Question for everyone : How are you using Create Text GPT in AI Builder?

[plarrue]

Hi everyone, 👋 From the AI Builder product group we want to hear from you.

👉Have you tried Text Generation in AI Builder already? How are you using it? For what use case? Is there anything you would like to see added or changed to make it even better?

👉We are interested in knowing about your AI Builder experience. Please share your story!

Thank you!

[David_MA]

I have explored some of the possibilities of the AI options (getting data from receipts and sentiment analysis) within the Power Platform, but I don't think it is ready for real-world use yet. I have also used ChatGPT to get assistance with developing flows when I am struggling with an expression. It often makes up its own commands. For example, yesterday I wanted an expression to get the current week number and determine if it was an odd or even week. It kept providing false information. Attached is the transcript of the conversation I had with the AI.

I was eventually able to figure out an expression by what the AI was trying to do (even if some of the commands were not in Power Automate), so it is helpful from that perspective. And when I showed it my result, it did a good job of breaking out the expression and explaining it.

Additionally, I am in a heavily regulated industry. I am concerned about the data processed by the AI being used to train it further and then that information getting used outside of the company. I have read some articles in the news already where companies have used AI and it has caused issues. I've seen that Microsoft plans to release a private version of the AI tools, but initial reports say this will be 10 times more expensive than the current version. See Microsoft could offer private ChatGPT to businesses for “10 times” the normal cost | Ars Technica for more information.

This is the expression I ended up with:

if(equals(mod(add(div(dayOfYear(utcNow()),7),1),2),0),'Even','Odd')

 And this is the explanation of it from ChatGPT:

Here's how the expression works:

• dayOfYear(utcNow()): This function gets the current day of the year.

• div(dayOfYear(utcNow()), 7): This expression divides the current day of the year by 7 to get the current week number.

• add(div(dayOfYear(utcNow()), 7), 1): This expression adds 1 to the current week number to ensure that the first week of the year is considered week 1, not week 0.

• mod(add(div(dayOfYear(utcNow()), 7), 1), 2): This expression calculates the remainder of dividing the adjusted week number by 2, which will be 0 for even weeks and 1 for odd weeks.

• equals(mod(add(div(dayOfYear(utcNow()), 7), 1), 2),0): This expression uses the equals function to check if the remainder is equal to 0, which indicates an even week. If it is, the expression returns 'Even'. If it is not, the expression returns 'Odd'.

Attached is the complete transcript with ChatGPT for creating the expression and at the end it tells why it didn't provide this expression to me in the first place.

If I have any advice for people using AI to help with workflow development, be persistent with the AI even if at first it gives you erroneous information. Feed it the error messages and analyze what it is trying to do. Additionally, always feed it generic information and nothing personalized or with company data unless you really know how it is being used and shared by the AI.

[JoeF-MSFT]

Hi @David_MA - thank you for sharing this feedback! Nice use of ChatGPT to help you build Power Platform solutions.

• Do you remember which limitations you faced when getting data from receipts and sentiment analysis? We're continuously improving these AI models. 
• Have you tried the AI Builder generate text with GPT action? How text generation in Azure OpenAI Service works (preview) | Microsoft Learn It uses the same AI model as ChatGPT. Your data is maintained in accordance with your expectations, and Microsoft does not use it to enhance the model or share it with OpenAI. 
• Power Automate has two AI-infused features that you might find interesting:
  • Describe it to design it. You can use it to create a flow just by describing it! Create a cloud flow from a description (preview) - Power Automate | Microsoft Learn
  • Format data by example. You can use it to generate expressions by simply giving examples of what you want to achieve: Use format data by examples to change the format of data with Power Automate - Power Automate | Micr...

Keep the feedback coming!

[David_MA]

For the sentiment analysis, I tried putting text from different news articles through the AI engine. The results seemed mixed at best to me. A couple articles I thought were positive it ranked as negative, and a couple I thought were negative it rated neutral. I thought this might be useful for analyzing things such as social media posts so a marketing/public relations team could be notified of very negative or highly positive posts. I am not sure the AI is ready for this. I suppose this type of thing is highly subjective, so I don't know how reliably AI could handle something like this.

