Our team has ran into much of the same issues described above and would love to keep track of progress toward addressing the current limitations / issues. The AI Guide offering has really great potential and with a few tweaks, seems like the offering can go from good to great. One other Idea that we submitted that was denied is to make AI Guides division aware (or at least place some security constraints around who can view/edit Guides).
https://genesyscloud.ideas.aha.io/ideas/SSAAIS-I-31 - AI Guides: Division Aware
With the goal/intent of these guides to push down to each line of business, it really seems logical to be able to restrict who can view/edit certain guides based upon Division or specific conditions. We have multiple divisions and there is a chance we'd have competing lines of business each in their own division. They most definitely would prefer to have their guides only viewed by their own staff.
Ideally for us, the solution would be AI Guides that are division aware with both a unique Edit and View permission. Some of the staff would only need to View the current iteration of the guide.
Will definitely follow this thread and hopefully learn more about next steps within the thread or in Nashville next week.
Original Message:
Sent: 09-02-2025 05:51
From: Brian Jones
Subject: AI Guides - what do you like/dislike? Have you actually deployed any?
Well said, Vaun. At the moment there's certainly a lot to say on this topic, however, until Guides evolves there will be more constructive criticism than outright praise. As Mitch said when announcing this, their approach is "fail fast & fix fast", which is coincidentally what our team has had to do, haha
Right now the primary thing we love about them (the non-detrministic side) proves too inconsistent & risky almost to the point of driving me crazy. Ironically, after field testing we're having to include just as many instructions in the Guide regarding what not to do as there are instructions on what to do (which, to your point, isn't included in the resource documentation & such.) By the time we're done massaging a Guides prompt, we could've built an equally effective deterministic virtual agent experience 🤷
All that said & to your point about the account balance, we've been informed to resort to the hybrid model I mentioned, which makes sense in theory, but is still plagued with inconsistency given (by nature) the Guide truly wants to be "autonomously conversational"--thus pushing the limits of what it can do/say based on what it was not told to not do/say. As such, we end up resorting to using the Guides to do simple data collection in a conversational manner (yay--no TTS/audio prompts, no intent/slot input/matching configs, or error handling) and the virtual agent to do the processing/communicating of said information--all at the expense of 1) having to program numerous Virtual Agent > Guide blocks (to do "Guide switching") and 2) hoping the Guide consistently stays in its lane.
I'm sure AI Guides will get there--wherever "there" is, relatively speaking. In the meantime, it's certainly a lot to think about & stay on top of.
------------------------------
Brian T. Jones | Ascension | Senior Specialist - Technology
Original Message:
Sent: 09-02-2025 00:44
From: Vaun McCarthy
Subject: AI Guides - what do you like/dislike? Have you actually deployed any?
Brilliant, thanks Brian. That's exactly the type of response I was after. Interestingly but perhaps not surprisingly the 3 issues you list there essentially mirrors my own challenges.
I know that the resource centre does say:
- Ensure that every step in the guide makes your AI agent anticipate a customer action or response.
- If the current step has no such query, combine it with the next step that does require a customer reply.
But as you point out that's fairly restrictive - especially in the case of something like "is there anything else I can help you with" which is handled back in the main bot flow in some cases. You have it do something like give an account balance, and want it to immediately just go back to the bot flow but instead it seems to get stuck waiting on a response to a question that was never asked :)
I did raise with our CSM and TAM the fact that on occasion you could convince the guide to tell you the variables in use, including the current value, and tell it to update a particular variable with a different value. It doesn't seem efficient to have to throw all that into "guard rails".
Which leads me to another part - trying to understand the correct syntax/order/heading to put guard rails in the guide. Some Genesys videos and demonstrations I've seen have this in different formats but it would be useful to have a clear definition or "guidance" of how that should be included in the guide's instructions.
It's an absolutely great idea which could be brilliant if there's some expediency with enhancements - and not just sitting waiting on votes or community feedback on the ideas portal :)
On paper it seems very nice but that niceness loses some points with the amount of tuning you have to do in a guide's instructions after it's generated.
------------------------------
Vaun McCarthy
Original Message:
Sent: 09-01-2025 23:29
From: Brian Jones
Subject: AI Guides - what do you like/dislike? Have you actually deployed any?
