AI with Microsoft 365: from Copilot to working processes

AI with Microsoft 365 only becomes useful when you have a clear picture of one work process: who carries it out, what information goes into it, which systems are involved and where someone needs to review or approve. You can then choose whether a fixed workflow, an agent or custom development fits.

Many organisations do not need to start again. Questions, documents, appointments and actions are already in Teams, Outlook, SharePoint or Excel. Microsoft 365 Copilot can work with information the user already has access to. That makes permissions and information management an important starting point. The question is therefore rarely which Copilot licence you should buy. A better question is which recurring piece of work deserves a better way of working.

Start with a work process

Take one process that currently takes time, needs a lot of attention, or is prone to manual error. For example: a request for a quote that needs to move from a shared inbox to CRM and ERP, or a recurring question about an internal work instruction.

Then describe six things:

1. The owner. Who decides that this process works well and who maintains it after the initial delivery?
2. The input and output. Which email, form, document or question starts the process? What needs to be recorded at the end?
3. The systems. Where is the source information and where does the result need to go?
4. The access boundary. Which employee or role may see which information and perform which action?
5. The fixed steps and those requiring judgement. Which rules stay the same? Where do you need context, interpretation or permission?
6. The checkpoint. What must never happen without an employee?

This description prevents an AI agent from looking for a solution to a problem that actually needs clear routing or a missing agreement. If you need a more thorough framework to work this out, the TWINE framework can help.

A fixed workflow for predictable steps

If the route is largely fixed, a workflow is often the calmest choice. Power Automate is designed to automate steps between applications and services, such as collecting data, sending notifications or synchronising files. That fits a request that needs to be complete, an approval that needs to reach the right place, or a record that may only move to another system after a check.

Design for the exception in a process like this too. If a customer number is missing or a match is uncertain, the task goes to a person. That makes it clear who makes the decision.

An organisational agent for knowledge and actions

Sometimes the whole route is not predictable. You can then combine fixed workflow steps with agent steps: the workflow controls the sequence, rules and approvals, while an agent might assess information from approved sources, interpret a question or prepare a next step. It also works the other way around: an employee can ask an organisational agent for help, and that agent then calls a fixed workflow to retrieve data or perform an allowed action. Copilot Studio is Microsoft’s low-code environment for connecting such agents and workflows to your own knowledge, systems and channels.

An agent like this needs more than a good initial instruction. Decide which sources are valid, which actions are allowed, how users sign in, when the agent must hand over to an employee and who checks the answers. Microsoft describes this as a cycle of planning, building, adoption, management and improvement.

Custom development when the process needs more engineering

A process can need more than a configured agent or flow. Think of your own business logic, a specific integration, your own user experience, tighter runtime control or an application that works outside Microsoft 365. We then look at a custom development approach. For an Azure environment, Microsoft Foundry can be one of the building blocks. The right approach depends on the systems, the risk and the way you want to manage it.

This step also calls for more precise technical choices around integrations, identity, storage, logging and processing. A general promise about where data is stored and processed does not tell you enough. This needs to be designed and assessed for each customer and each work process.

The work only starts after the first version

A process only works in daily practice when people know when they can rely on it, when they need to intervene and who to report a problem to. Plan for a process owner, a small test group, a way to report exceptions and a regular review of use, quality, costs and changes in the source processes from the start.

If your team already works with Microsoft 365, it can make sense to build the solution within that environment too. We then work with familiar systems, permissions and management. If Microsoft does not fit the process well, we choose other technology or a custom approach that suits it better.