For processing receipts, it did quite well on this but not good enough for production use I thought. It had a very hard time with figuring out the taxes as that was where the majority of errors were found. I thought this could be useful for setting up a hot folder where people could upload their receipts and the AI could process them. Attached is a PDF of the receipts I used (downloaded images from the Internet) and a spreadsheet where I wrote the processed data to. I highlighted fields in yellow where it got the result wrong.

It only got one date wrong, and I think it was because the receipt had the date in the format of mm-dd-yyyy rather than the usual North America format of mm/dd/yyyy.

[David_MA]

Regarding using the feature to describe a flow and then let the AI build it, I have never had much luck with this feature. It may be good for basic workflows or for beginners, but I have found so far that what it creates requires too much rework than if I had just started it from scratch or it cannot create anything.

For example, I asked it to build this: When a new MS Forms response is submitted, check to see if it is a request for deleting an account. If it is, then check to see if it is for deleting the account of the user who submitted the request, or another user as specified in the request. If it is for a user specified in the request, then check to make sure the account is valid in Azure AD. If the account cannot be found, Post and adaptive card in a Teams channel so the IT team can look up the account and provide the correct e-mail. Finally, send a confirmation e-mail to the requester. If the user account cannot be found for the other user, the confirmation email needs to tell the submitter to contact IT to review the request.

It came back with this:

All of the actions to do what I want exist, as I have built this flow, which is the following so far:

[willtwilson]

I'm keen to try it but sadly it doesn't appear to be available yet in our region. Is there a list of Preview regions? I couldn't easily find this information.

Thanks.

[plarrue]

Hi @willtwilson ,

Thanks for reaching out - At present, access requires a Power Platform environment located in the United States. However, we are actively striving to ensure accessibility for everyone.

Regards.

[takolota]

GPT PDF & Image data extraction template: https://powerusers.microsoft.com/t5/Power-Automate-Cookbook/Extract-Data-From-PDFs-and-Images-With-G...

It doesn’t require specifying certain document areas, wordings, styles, data, etc. It just OCRs the file, converts it to a replica text (txt), and passes it to a GPT prompt where you can ask GPT to do whatever you want with the document data.

Even if you have some documents where the traditional models work better, you could set some criteria for your fields that may indicate a mistake & have it then run this method on it to double check if the custom model results agree with these results.

[JoeF-MSFT]

This is incredibly cool! Thank you for sharing @takolota! The spectrum of possibilities that GPT brings to the table is nothing short of spectacular!

[ARB_wcc]

Hi @JoeF-MSFT,

Se below 2 prompts as an example, the use cases are endless but I find myself mostly using it to extract data or even for cognitive analysis like reading email body and evaluating which invoices have been approved or declined. The model is good but I think it can be improved, but that's probably because I'm used to GPT4 now!

[JoeF-MSFT]

Thanks for sharing @ARB_wcc. Really powerful prompts! And love the use of emojis. 😍

[ARB_wcc]

Hi @plarrue,  I wanted to share my positive experience with the GPT connector. It has exceeded my expectations, delivering an impressive 93% accuracy rate with the prompt below. Moreover, it helped me simplify the process by replacing multiple formulas with a single, straightforward prompt. However, I have noticed the AI occasionally assumes missing data, which, in some instances, leads to inaccuracies. While this does not dramatically affect the overall performance, I am curious if there's a way to configure the AI to only interpret the input as-is, without making assumptions.

See example testing animation below:

Prompt Testing Animations

And the results after testing the prompt with 50 random examples:

[plarrue]

Hi  @ARB_wcc 

Wow great work. Your prompt is terrific!

To answer your question on how to configure the prompt to avoid any assumptions, you can try by adding an escape instruction to your prompt.
If the text below has less than few words or looks like a placeholder text, respond "Unknown", otherwise respond with the shipment number.

PS: Please feel free to share your flow in our AI Builder Cookbook https://powerusers.microsoft.com/t5/Power-Automate-Cookbook/bd-p/MPA_Cookbook?filter=ai%20builder so other person in the community can benefit of your great work.