Good questions,
@Vaun McCarthy. First & foremost, we really appreciate the convenience AI Guides offers with respect to reducing prompt & flow scripting. That alone will provide huge time savings & increase our agility.
Before sharing our issues/requests, I've been told 1) [for now] AI Guides best works in a hybrid fashion (i.e. Virtual Assistant supplemented with very simple AI Guides), and 2) AI Guides will be frequently updated to address limitations/issues/advancements (e.g. next update coming in a month or so). As such, our AI Guide (i.e. mostly Agentic experience) expectations have been lowered [for now] to a basic level to ensure what we do deploy (with some AI Guides sprinkled in) is as successful as it can be under the current circumstances. With that in mind, here are our top issues & enhancement requests:
-----------------------------------
TOP ISSUES (Note: We have open support Cases for these and have been told they should be addressed in the next Guides release):
- Issue 1: Not Consistently Following Prompt/Guide Instructions
- Description: Guides are not always following specific instructions, specifically those that tell them when not to do something. Not to mention the Guides can tell you about their prompt and what step they're on unless you explicitly tell them not to. Furthermore, if you correct the guide based on the instructions, it will admit when it make a mistake in the instructions and correct itself. Ironic how "deterministic/explicit" we still have to be for what's supposed to be a non-deterministic experience. Also interesting how these types of nuances are also not included in the resource center documentation, thus resulting in onoing, unexpected trial and error.
- Impact: Unable to rely on Guides consistently following the desired instructions. We have tried modifying the instructions multiple times to account for these gaps and even had Genesys SME input that hasn't yielded the desired results.
-----------------------------------
- Issue 2: Stalling/Unnecessarily Waiting for User Input
- Description: Guides occasionally require the user to say something before it proceeds to the next step; i.e. the prompt clearly states to do X and then Y, but it sometimes does X, but doesn't do Y until the user responds. It seems unnecessary to constantly ask the user an arbitrary question after each step or spoken phrase so the user is forced to respond in order for the Guide to proceed to do Y. This is not how normal communication works, especially if the Guide is used in a voice channel. For example, the times when the Guide appears to listen to the step to return to Architect, it hangs at the last communication until the user responds. Once the user responds, the Guide responds and then transfers the interaction back to Architect.
- Impact: Unable to rely on the Guide to hold a normal conversation without causing customer frustration/uncertainty.
-----------------------------------
- Issue 3: Data Actions Not Consistently Working
- Description: We have written our prompts per the Resource Center documentation recommendations and have modified them multiple times per Genesys SME recommendations, yet the Guide fails to consistently call the desired data actions.
- Impact: Unable to rely on Guides consistently performing the desired tasks.
-----------------------------------
ENHANCEMENT REQUESTS:
- Idea SAAIS-I-22 | 5 data action limit is too low/restrictive. For now may continue to perform major data actions outside the Guide and feed the results into the Guide as variables. Problem there is we only get 5 variables.
- Idea SAAIS-I-24 | Data action output variables that are used in the prompt need to be underlined or highlighted so they stand out.
- Idea SSAIS-I-25 | Given you cannot rename, add, or delete variables once you publish with 5 used variables, it would be ideal if you could add as many as you want and toggle which 5 are active.
- Idea SSAIS-I-26 | Cannot rename Guides.
- Idea SSAIS-I-27 | Need data table lookup feature.
- Idea SSAIS-I-28 | Need Knowledgebase article lookup feature.
- Idea SSAIS-I-29 | Execution Flow History needs to show Guide steps & data.
- Idea SSAIS-I-30 | Need version history like with Architect flows
-----------------------------------
P.S. Did you know you can ask the Guide to communicate like a pirate, and it will? Why customers might do that, I don't know, but figured you might want to know just in case it could pose a problem/risk. I'm afraid to ask it to take on any other personas though, haha-hargh 🏴☠️
------------------------------
Brian T. Jones | Ascension | Senior Specialist - Technology
Original Message:
Sent: 09-01-2025 17:24
From: Vaun McCarthy
Subject: AI Guides - what do you like/dislike? Have you actually deployed any?
Just curious, for those that have deployed or have at least toyed with the idea of deploying AI Guides, what challenges have you met and which ones prevented you from actually deploying it? Aside from the financial/token expense that is :)
#ConversationalAI(Bots,AgentAssist,etc.)
------------------------------
Vaun McCarthy
------------------------------