WonderLaura @iwmentor – Laura Rogers

For this post and video, I put together ten cool things you can do with AI Builder in Power Automate. I built demos for almost all of them.

🧾 1. Extract data from invoices

Microsoft already knows what an invoice looks like. Because of that, there is a whole prebuilt action called Process invoices. Your trigger can be anything. For example, a file landing in a SharePoint library or an email attachment arriving in your inbox.

In my demo, I did this:

1. When a file is created in my AP Invoices folder
2. Get file content
3. Process invoices, pointing at that file content

With that action, you’ll get over 200 pieces of data. Amount due, vendor address, customer info, tax, line items, all of it.

Here is the part people miss. Almost every field also has a confidence value. My invoice total came back at 0.535, so roughly 54% confident. That is a number you can act on.

💡 Fun trick: I had Copilot generate a fake bicycle parts invoice to test with. It works great for demos.

📇 2. Capture business cards

This one is a prebuilt model too. It pulls name, title, company, phone, and email off a photo of a card.

Go to the AI Hub, choose the business card model, then click Use in a flow. It hands you a ready made template. Alternatively, use it in a Power App so you can snap cards with your phone at a conference.

Then write the results into a SharePoint list called Contacts, send in an email or add to a database. That is your whole conference follow up problem solved. In this example, it’s a manually triggered flow where the image is an input control.

😀 3. Score sentiment

Sentiment analysis scores text as positive, negative, neutral, or mixed. Think surveys, help desk tickets, reviews, and form replies.

I built mine to trigger when I flag an email. I did not want it analyzing every single message that arrives.

You get an overall sentiment, and you can also get a score per sentence. Next, do something useful with it. Post a Teams message on negatives, or add a Sentiment column to your survey results list.

In my example, I converted the email body to text with the “HTML to text” action. Since the analyze sentiment action needs to be provided less than 5000 characters, I did a compose action to get a substring of the email’s text. Then, you provide that to the Analyze sentiment action, and the sentiment is an output.

🗂️ 4. Classify into categories

The Classify into standard categories action uses a fixed list: issues, compliment, customer service, documentation, price and billing, and staff.

If those do not fit your business, build custom categories in the AI Hub instead.

In my demo, that T-Mobile advertisement email came back as customer service. That actually makes sense. After the classification, you could add a Switch action and route each category down its own path.

I also added Generate key phrases to the same flow. It pulled out things like “five year price guarantee” automatically.

📸 5. Detect objects and describe images

There are three image actions worth knowing: detect objects in images, describe images, and recognize text in an image.

Describe images works out of the box. Detect objects requires a custom model, which means uploading and tagging your own product photos first. This example is a manually triggered flow where the image is uploaded as an input parameter.

I tested the Generate description of an image action with a picture of a Bluetooth speaker. It confidently told me it was a black coffee mug. Live demos keep you humble.

🧩 6. Start with the templates

This is my biggest recommendation. Search the templates for AI Builder before you start building actions from scratch.

I picked a random one that analyzes tech articles from an RSS feed and posts results to Teams. It already had a feed built in. All I had to supply was my team and channel.

It ran and posted key phrases, organizations, and summaries for ten articles. Meanwhile, I learned more from reading that template than from an hour of trial and error.

✅ 7. Combine extraction with approvals

Here is where this gets real. Check the confidence level in a condition.

If confidence is high, let the flow proceed automatically. However, if it is low, send it to a human through an approval. That way, a person reviews only the questionable ones. That is AI analysis and human judgment working together.

📄 8. Build a custom document model

For custom documents (beyond just invoices), go to the AI Hub and upload around 25 to 30 sample files. Then map every field on each one. It takes time, but afterward it recognizes your specific documents every time.

I used this for a client with veterinary lab results. Those documents arrived from many different vets in many different layouts, and it still worked when I spent a lot of time uploading many different example files and training AI.

🔍 9. Recognize text with OCR

The Recognize text in an image action reads handwriting, labels, signage, and old scans. Despite the word “image” in the name, it also handles PDFs. It returns the text plus position data. As a result, you can slice regions or feed the output straight into a classification step.

💬 10. Run a prompt

The Run a prompt action gives you AI Classify, AI Extract, AI Reply, AI Sentiment, and AI Summarize. You can also author your own prompt and test it in Prompt Builder.

Here is my favorite build from the whole episode. A drop library where files sort themselves:

1. When a file is created in a library that I named “File Drop Library”
2. Get file content
3. Recognize text in an image to pull the full text out of the PDF
4. Run a prompt with AI Summarize on that text
5. Classify into standard categories
6. Switch on the category, then Move file to the right SharePoint site, which could be different for each category.

People drop files in one place. Then the files disappear and land where they belong. Send the person an email so they know where it went. This action Recognize text in image or document works on images and PDFs.

⚠️ The tip that saves the most troubleshooting

Read the action’s input label carefully. If it says invoice file or image, you can pass file content. If it says text, you cannot.

Text inputs cap out at 5,000 characters. Therefore, I add two steps in front of them every time:

1. HTML to text on the email body or document
2. Compose with a substring() grabbing the first 4,999 characters

That one pattern fixes most of the errors you will hit while using AI Builder in Power Automate.

💰 Watch your credits

AI Builder in Power Automate consumes AI Builder credits. Check your usage in the Power Platform Admin Center under Licensing, then Capacity add-ons.

Yes, there is a cost. But compare it against the hours your team currently spends typing invoice numbers and filling in metadata by hand.

🚀 Where to start with AI Builder in Power Automate

Start with a prebuilt model and a real document you already handle. Turn on the approval setting. Only go custom once the prebuilt path has proven its value. Then test the heck out of it.

Here is the associated demo video:

🎓 Ready to build this yourself?

Power Hour — Live every Wednesday, 11am Central. You vote on the topic. See what’s next on our iwmentor.com schedule.

IW Mentor courses — Power Automate, Power Apps, SharePoint, Copilot Studio, and more. Built for business users. No developer required.

Ultimate plan — All courses, plus live Q&A with me every three weeks. Bring your broken flow. We’ll fix it together.

Which of these ten tricks are you trying first? Tell me in the comments.