Recent searches


No recent searches

Using AI to generate replies in an AI agent for messaging (Legacy)



image avatar

Aimee Spanier

Zendesk Documentation Team

Edited Feb 03, 2025


1

96

96 comments

image avatar

Daniel Aron

Zendesk Product Manager

Hi Fiona, using content blocks in articles should not affect the bots ability to generate replies. We have however found an issue relating to article titles that is likely the cause of failure. A fix is in progress, and once it is released I'll let you know and you can retest this example.

0


I am getting "I apologize for not being able to naswer your question. Hear are some topics that might assist you." when typing in the title of a KB article. I do not have any restrictions set up with the exception of all articles are required to be signed-in to be able to view. I am hoping that this is just an error in me setting up the bot but essentially have followed everything I have seen. 

0


Hi Daniel Aron, and thanks for letting us know about this! I just returned from a sabbatical, so we'll be setting this up in the next week or two and giving it a try. I received a survey, but as I've been out of the office, nothing to report just yet. 

0


Hello,

I have created a new Test Bot to evaluate the replies generated by AI. The bot seems to be suggesting relevant answers; however, the link to the article is incorrect and points to a Knowledge Base that is not associated with the brand for the bot. The Read More button links to a different article and it does seem to be linking to the correct brand/article.

Do you know why I am getting this inconsistent behaviour?

0


Daniel Aron Brett Bowser

Is there anything on the roadmap for either of these?

- Option similar to the "Leave a message" functionality on Web Widget (Classic), so that our users can open a ticket rather than opening a chat

- Option to make the subject line more descriptive than "Conversation with <User Name>" so that we can see what the conversation is about

Thanks!

0


image avatar

Daniel Aron

Zendesk Product Manager

Hi Chris Wilson can you confirm that the necessary pre-requisites were met in order for the bot to reference restricted content? The user must be authenticated using the messaging JWT authentication and the sunco appuser id must be mapped to the zendesk id via external id. Also note that restricted content isn't supported in the "test bot" feature, so it needs to be tested in a live environment.

0


image avatar

Daniel Aron

Zendesk Product Manager

Hi Gaby Torres we've seen other examples like this where generative AI goes outside the bounds of its directive. The team is constantly looking at ways to minimise this but it can not be completely avoided. If the use-case is a concern, we suggest creating a bespoke answer flow for that intent and linking to that rather than relying on a generative reply. cc Monika Vogel  

2


image avatar

Daniel Aron

Zendesk Product Manager

Hi Alan Pugh thanks for the feedback. The short answer is no, there isn't anything planned for either of those. However, I think the first one can be easily address by configuring the bot to present a simple greeting message with an option to "leave a message" and then link to an answer that transfers to an agent. You can set expectations in the bot messages for when they might expect a reply from the agent and the fact it will come via email if that's your preference. This article may help.

0


image avatar

Matt Russell

Zendesk Luminary

Hello.  Two issues we are running into with our chatbot GenAI.  After speaking with our ZD Service Consultant, she recommended I raise these two points here.

  1. With Generative AI replies on, we are routinely seeing the first answer fail (something to the effect of "sorry, can you rephrase that") but if we copy/paste the exact same question, with the second try a lovely generative reply will be given with the expected answer and link to an article within our KB.  Why does it take this 2nd nudge to get the GenAI to kick in?

  2. Often times, when we get a GenAI response that includes a link within our KB, we see a URL written within the text (Tatiana also mentioned this above) and additionally we get a Read More button.  The issue we have is that the written URL and the hyperlink of the Read More button are for different articles within our KB.  We have found the path in the Read More button is the article we would expect, but the URL within the body of the text is not the best article, but more confounding that they are different altogether.
    Would be ideal to remove the written hyperlink, and only use the Read More.

Thank you in advance, -Matt Russell

3


I'm also seeing the same as Matt Russell.

I just tested for a question that is directly answered in our articles and it took 3 attempts and very slight rewording for the bot to understand (see screenshot). When I later attempted the second variant, it understood and answered correctly...

0


We're also seeing the same issue of incorrect articles (but otherwise correct summarisation)

 

https://support.zendesk.com/hc/en-us/articles/6138268212634/comments/6856225169562

0


Thanks for your reply! @Daniel Aron

https://support.zendesk.com/hc/en-us/articles/6138268212634/comments/6853511762586

I also noticed that some anwers include a link to the article as well as the Read More option and some answers only include the Read More button.  What determines for the answer to give the link?  Is there a way for the link not to show and just keep the Read More? The options suggested in Read More seem to be more accurate than the link. 

Thanks!

 

 

0


image avatar

Daniel Aron

Zendesk Product Manager

Hi Matt Russell and Fiona in regards to the first issue highlighted

Why does it take this 2nd nudge to get the GenAI to kick in?

