Create Text With GPT - Manual approval action required now?

[CGX_bwang]

Hi,

We recently started running into an issue where the "Create text with GPT" action in Power Automate is requiring an approval action afterwards (each time the action is used) or it will give the error message below.  Previously, this was just a warning but now it appears to be mandatory and won't let you save the flow.  It also appears to be blocking previously built flows that used this action.  The error goes away if we add in the Start and wait for approval.  

Was wondering if anyone else has ran into this (or have any work-arounds) but for us requiring a manual approval each time the action is used would eliminate many of the use cases to automate certain tasks.

Thanks!

[takolota]

@CGX_bwang 

This doesn’t make sense & makes MS look out of touch for not understanding that many use-cases may involve reviews & approvals outside the immediate flow.

In the meantime, I’d suggest trying to just add the content approval to the very bottom of the flow & set its run after setting to only trigger if the previous action times out. Maybe if the action is present it will remove the error even if it never gets triggered.

[CGX_bwang]

@takolota  Thanks for the suggestion, I had tried setting the "configure run after" for the approval action to only when the previous action (GPT action) was failed/timeout/skipped but the error still picked it up. 

The approval requirement appears to actually apply on each use of the Create text with GPT action, meaning that if you have multiple GPT actions, it's requiring the approval action after each one or it'll show the error and refuse to save (meaning you can't just have one approval at the end sadly). 

Just seems unfortunate that we just started experiencing it this week so wasn't sure if it was a change on our end or from MS.       

[takolota]

Ridiculous.

The other thing I would think to do is set a Dataverse action to auto-approve the approval, but that’s a lot more work that shouldn’t be necessary.

https://youtu.be/bqOftVN7J_Y?si=rOuyxDZL2egXe-Q0

[takolota]

@plarrue @JoeF-MSFT @Ashbhatia 

[BadToro · Accepted solution]

Hey, I also don't need to check the output generated by GPT for what I'm doing. Here's a trick I found:

1. After the GPT step, throw in a 'Start and wait for an approval of text' action.
2. Make sure to turn on the 'Static Result' option on the approval action.

It'll then run the approval without having to sit and wait for it. 

[CGX_bwang]

@BadToro  That worked!  Would've never thought of using that feature, thanks so much! 

Not sure if MS intended for us to have to use that work-around, seems like it would be easier to just make the approval requirement a toggle or option.    

[ARB_wcc]

I just noticed it today after trying to modify a prompt and it wouldn't let me save it again. Making it mandatory to have an approval action doesn't make sense at all, specially for data extraction.

If anything, we can review the data, after it's been extracted, no need to be approving every single action.

¯\_(ツ)_/¯

[ARB_wcc]

@BadToro Thanks!! 👍🏻  'Start and wait for an approval of text' action with a 'Static Result' option is like setting up a cardboard cutout of a bouncer at the club entrance. Looks intimidating but lets everyone waltz right in 🕺💃😅

[geo1]

Are you still able to set a static result today?  I cannot enable it, say it is not allowed for that action. 

[BadToro]

Before you can turn on the 'Static Result', you've got to run the flow at least once. Cheers.

[ARB_wcc]

@takolota  / @plarrue / @JoeF-MSFT We have a new solution where we are using the GPT connector to "parse" folder names, extract relevant information from the name and populate file attributes.

Example below:

See below the prompt:

Now, let's talk numbers: we're dealing with a whopping 38,000+ folders. Just picture having to manually approve each file to be renamed in this sea of folders. It would be like playing a never-ending game of "Whack-a-Mole," but with folders instead of moles! 😂 📁📁📁📁

[Ashbhatia]

Hi there,

We understand your frustrations and concerns regarding the recent changes requiring manual approval for the "Create Text with GPT" feature in Power Automate. It's clear that these changes may seem restrictive, particularly when the intended use cases involve lower variance, like classification.

We want to assure you that the intention behind this requirement is not to undermine the expertise of our developer community or impose a top-down approach. Rather, it's a response to the unique challenges posed by Generative AI. These models, while incredibly advanced, sometimes have a tendency to fabricate information or are susceptible to prompt injection.  

As part of our commitment to responsible AI, we already have measures in place to moderate content and prevent the generation of inappropriate responses. However, controlling the tendency of the model to fabricate responses isn't as straightforward. Even though we guide the model to stay as factual as possible, it can sometimes deviate and produce misleading outputs. Moreover, there's also a significant risk related to 'prompt injection' in the data. Even with well-crafted prompts and automation, compromised incoming data can lead to unexpected results.