Thanks!

[takolota]

Hello @JoeF-MSFT & @plarrue 

I shared my template above, but I am not yet able to implement it as there are challenges with our head of IT security.

He was not convinced there is enough data privacy when using GPT. I shared some documentation stating the data isn’t shared with anyone (https://learn.microsoft.com/en-us/legal/cognitive-services/openai/data-privacy).
But he says it is just an assertion & the situation is still developing. He has yet to see any independent verification of data privacy & security while using GPT.

Is there any way you can help with this?

Is there more documentation & independent verification of GPT & these MS AI Builder actions yet?

[JoeF-MSFT]

Hi @takolota - that's a fair concern, but rest assured that we take privacy and security very seriously at Microsoft as reflected in our products terms of use and privacy policy. 

In addition to the documentation link you shared, I can reiterate that:

• Microsoft doesn't use customers’ data to train LLMs. We believe the customers’ data is their data, aligned to the Microsoft' data privacy policy.
• The foundation models aren't improved through your usage. This means your data is accessible only by authorized users within your organization unless you explicitly consent to other access or use.

FAQ for Copilot data security and privacy in Microsoft Power Platform - Power Platform | Microsoft L...

[UmarFarooq18]

I've been using Create Text GPT in AI Builder to help me generate a wide range of written content. Whether it's drafting blog posts, creating marketing copy, writing emails, or even coming up with ideas for stories, Create Text GPT has been an invaluable tool.

I start by providing a brief description of what I need, and then I let Create Text GPT work its magic. It quickly generates paragraphs of text that I can review and fine-tune to fit my specific needs. The ability to customize the tone, style, and length of the generated text is really useful, as it allows me to align the content with the brand voice or context I'm aiming for.

Of course, I always review and edit the output to ensure it's exactly what I'm looking for. This collaborative process between AI and human creativity has been a game-changer for me. It saves me a lot of time and gives me a solid starting point for any writing project.

Overall, Create Text GPT in AI Builder has become an integral part of my content creation process, helping me generate high-quality, engaging content more efficiently than ever before.
We can also use AI in Recruitment.

[plarrue]

Hi @UmarFarooq18 ,

Thanks for sharing your experience with us. I'm glad to hear that Create Text GPT in AI Builder has been such a valuable tool for you, saving you a significant amount of time and providing a strong foundation for your writing projects. 
It's good to see that you're also recognizing the importance of ensuring the accuracy and appropriateness of AI-generated content before using it.

Regards.

[Gabriel_Mayorga]

It worked well for less than a month 
I created 3 flows to extract information from emails and conver into a JSON array, and I was able to use the information to fill my Sharepoint list.
It was working well BUT suddenly since last week my responses are:
* Sorry, I can't respond to that request.

* "  I understand that you would like me to extract the following information from the given text: ......... output to a JSON array. Is this correct?  "  how can I respond Yes ?
* Im sorry, I'm not able to extract the email addresses from the text you provided.

Also I've noticed that my AI Builder / Models  screen shows "You've consumed all of your AI Builder credits. Contact your administrator to get more capacity."   but my Environment have 10,000 credits available and not used.

Really not sure what is causing this problem since this was working OK and since last week nothing is working ....

Any advice ??

[ARB_wcc]

Haha the same happened to me, I had a prompt to translate emails to JSON but it stopped working a few weeks ago, for some reason it was refusing to process my prompt just because it had the word "email" in it, so I replaced the word email by "text" and now it's working again. I guess you just need to be more creative with your prompt now 😅

[takolota]

I’ve now updated that PDF & image to JSON data extraction template

https://powerusers.microsoft.com/t5/Power-Automate-Cookbook/Extract-Data-From-PDFs-and-Images-With-G...

It’s now more accurate & much faster.

But I do look forward to being able to switch to a larger context window model on the Create text GPT action as that will enable a little bit better accuracy & processing of even larger files.

Not to mention being able to switch to even faster or more accurate LLMs beyond GPT 3.5. With that I think we could get the processing time down to around 10 seconds & the accuracy up to 95-99%.