so it's free.
How we approach it
We define the MCP, the server-side validation, and the permissions before the first line of code is written, whether we're building fully custom or on the Power Platform. That makes an application maintainable today and usable by an agent tomorrow.
For anyone who wants to get started with their own MCP server, we've published an open-source template in .NET with Entra ID. You can read more about it in our article on MCP with Dynamics 365 and Power Apps. To see what an agent can actually do for your team, check out our AI & Agents page. And to see how we build applications, go to Custom Development & Power Apps.
Soon, anyone can build fast
Building with AI makes an application faster and cheaper to build. That's good news. But whether it's future-proof isn't something you decide with a prompt. That comes down to the MCP, the server-side validation, the permissions, and the controls underneath. Soon anyone will be able to vibe code. We make sure it also scales, stays secure, and keeps working.
Want an application that's ready for AI?
Tell us what you need. Together we'll look at what the application has to do, including for AI agents, and build that into the project from the start.
Discuss your projectThis article is part of
AI and Agents
Put AI to work for you: a digital coworker that carries out tasks in your Microsoft apps, follows up, and talks to your data.
Explore AI and AgentsKeep reading

CRM AND ERP
A CRM alongside your ERP: why you don't have to keep track of anything twice
You've been working in an ERP like Business Central for years. Adding a CRM sounds like a second system where you keep track of everything twice. It doesn't have to be: agree on which system manages each piece of data, and each system does its own part.

AI IN PRACTICE
Beyond digitization: working with an AI assistant on top of your business Applications
A customer email comes in. Five minutes later, the project is set up in Connect, our in-house application, and colleagues can log hours on it. Add the tasks in Jira and it takes fifteen minutes. Meanwhile, I was drinking my coffee. That's how we've been working at Profility for a year now. This is the step after digitization.

POWER APPS IN 2026
Are canvas apps dead? AI changes the trade-off
For years, writing less code was the case for canvas apps. But what if AI writes that code? Are canvas apps dead? No. But choosing canvas by default is.
