Moving from me to us is about ownership. Here is what separates the agents the team uses from those only I use.
The guide to building your first AI agent in under an hour now has over 2,500 comments on LinkedIn. It shows how to build the Document Agent in Claude, ChatGPT or the Copilot you already have.
Then comes the hardest question: how does it go from working for you to being used by your colleagues?
I have wrestled with that question for a long time. I have built hundreds of agents in recent years. A couple of dozen now work for an entire team. The rest work for me. Even the best do not yet run quite as I would like.
In the previous edition, I asked whether the next step should be more cases by industry or my method. Those who replied wrote METHOD. Here is the first part: how an agent moves from me to us.
1. A personal agent and a shared agent are two different products
You forgive your own agent. If it answers badly, you adjust the prompt. If the document is outdated, you know. If it forgets something, you remember for it.
A colleague does none of that. She tries the agent twice. If it answers incorrectly the second time, she never opens it again.
From me to us
Why a shared agent has different requirements.
| Figure | Meaning | Context |
|---|---|---|
| Hundreds | agents built in recent years | Stefano's own agents |
| A couple of dozen | now work for an entire team | the rest work only for him |
| 2 | attempts a colleague gives the agent before abandoning it | illustrative example |
A personal agent is forgiven. A shared agent is not. Moving from me to us is mostly a leap in ownership.
Model 1. The distance between an agent that works for you and one the team uses.
In the Copilot guide, we put the employee handbook directly into the agent's instructions. That works fine as long as it is your agent. When the handbook changes, someone must update the instructions, and that someone is you. Every time.
Principle
Moving from me to us is therefore mostly a leap in ownership. There must be an owner, a source that stays updated, access that follows the user, and someone who notices when the agent breaks.
2. Successful agents live where the team already works
The most successful agents I have been close to were built by developers themselves and shared with each other. They help ship code and connect to Jira and Confluence, so tickets and documentation update automatically.
When I look at why these took off, I see three things.
1. The work itself wakes them up. Nobody needs to remember to open them. When the code moves, the ticket moves too.
2. The result lands in a system everyone already checks. The whole team can see and correct an updated ticket. Nobody can do that with an answer in a private chat.
3. The team has agreed what finished means. Developers agreed this long before the agent arrived. They have spent years documenting when a task is done, and the agent inherits that agreement.
The last point is what most other teams lack. An agent cannot invent a workflow the team has not agreed on itself.
3. Five things break on the way from me to us
This is the list I go through with customers when an agent needs to be shared. Number one surprises almost everyone.
1. The connections are yours. In Copilot Studio, an agent's connections to external services are tied to an individual user's identity. Microsoft itself writes that this explains why an agent may work for you but not your colleagues. Sharing the agent does not automatically share the Power Automate flows it uses either. The solution is typically environment-level connections or a technical account (service principal), and that means a conversation with IT.
2. Knowledge and permissions do not come along automatically. When you share an agent from Agent Builder in Copilot, the SharePoint files and folders you added are shared only if you have permission to share them. If that fails, the agent is shared anyway, just without the files. When you update the agent's knowledge, you must share it again. And removing a colleague's access to the agent does not remove her access to the files.
3. Nobody owns it. Who in the business answers when it makes a mistake? If it is you because you built it, the agent stalls when you get busy. If it is IT, it is the wrong name. The owner is the person whose process the agent works in.
4. Nobody measures it. Use the same three figures as in the previous edition's cases: value in one unit, a quality threshold that must not deteriorate, and how much of the work actually runs through the agent. Without them, nobody can defend it when budgets are allocated.
5. Nobody knows it exists, and nobody can switch it off. When everyone can build, numbers grow faster than oversight. Who knows where your agents are? Who presses stop when one does something it should not?
The three figures
Value in one unit. A quality threshold that must not deteriorate. And how much of the work actually runs through the agent.
4. The platform makes the leap possible, within your own walls
This is my daily work. At the companies I advise, I typically set up the agent platform on Microsoft: Agent 365 and Power Platform, with Copilot Studio and Agent Builder as the tools.
Honestly, I think it is one of the most exciting things happening in companies right now. People who have never built anything digital can build an agent themselves, behind the company's firewall, using data they already have access to. Digital tools used to be something IT or a vendor made for you. Now you can make them yourself.
The platform is what lets that happen without chaos.
