AI agents in SharePoint: helping employees find company documents and knowledge faster

Take a hypothetical engineering company with 150 employees. A salesperson needs the current product sheet for a customer. SharePoint has three files with similar names, from different years, on three different sites. Search shows all of them, so the salesperson writes to the technical department and waits until tomorrow for an answer.

This is not a rare case. A 2012 McKinsey Global Institute report estimated that office workers spend nearly 20% of their working week searching for internal information or for colleagues who can help. The figure is over a decade old and is an estimate, but anyone who has asked a colleague “where is the latest version?” knows the problem.

Traditional search looks for words. If a policy is called “Rules for settling business travel expenses” and an employee types “business trip”, they may not find it. On top of that, search cannot tell the version in force from a draft from three years ago.

An AI agent is an assistant you ask a question in an ordinary sentence, for example “what is the limit for overnight accommodation on a domestic business trip?”. The agent finds the matching documents, reads them and answers in a single paragraph with a link to the source. It can combine information from several files, for example from the policy and from the latest management directive.

The most important security rule: an agent in Microsoft 365 sees only what the person asking has at least read access to. The production manager will not get the content of board contracts from the agent if they do not have access to them. But the same rule also works the other way round, as we explain below.

An agent can also make mistakes or misunderstand a question. That is why a well-configured agent always gives its source, and for important matters the employee should click the link and check the document.

Microsoft offers three routes, which are often confused. They differ in who uses them, how widely they search and what they cost. We compare them in the table below.

In practice, many companies combine two of them. Managers with a Microsoft 365 Copilot licence use Copilot chat and agents in SharePoint. For other employees, a single agent in Teams is built in Copilot Studio, limited to selected sites.

In many companies, SharePoint permissions have built up over years. At some point someone shared a folder with “everyone in the organisation” because it was quicker. As long as you had to know where to look, nobody noticed.

An agent changes this. If a payroll file sits in a folder open to everyone, the answer to “how much does the warehouse manager earn?” is no longer a theoretical question. The agent does not break security – it simply uses the permissions someone granted long ago.

The good news: companies with at least one Microsoft 365 Copilot licence get SharePoint Advanced Management tools. These let you check and limit overly broad sharing before the agent reaches people:

An agent is only as good as the documents it reads. If the current price list and three old versions sit side by side, the agent may quote any of them. Preparing the knowledge is usually the longest part of the project, but the order it brings will also help people.

The steps below are worth doing for the area you are starting with, not for the whole company at once:

Metadata is extra information about a file, stored in library columns: the counterparty, the contract end date, the document type. It makes filtering and setting reminders easier. The problem is that nobody likes filling it in by hand.

SharePoint has a feature for automatically filled columns (autofill columns). You assign an instruction to a column, for example “give the contract end date”, and AI reads it from each new file. The service is billed by usage and supports the languages available in Copilot.

Take a library of supplier contracts. The “counterparty”, “value” and “end date” columns fill themselves in, and Power Automate sends a reminder 60 days before the contract ends. Microsoft does, however, recommend checking the results, using no more than 10 such columns per library and files of up to 65 pages. Bulk processing of files already in the library is still to come.

An agent built in Copilot Studio can carry out a simple action after answering. For example, a salesperson asks about discount terms, and the agent gives the rule and immediately offers to send a discount request to the manager for approval. The approval request reaches Teams through Power Automate.

This is useful, but it widens the project: you have to decide what actions the agent may take, on whose behalf and with what record. It is sensible to start with an agent that only answers, and add actions after a few weeks of operation.

Choose one area where people often ask experts and where the documents have an owner. Good examples are quality procedures, product information for salespeople or internal regulations. Before you start, count how many questions a week go to the experts in this area.

After a few weeks, check three things: whether the number of questions to experts has fallen, how many answers from the test list the agent gets right and what users say. If the agent keeps getting a particular topic wrong, the document is usually to blame, not the model.

If a company has a few hundred documents and two people who know where everything is, a well-designed folder structure and a home page with links will achieve more than an agent. The same applies when the most important documents are on a network drive or in email – they first need to be moved to SharePoint.

An agent also makes no sense when nobody wants to be responsible for keeping documents up to date. Within a few weeks, employees will notice wrong answers and stop using it. Rebuilding trust in a tool is harder than launching it.