Copilot Studio in document workflows – how an AI assistant helps approve invoices, contracts and requests

An electronic workflow does not in itself shorten decision time. The document reaches the right person faster, but they still have to read it, check the context and judge whether anything is out of the ordinary. An AI assistant takes over the preparation for that decision.

Take a hypothetical distribution company with 150 employees. The purchasing manager approves several dozen invoices and purchase orders a week, often on his phone between meetings. The list below shows how the assistant helps him.

The most common design mistake is trying to make the assistant do “everything at once”: archive, approval engine and chat window. Such a solution is hard to maintain, and every change to the process means rebuilding the assistant.

It is better to split the responsibility across four elements that already exist in Microsoft 365. The table below shows them using the example of approving a purchase invoice. With this split, a second and third assistant can use the same foundation.

The assistant answers on the basis of the sources you point it to: SharePoint libraries, lists with process data, selected intranet pages. The most important decision is the scope. An invoice assistant does not need access to personnel files, and a legal assistant should not read invoices.

The narrower the scope, the more accurate the answers and the lower the risk of someone seeing data they should not. Connecting “the whole of SharePoint” to the assistant is a sure way to get answers based on outdated versions of procedures.

The second rule: every answer should include a link to the source document. The approver can check with one click where the assistant got the information. This is also the best protection against the AI model making things up, because any error is immediately visible.

The assistant can not only answer questions but also trigger actions: a Power Automate flow, an update to a document’s metadata, fetching an order status from the ERP system. This is where the most important security boundary lies.

We divide actions into two groups. Informational actions, such as “show the order status”, can be made available from the start. Actions that change data, such as “reject the invoice” or “forward to the director”, require the additional safeguards described below.

The best places to start are documents that are numerous, repetitive and assessed against similar criteria. The examples below are illustrative – in each company the scope and rules have to be agreed with the process owner.

With invoices, remember KSeF, Poland’s National e-Invoicing System. Most invoices from businesses now arrive as structured data, so the assistant does not need to read scans. Reading the image of a document is still needed mainly for foreign invoices, receipts and documents outside KSeF.

Copilot Studio is not licensed per person. The company buys credits: a pack of 25,000 credits for USD 200 a month, or pay-as-you-go for actual use through an Azure subscription. Each answer and action by the assistant uses a number of credits that depends on how complex the task is.

When someone with a Microsoft 365 Copilot licence uses the assistant in Teams or SharePoint, some answers based on company data do not consume credits. Power Automate flows triggered outside the assistant may require separate licences, so check this before you start.

Beyond licences, the cost consists of design work, connections to systems (e.g. ERP) and your own people’s time to describe the rules and test. Microsoft provides a credit usage calculator. It makes sense to estimate usage based on a pilot rather than in advance.

The first assistant is usually a project of a single department. Problems start with the third and fourth, when nobody knows who is responsible for which one and what data it uses. That is why it is worth setting rules from the start, even if they seem excessive.

Choose one type of document with a large number of approvals and one group of approvers. Before you start, measure three things: the average time from receipt to decision, the number of documents sent back for clarification and the number of errors found after approval. Then work through the steps in the list below.

A simple estimate of the benefit: number of documents per month × minutes saved per document. For example, 400 invoices × 3 minutes = 1,200 minutes, or 20 hours of managers’ time a month. Plug in your own data and check after two months whether the assumption holds. Our case study from a manufacturing plant shows how a knowledge assistant built on Copilot Studio works in practice.

An assistant will not fix a process that does not exist. If nobody knows who approves invoices above a certain amount, you first need to set the rules and build an ordinary workflow. Only then is it worth adding AI.

It also makes no sense when there are few documents, e.g. a dozen or so contracts a year. In that case the set-up cost will outweigh the benefit, and a good document template and a checklist will work better. The same applies to decisions that depend mainly on judging the relationship with the other party rather than on analysing a document.