The landscape of risks associated with automation using Generative AI is a relatively new and not yet fully understood territory. As we navigate this space, we're taking a cautious approach to ensure human oversight and control over AI-generated content.

That said, we're actively committed to developing robust safety systems, and, as we gain a deeper understanding, restoring as much control as possible back to developers. Our ultimate goal is to evolve this capability collectively with you, the developer community.

Your feedback regarding the stability of prompts and the potential for versioning control is incredibly valuable. We recently upgraded to the latest gpt-3.5-turbo-16k-0613 model. Our goal to bring the latest and greatest of the GPT models to our citizen developers. The latest model also quadruples the token window from 4k to 16k. Model update might be a reason for the variance in response that you were experiencing.

Please bear in mind that this feature is still in preview. As we gather more insights and navigate this emerging field, we're continually refining and tweaking our offerings to improve the user experience. 
Your patience, understanding, and input are greatly appreciated as we work towards creating an environment that is responsive to the knowledge and feedback of developer community. Thank you for being a part of this journey with us.

Ashish

[Happymadness]

@Ashbhatia I just experienced the issue that OP is reporting. That is, a sudden change of the "action doesn't have a content approval action after it" message from Warning to Error, now preventing me from saving/testing/running my flow.

While it would be ridiculously annoying to manually approve EACH flow, especially since my use of 'Create text with GPT' is simply capturing verbatim text from .txt files and extracting to an organized .json object. However, it's something I am willing to do since my flow works like magic to automate my menial, hours-long, copy/pasting tasks.  

However, I'm unable to even implement ANY sort of approval in my environment since my Microsoft account is Enterprise level. Now, I'm plagued with this error: "This operation violates admin data policy 'MyCompany's Standard Policy', which restricts the use of business connector 'shared approvals' with non-business connector 'shared commondataserviceforapps'".

So, this mandatory requirement to manually approve is basically preventing me from using the 'Create text with GPT' action or any other AI builder content creation tool AT ALL! Like I said, this was a warning before today, and I had no issues before it went from Warning → Error.

I know I cannot use any business connectors with non-business (eg, AI builder tools). This is not an issue and never has been. My Enterprise admin's data policy is reasonable since my company works with sensitive data, and I would never open connections between GPT & company data on Power Automate. My flow fetches publicly available, published text and uses GPT to classify, structure, and reorganize the text word-for word so that it can be used easily in spreadsheets/tables. Nothing outrageous or morally questionable about it.

I understand Microsoft's rationale for mandating human approval & promoting responsible AI, but I hope Microsoft can revisit and optimize this policy since it can be seemingly more harmful than helpful. And I need this resolved- please advise on how to proceed! This is really harming my established, automated workflow. Thanks!

[miamoore]

@Ashbhatia Please relegate this back to a warning. It is unrealistic to expect someone to approve many individual AI responses and have the flow run for long periods of time waiting for approvals. Especially in scenarios where we know that it doesn't cause a harmful effect if the information is incorrect or the AI is simply just extracting text from a document. We deserve the choice. This has impacted my current flows and I need this fixed ASAP as this is disrupting my company's workflow.

[takolota]

EDIT: Found a way to make the connection work. Still using the previously mentioned static result work-around though to negate this needless approval requirement across 8 different GPT actions in the workflow. This is just stupid, I've been fuming most of the morning after both this & the GPT action just not working for a larger 4 page PDF document with a bad request / timeout error.

[Happymadness]

@Ashbhatia @plarrue @jk_one @Antrod @JoeF-MSFT @Chris @antoinec @Antoine2F 

Can we get a solution for this? The approval requirement is evidently doing more harm than good. Would appreciate hearing from AI Builder staff as soon as possible.

[Happymadness]

Please share if you have the time! I'm still stonewalled by the business vs. non-business connection issue. Feels hopeless...

[takolota]

@Happymadness 

For the solutions connection issue you have to select to manually test the flow once so it confirms your use of the connections. It just can't automatically run past flow runs before the new connections were added.

On the other issue I am having with the Create text with GPT timeout / bad request error, I've had to set something in a parallel branch to run if the 1st data extraction is taking too long. That will take only the 1st 5000 characters of the document text for the prompt & extract data from that 5000 characters. Then it returns that data to the app & terminates the flow so the other initial GPT action doesn't take 2 minutes before failing. 

[ARB_wcc]

This has become the most popular thread on the forum, +2k views 😂