What Copilot Studio can and cannot see in your business

What it can see

An agent answers mainly from the knowledge you give it and the systems you connect to it. Microsoft documents these ways to give an agent knowledge (Microsoft Learn):

• public websites you name, searched for the agent;
• SharePoint sites and files;
• files you upload, from your computer, OneDrive or SharePoint (Microsoft Learn);
• tables in Dataverse, Microsoft's store for structured business data;
• other systems through connectors, which let an agent look things up in, or act on, the tools you run (Microsoft Learn).

Depending on its settings, an agent can also draw on the AI model's general knowledge or the public web, so someone should check what is switched on.

Permissions can travel with it. When user sign-in is set up, an agent answers each person only from the SharePoint, Dataverse and connected knowledge that person is allowed to see, and files brought in from SharePoint or OneDrive are checked against each person's access when they ask. Public websites and files uploaded from a computer have no such check, so anything you upload that way can show up in answers for everyone who uses the agent. Connectors used as tools can be set to act with the builder's own access instead, so ask whose access the agent is using.

Copied, or read live?

How a source is added changes how fresh its answers are (Microsoft Learn):

• SharePoint files added through the file upload option are copied into Dataverse and refreshed every four to six hours.
• SharePoint added through its own connector is read where it sits, in real time, as far as SharePoint's search has indexed it.
• Business systems added as real-time knowledge (Salesforce, Dynamics 365, SAP, Oracle Database and others) are queried when the question is asked, with only their table and column names indexed in advance. Microsoft marks this option as a preview, not meant for production use yet (Microsoft Learn).

So an agent asked about this morning's orders may be reading this morning's orders, or a copy from several hours ago. Choose the option to suit the question.

What it cannot see

Everything above is something an agent can be connected to. The problems below are things nobody has connected. None of them is specific to Copilot Studio; any AI tool has the same blind spots.

• What was never connected. The spreadsheet on the finance director's laptop, the price list in someone's inbox, the orders taken by phone and written down later.
• What was entered late, half filled in, or not at all. An agent reads records; it cannot know that a payment has arrived but has not been matched yet.
• Which of two numbers is right. If sales and finance hold different figures for the same month, the agent will answer from whichever it reaches, with the same confidence either way.
• What your words mean. Revenue before or after credit notes? An active customer in the last 30 days or the last 90? If nobody has written it down, the agent cannot follow it.
• History that systems overwrite. Many tools keep only the current state. Ask how a customer's balance moved over a year and the answer may not exist anywhere.

An example

An illustrative company, not a client.

A distribution company of 50 people builds an agent for its managers. The first question someone asks it: "Which customers are paying us late, and how much is at risk this month?"

Invoices sit in the accounts system, which the agent can reach. Credit notes are kept in a spreadsheet, because that is how the team has always done it. A large customer paid by bank transfer on Friday, but the payment will not be matched to its invoice until Monday. The agent gives a clear, well-written answer, built on the accounts system alone. It overstates some balances, lists a customer who has already paid, and nobody in the meeting can tell.

The same agent, pointed at one checked table of customers, invoices, credit notes and payments that every system feeds, gives an answer finance can check against its own records. Nothing about the agent changed. What it could see did.

When it is enough, and when it needs more under it

Before you switch an agent on

1. Match each question to a source, and to how fresh the answer must be: a copy refreshed every few hours, or a live read.
2. Set up sign-in and check whose access the agent uses, and whether general knowledge or web search is switched on. Someone who looks after your Microsoft 365 settings will need to help.
3. Test it with the questions your managers really ask, against answers you already know to be right. Where it disagrees with finance, the problem is usually upstream, in the data.

If you are still deciding where AI should start in your company, our guide on where to start with AI in a company of 50 people comes first.

Where we come in

When the questions turn to numbers that span systems, an agent needs one checked layer of data under it. That is what we build. We connect your whole business and put AI to work making sure every entry goes in properly, so Copilot Studio, your reports and your team read the same numbers. Your systems stay where they are; we add the layer that finally makes them speak to each other, and the command centre your leaders read every morning.

Common questions

Checked against Microsoft's documentation on 28 October 2026. Microsoft updates it often.