I'll share this with the team to investigate. In the meantime, I suggest disabling the If the bot needs clarification feature in the bot behavior settings to prevent this from happening.

2


image avatar

Daniel Aron

Zendesk Product Manager

In regards to the second issue Matt Russell Patrick Curry

written URL and the hyperlink of the Read More button are for different articles within our KB

Thanks for this feedback. We'll take this on board for further consideration. cc/ Monika Vogel

4


@Daniel Aron Hi. We have some scenarios, where our HC article recommends contacting customer support. So when a customer is asking he gets an answer from the bot "Please contact customer support" with a follow-up question "Was it helpful?'. Of course, customers choose 'yes', but actually it doesn't solve their problem. If there is a separate intent for such cases, it's possible to create a custom answer with a step that transfers to an agent. But if there is no separate intent, we're at a dead end. Do you have any recommendations? 

0


Generative AI in chatbot not generating answers when connected to one brand but works fine with other. Both the brands have help center active and have articles also....

 

0


Hi Daniel,

I am also getting the problem Chris reported. The bot does not seem to have access to the articles (they are all restricted). Let me give you some details:

  • All the articles of the help center are restricted to agents and admins. It's a knowledge database for agents.
  • Same result if I test the bot than if I execute the same test live in the HC.
  • The bot always answers with the fallback response. 
  • I am testing it with an agent and with an admin, they are authenticated in the system of course (Zendesk authentication). I do not think, other kind of authentication is necessary, right?
  • Language is German
  • Configuration is pretty basic, the customer just wants to test the generative replies.

0


Hi,

I wrote before about the "ugly" url in bot answers earlier :) it would be great, if it showed Text on link that was linked instead of the linked URL.

This is related, as I have articles with links to headers within the article. In this case, the bot refers to the header #, which makes no sense for the user. I hope the team can do something about that. Thanks in advance. 

 

3


We have the same issue where the messaging bot's behavior is set to generate a reply but does not answer with any meaningful response.

We have both public and private KBs in our help center, and are English language only.

Support stated we also need to enable the Answer Bot - is this not for the classic web widget only?

0


When Dutch customers use the Bot the "Was this helpful?" question, after a generated answer and article recommendation of the Bot is translated to "Wat is dit slattering?". Translated in English this is "Is this slutty?". We had a few Dutch customers complain about this "vulgar language" now. :( Could you fix this? It is the generated question, nothing we wrote in a flow, that is translated incorrectly.

0


Hi. 

 

We have a restricted-access help center and within it several different user segments. Apparently our bot only replies autonomously with answers based on articles that are bound to native user segments like “Signed-in users”. If the articles are restricted to custom user segments like “Partners-only” or similar, then the bot ignores the articles altogether. 

 

Is there documentation on the relationship of generative replies with the help center's user segments that I can consult?
Also, and above all, does the bot handle restricted-access articles properly or are there known limitations (I think I saw a comment somewhere that handling restricted-access articles was going to be released Jan 2024)?

 

Thank you.

0


Hello,

Is it somehow possible or on the roadmap that Generative Answer looks as well in Gather / Community section and articles ? 

Thank you

1


Is there a way to turn off having the “Read More” link presented in an answer? 

1


0


image avatar

Daniel Aron

Zendesk Product Manager

Hi Leslie, enabling generative replies to draw from content beyond help center articles (inc. website content) is still on the roadmap. However, we are currently focussed on some other priorities and we don't expect to be able to deliver this feature until 2025. If this is of immediate importance to your business, I suggest looking into Ultimate which currently supports this with uGPT. I went ahead and made your account representatives aware that you might be interested in Ultimate. 

0


image avatar

Daniel Aron

Zendesk Product Manager

Hi JJ F. no it isn't possible to remove the “read more” link with generative replies. Can you share why you'd like to be able to do this?

0


We are experiencing something weird with our bot. They are recommending our articles correctly in most cases, but sometimes, it recommends things like this:

When we click the link, it doesn't send us to our FAQ, but to some other site from a different company.  Also, important to mention that we do have articles about the subjects in the messages. Do you know how to fix it? 
 

 

 

1


@Thais Souza da Silva - Thank you for sharing your experience with the bot. I'd like to follow up on your post and question about the weird behavior. Is your screenshot an actual response…or have you edited it to use the “example.com” and “company.com” domains as placeholders so as not to share the actual URLs? And also, is the bot responding in English to a question in another language?

0


Sally Anne Dishong Thanks for answering. These are real answers from the bot, with “company” and “example” in the middle of the suggested links. And it is answering in the correct language to the users, this is not a problem. Here you can see another example:

 

1


Thanks for your response, Thai! Our Customer Support team is currently evaluating the use of the bots and your experience is helpful as we continue our research.

0


Please sign in to leave a comment.