Environments in Power Platform separate where the agent is built from where it is used. Data policies determine which connections an agent may use. If a connection is blocked, the feature is greyed out for the builder, and the agent cannot be published until it is fixed. The rules sit inside the tool itself, and nobody needs to remember a policy.
Microsoft Foundry provides access to over 10,000 models, including Anthropic's Claude. An agent can therefore move to another model when prices or needs change. I always build so the customer can switch.
Agent 365, generally available since May 1, collects agents in one registry, with access and security through Entra, Purview and Defender. An administrator can block an agent across the organisation or assign a new owner. That answers point five: where are they, and who can switch them off?
Four building blocks in the platform
What each part solves on the way from me to us.
| Building block | What it solves | Effect |
|---|---|---|
| Environments | Separate building from use | Power Platform environments separate where the agent is built from where it is used. |
| Data policies | The rules sit in the tool | A blocked connection is greyed out for the builder, and the agent cannot be published until it is fixed. |
| Microsoft Foundry | Over 10,000 models | Including Anthropic's Claude. An agent can move to another model when prices or needs change. |
| Agent 365 | One registry for agents | Access and security through Entra, Purview and Defender. An administrator can block an agent or assign a new owner. |
You must take ownership yourselves. The platform makes ownership possible.
5. Build for the top line, and the agent gains supporters
Which agents make it from me to us? In my experience, it depends less on the technology than on what the agent is meant to do.
An agent that saves a few hours has no defenders. Its builder is happy. The rest of the team sees another thing to learn and a question about who will then do what.
An agent that helps win customers gains champions. Faster quotations, better customer answers, sharper marketing, better products. People want to participate when it is about growth.
Principle
That is why I call chasing savings a loser's game. You realise a saving once. If the company grows, nobody needs to fear for their job, and the desire to build spreads by itself.
6. From me to us in five steps
If you have an agent that works for you, do this in the coming week.
1. Write down the owner's name: the person in the business whose process the agent works in. If you cannot name someone, the agent is not ready to share.
2. Move knowledge out of the instructions into a shared source with permissions, typically a SharePoint folder the owner keeps updated.
3. Ask IT whether the agent's connections run on your account. If they do, have them moved before sharing.
4. Share with three colleagues before the whole organisation. Give them the three questions they ask most often, and agree the three figures you will review after 30 days.
5. Register the agent where your agents are collected, and write down who may switch it off.
I have collected the five steps in a one-page Team-Agent Canvas, a continuation of the canvas many of you used for your first agent. Write TEAM in the comments under my LinkedIn post and I will send it to you.
A question for you
Which of your agents would a colleague miss tomorrow if you switched it off today?
If you cannot name one, you know where next week's work lies.
/Stefano
Sources
Product features and licences are described based on Microsoft's own documentation:
- Microsoft Learn, Share agents with other users (Copilot Studio): learn.microsoft.com/microsoft-copilot-studio/admin-share-bots
- Microsoft Learn, Share and manage agents (Agent Builder): learn.microsoft.com/microsoft-365/copilot/extensibility/agent-builder-share-manage-agents
- Microsoft Learn, Data policies for Copilot Studio agents: learn.microsoft.com/microsoft-copilot-studio/admin-data-loss-prevention
- Microsoft Learn, Foundry Models overview: learn.microsoft.com/azure/ai-foundry/concepts/foundry-models-overview
- Microsoft Learn, Overview of Microsoft Agent 365: learn.microsoft.com/microsoft-agent-365/overview
- Microsoft Learn, Governance and lifecycle actions for agents (Microsoft 365 admin center): learn.microsoft.com/microsoft-365/admin/manage/agent-actions
- Microsoft Learn, Tenant environment strategy (Power Platform): learn.microsoft.com/power-platform/guidance/adoption/environment-strategy
- Stefano Vincenti, Build your first AI agent in under an hour (Sep. 2026), Build an AI agent in the Copilot you already have (Oct. 2026), Where AI actually makes money (Oct. 2026): stefanovincenti.substack.com
© 2026 Stefano Vincenti · aitrainer.dk · Status as of October 2026.
This is general practical guidance, not legal or compliance advice, and I am not a lawyer or compliance specialist. Do not use it as the sole basis for decisions. Always consult your own compliance or legal department and your DPO, and confirm dates and status against primary sources. Product features and licences change frequently; always check Microsoft's own documentation. Status as of October 2026.