AI

I have a messaging AI Agent which pulls from our Knowledge Base (customer-facing support center with all our policies and procedures and FAQ). Our Terms of Sale have very specific legal information about our policies whereas our more FAQ focused articles have more generous and customer-friendly wording around our policies. The AI Agent uses the ToS to respond about our policy provisions but I’d like it to use the customer-friendly articles as the source instead. I’ve tried to tell it in an Instruction but that doesn’t seem to be working. Is there another way to dictate priority when looking at multiple sources or do I need to simply find a way to get it to not look at certain articles in our Knowledge Base (and risk it losing access to other information a customer might ask about)?The example is our return policy which is 14 days in the ToS but 30 days in the customer-friendly article (which is correct). 

Zendesk intelligent triage low accuracy after a migration usually means the AI is working as designed, on data it should never have seen. Zendesk’s AI learns from resolved interactions and applies those patterns to new requests. After a bulk Zendesk import, it learns from everything you moved, including parts you didn’t want it to repeat. Zendesk AI agent accuracy after migration drops for the same reason: the model can only be as good as the tickets and articles behind it.The main challenge is telling a data problem from a setup problem. When Zendesk AI fails after migration, both look the same in the agent queue: wrong routing, weak suggestions, and replies that sound confident but are wrong. Teams that guess often retune settings for weeks while the real cause lies in the imported records.If you haven’t migrated yet, my earlier post's two-step approach prevents most of this. This one is for when the import is already done.Symptoms and What They Usually MeanMatch what you’re seeing t

Hello everyone,I am Aarti Jangid, working at Dev Technosys. I am interested in learning how businesses can use AI chatbots and customer support automation to improve customer experiences.AI chatbots can answer frequently asked questions, provide instant responses, and help customers find information quickly. However, complex problems often require human assistance.My question is: What are the best practices for integrating AI chatbots into customer support workflows while ensuring customers can easily connect with human agents when needed?I would love to hear your experiences, recommendations, and insights on improving customer satisfaction through AI-powered support solutions.Thank you!

I am in the process of migrating from AI Agents-Essentials to the Agentic AI platform. Our Agent SOP leverages triggers & automations driven by tags Essentials adds to the tickets. This tagging behavior doesn’t carry forward to the Agentic AI replacement, so I’m in working on updating the triggers & automations associated with our Agent SOP for the new platform.Essentials adds a simple “ai_agent_automated_resolution” tag to a ticket that winds up consuming an automated resolution. But for the Agentic AI platform, I need to leverage the “Resolution type” and “Resolution tier” fields. It appears that the equivalent to the “ai_agent_automated_resolution” tag is “(Resolution type= Automated) AND (Resolution tier=Verified resolution)”.When I access these fields in trigger conditions, they have more choices than are currently documented (About ticket fields):“Resolution type” is documented as “Automated or null”, but the trigger condition has an “Assisted” option in addition to “Auto

Our agents currently cannot view their AutoQA reviews, while Admins/Reviewers can.This is a challenge for our QA process because agents are responsible for identifying and raising discrepancies in their AutoQA results, and AutoQA is also part of their KPI. Without visibility, agents cannot determine which category was marked down or validate whether the result is correct.We would like to request:Agent access to their own AutoQA scores and failed categories. Visibility of the Deep Evaluation explanation where available. An admin-controlled “Allow Agents to see AutoQA” setting. A conversation-level report showing Agent | Conversation ID | AutoQA Score | Failed Categories.This would help improve QA transparency, coaching, KPI management, and dispute validation.

Hi everyone,I am currently testing and rolling out Zendesk AI Agents via Email for automated response generation for and knowledge retrieval. While I see great potential, I am navigating two key operational challenges where I’d love to gather insights, best practices, or workarounds from the community.1. Controlling & Limiting AI Agent Response Length The Challenge: In my testing, response length varies widely depending on the number of knowledge base sources retrieved and the nature of the customer question (simple quick answers vs. complex troubleshooting). The Impact: Verbose, multi-source answers can overwhelm customers who need quick, direct resolutions, creating unnecessary noise in email confirmations and replies. What I’m Exploring: How are other teams setting boundaries or encouraging conciseness in AI Agent responses upon KB retrieval to ensure consistent output formatting? I’ve tried specifying an instruction to keep answer concise (also in Tone of voice) but it does

Hi everyone!We're using a Zendesk Messaging AI Agent with our Help Centre as a knowledge source. Some Help Centre articles contain hyperlinks to external resources (for example, another customer portal or documentation site). The links display correctly in the Help Centre article itself.However, when the AI Agent answers a question using content from those articles, the response often includes only the link text and not the actual clickable URL. As a result, users don't receive the link they need to continue.Has anyone experienced this? How are you handling external links in AI-generated responses? Are there any best practices for authoring Help Centre content so that hyperlinks are consistently surfaced by the AI Agent? Thanks.

