Answered

How could my flow be improved?

(1) Share

Report

Posted on by JS-20021447-0

I have created the following flow below. The first AI builder action (second action down) reads emails from a shared inbox and decides using my custom prompt whether the email concerns a single manuscript with ethical issues. If it does, it outputs the word "single", if it does not it outputs an explanation of why it believes the email does not concern a single manuscript with ethical issues. The output text is used as a string variable.

If it outputs "single" it goes along the first flow on the left and creates a report from the email along with a teams message and excel log. If not, it goes to the default path and logs the email details and explanation.

I am very new to PowerAutomate and I now wonder whether I used the best AI model for the first AI builder (create text using GPT). Can I train this model somehow? I can't see any option to. Should I have used the prediction model instead? I can't figure out how to train my model and have relied on adjusting the prompt. Does anyone think I should use a different AI builder model and if so, where do I train it?

Thanks for any help! I'm really enjoying creating but just am a total newbie.

Categories:

AI Builder

I have the same question (0)

All responses (2)

Answers (1)

Sort by

Verified answer

Ravi-Prajapati 416 Moderator on at

Like (0)

Report
Copy link

Link copied!
You're on the right track using AI Builder's "Create text with GPT" model for classification based on email content. However, this model is not trainable in the traditional sense—its behavior is refined through prompt engineering.

If you need a model that can be trained and improved over time, you might consider using the Prediction model in AI Builder. This model allows you to label past emails manually, train a custom classifier, and improve accuracy with more data.

To train an AI Builder Prediction model:

Go to AI Builder in Power Automate or Power Apps.

Create a new Prediction model and choose "Category classification."

Upload past emails and label them as "single" or "not single."

Train and publish the model.

Use it in your flow by replacing the GPT action with "Predict" from AI Builder.

If your use case involves structured data rather than free-form analysis, the Prediction model could offer better control.

Was this reply helpful? Yes No
JS-20021447-0 9 on at

Like (0)

Report
Copy link

Link copied!

Thank you Ravi, I really appreciate you taking the time to help me out. I will train up the prediction model and see if I can get that to work a little more reliably. Thank you!

Was this reply helpful? Yes No