Let us build a AI from scratch for the following business case.
Step1: Create a new AI model
Within Power automate framework, there are many delivered templates that can be used or create one from scratch.
Navigation: Power Automate> AI Builder
Choose the Custom model
Step2: Type of documents – Select the option for structured documents. You can teach the AI model to extract the data from different layouts.
After selecting structured documents, list all the key fields you want to extract from the documents. For example, Vendor name, Invoice number, date, hourly rate, Amount etc. You can add fields as below and select the type
Step3: Create Collection of documents – After selecting the document type and the information to extract, create a collection of similar documents. These are the documents (in our case invoices) which will be used to train the AI model. Minimum 5 documents should be attached. To get a better accuracy of the AI model, add more documents to the collection.
After the documents are uploaded, identify the individual fields which you want to extract from the template.
Step 4: Train the Model – After the documents are uploaded and key fields identified for all the documents, the system is ready to get trained
System will take a while to Train the model and you can see the status when Model is ready
Step 5: Review the accuracy score – Depending the on the number of documents and similarities in the document formats, the system will predict a score which gives you good indication about recognising all data. If the score is low, there are ways of improving the score by adding more documents and using other techniques as explained below.
Step 6: Prepare a sample invoice – As the model is ready for testing, prepare an invoice.
Step 7: Create a Power automate process – To test the AI model create a Power automate process.
e.g – You can use these steps to read Invoice from the attachment and send extracted data by email for approval or post it to a teams channel.
5.1 – Email arrives with Invoice attachment
5.2 – For each attachment, run the model and extract data
5.3 – Send extracted data in text format by email or post it to a teams channel.
As the data is stored on Dataverse, you can use hundreds of connectors to process the extracted data.
Email received the extracted data