Hey Community!If you’re using Zendesk Copilot, I’d love to hear how you’re approaching the rollout with your agents. What role have you played in introducing Copilot to your agents? How are you helping your team build confidence and incorporate it into their workflows? What has worked well so far, and what tips would you share with other teams? Is there anything you’re still figuring out? Looking forward to hearing what’s working, what you’ve learned, and any advice you can share with others getting started with Copilot! 

I’m gauging the feasibility and also looking for inspiration on building a custom agent (not a Zendesk agent, but an agent you’d build outside of Zendesk and then house in Zendesk. Any guidance on feasibility, or ideas on what others have tried would be greatly appreciated!

We use Zendesk and most of our tickets are created/generate via email coming into Zendesk, so bypassing all the help articles on Zendesk. I have read/know that you can use the new AI agent for email to scan the emails coming in and see if there are any help articles that could potentially answer the queries within the emails and offer these to the requester to see if those answer their queries, and if not we would like these tickets to go into the queue for human operators to pick up.Has anyone implemented this, know what triggers and automations you/we need to set up implement this?We do not have access to sandbox testing in our package, so it possible to test the email AI agent on certain email/tickets for example setting a tag on certain requesters (email addresses) so only emails coming from these requesters are processed/picked up by the AI agent, before we roll it out fully to pick up all emails coming in?Any advice and pointers would be greatly appreciated - thank you.

Hi everyone,I’m working on building procedures for Email AI Agent and looking for best practices on making responses truly dynamic rather than static or hard-coded.Specifically, I'm trying to solve two main challenges in single-touch email workflows:1. Tag-Driven Conditional LogicWe want the generated response to conditionally adapt based on ticket tags or custom fields (e.g., requesting an invoice only when a specific tag like warranty_oow exists on the ticket). What are the best patterns for structuring procedure nodes or prompts so metadata triggers the right output without internal evaluation logic leaking into customer emails? 2. Context-Aware Intake (Asking Only for Missing Info)When collecting triage details, we want the AI to analyze the customer's initial message and dynamically request only what's missing—skipping questions for information already provided (like model name, serial number, or troubleshooting steps already tried). How do you reliably prompt or structure procedu

Today I got a 2nd email announcement about the coming end-of-life for AI Agents-Essential. This email notes that the changes are rolling out starting April 27 (i.e., 2 days ago). But if you link to the announcement post, you’ll note that the rollout has slipped a month (to May 25-June 12) with no change to the end-of-support/end-of-life dates.The post also notes “In the coming months, we’ll share detailed timelines, transition plans, and guidance to help you transition to the upgraded AI agent experience.” (emphasis added). Those details, plans, and guidance should be in place now. The schedule has barely 2 months from the end of the rollout window until the end-of-support date.We have a Professional instance - no sandbox. Any changes we need to make are to our production environment. This is fraught with risk. Setting up AI Agents-Essential for email was trivial and works well. Reading the documentation for email for AI Agents-Professional makes my head spin with the level of complexi

I am trying to prepare for the end-of-life of AI Agents-Essential, and since we are Suite Professional, not Suite Enterprise, I have to do this in my production environment. So the risk of breakage is very high. I’ve built a set of automations and Agent SOP around the AI Agents-Essential behaviors and I cannot find sufficient documentation on the new AI Agent email behaviors to have any idea how to make things work the same way.For example - I have triggers for the ar_suggest_true and ar_suggest_false tags being set on a ticket. to guide my Agents? When a ticket is tagged ar_suggest_true, the Agent reviews the AI Agent’s response and unless it’s clearly something they need to manually address, move the ticket to a custom Deferred status, taking no further action unless the end-user indicates that they still need help. Conversely, a ticket tagged ar_suggest_false gets priority because no response has been sent to the end user.Do these tags still exist with the new AI Agent? Is there som

Hello Everyone,I am running into issues when trying to implement a phased roll out of our Ai agent for email.I am trying to roll out specific use cases and have the Ai agent only respond to those.The issue I am seeing is that the generative reply aspect we cannot turn off. So my Ai agent when I place it in production replies to all emails. Rather than just the specific use cases activated.I have tried intents and disconnecting my knowledge sources for email channel.Would love any suggestions or feedback from you all. Thank you,Brandon

Hi Zendesk Team,We are currently using Zendesk AI Agent together with human agents to handle customer support tickets.We have encountered an SLA management issue when an AI Agent escalates a ticket to a human agent.Current SituationWhen a customer contacts our AI Agent, the AI Agent may respond first and then escalate the ticket to a human agent when it cannot resolve the issue.However, because the AI Agent has already sent a public reply, the First Reply SLA may already be satisfied.After the ticket is transferred to a human agent, there is no new SLA countdown specifically for the human agent's response.As a result, the ticket may remain in the human agent's queue without an effective SLA constraint, even though the customer is waiting for a human response.ExampleThe process is:Customer → AI Agent → AI replies → Escalate to Human → Human AgentThe problem is that after “Escalate to Human”, we would like Zendesk to start a new SLA timer for the human agent.For example:AI Agent escalate

Businesses are moving to cloud contact centers because traditional systems can be costly, complex, and difficult to scale. A Cloud Contact Center Solutions gives teams more flexibility to manage customer calls and interactions from anywhere.Some key benefits include:Lower infrastructure and maintenance costs Easy scalability as teams grow Better call routing and IVR Real-time reporting and analytics CRM and business tool integrations Support for multiple communication channelsPlatforms such as MCube are also helping businesses move their customer communication to the cloud. However, the right solution ultimately depends on the business's call volume, team size, integrations, and customer service requirements.

Hi,One challenge I’m seeing with Zendesk Guide is keeping the Knowledge Base accurate and useful as content keeps changing.With AI and generative search becoming a bigger part of the Help Center experience, outdated or duplicate articles can become an even bigger problem.The challenge is not just creating new articles. It is also knowing: Which articles need updating Which content is outdated or duplicated What information customers are actually searching for How to keep content consistent across different knowledge sources For example, if a process changes, it is not always easy to identify every related article that may need an update.I am interested in how other teams are managing knowledge maintenance and making sure their Help Center stays AI Ready as it grows.

Hi,When rolling out an AI Agent for email, I think it’s better to start with a few specific use cases instead of enabling it for every email.A few things you need to check first: Review the email automation trigger and its conditions Make sure only the intended emails reach the AI Agent Define what should happen when an email doesn’t match the use case Use the Default reply and escalation flow to hand off cases that AI shouldn’t handle Test both in-scope and out-of-scope emails before going fully live  

O maior desafio da IA no atendimento não é a IA. É a definição dos processos.Quando se decide implantar IA no suporte, a conversa começa pela ferramenta: qual bot, qual integração, qual modelo. Mas o travamento quase nunca está aí. Está no processo que ninguém consegue desenhar no papel.Na prática, o que aparece é isso:Fluxos que existem só na cabeça de duas ou três pessoas do time. Base de conhecimento com artigos de dois anos atrás descrevendo procedimentos que já mudaram. Vinte macros criadas ao longo do tempo, metade em desuso. Regras de escalonamento que o agente sênior "sabe", mas que nunca viraram critério objetivo.Automação não corrige nada disso. Ela executa o processo ruim mais rápido e em escala maior. Se o fluxo de reembolso tem três exceções não documentadas, o bot vai errar nessas três situações todos os dias, sem cansar. E uma IA lendo uma base inconsistente produz respostas inconsistentes — não é limitação do modelo, é reflexo do insumo.O trabalho que vem antes: mapear

For those using Advanced AI, I’d love to hear how you’re managing widespread or known issues to help prevent those tickets from unnecessarily reaching a human agent.For example, when there’s a known issue impacting a large number of customers, what have you put in place so Advanced AI can recognize the issue, provide the appropriate messaging or workaround, and deflect the ticket when human intervention isn’t needed?Are you using specific Knowledge Base articles, intents, procedures, temporary content, triggers, or another approach?I’m especially interested in how others balance quickly responding to widespread issues while still making sure customers who genuinely need additional help can reach an agent.What has worked well for you?

A short summary of what the response covers: how Nuvirateck started using AI agents to cut down repetitive sales and delivery work like outreach drafting and quotation checks, what data and templates we had to clean up first, how we kept a human-review rule so the team trusted it, a real example of the agent catching a pricing error on a client quotation before it went out, and the advice to start small with narrow, repetitive tasks before trusting an agent with judgment calls. We started looking at AI agents because our team was drowning in the repetitive parts of sales and delivery work: drafting outreach at scale, reviewing quotations line by line, cross-referencing inbox replies against campaigns, and putting together decks and proposals under tight deadlines. None of that work is hard, but it eats hours that should go to actual conversations with prospects and customers. That was the real trigger for us, not curiosity about the technology for its own sake.Before we could really us

We have a small pilot users testing with the AI agents.  They are having conversations with the AI agent in the help center using the chatbot.  I was expecting to review their conversations in the “View unresolved conversations” in the AI agent settings.  I only found conversations 8 days ago (April 1 and before).  Nothing between April 2 and April 8. Is there a setting that I need to configure?

Hi,For us, the main reason to explore AI Agents was to see how we could handle repetitive support requests more efficiently while still keeping the customer experience personal.Before experimenting, we focused on a few important things:1. Clear goalsWe first identified which type of requests would actually benefit from AI. The goal wasn't to automate everything, but to reduce repetitive work and give agents more time for complex conversations.2. Good documentation and knowledgeWe reviewed our Help Center content and made sure the information was accurate and easy to understand. This was one of the biggest lessons for me—AI needs good knowledge to give useful answers.3. Clean processes and workflowsWe looked at existing ticket fields, tags, triggers, and automations before introducing anything new. This helped avoid adding AI on top of an already complicated workflow.4. Team alignmentIt was important that everyone understood what AI would handle and when a conversation should be passed