A 40-person support team at a mid-market SaaS company turned on Zendesk AI six months ago. They expected 60% automation. They got 23%. The bot answered password reset questions fine. It told customers to check the help center when they asked about refunds, billing disputes, or anything requiring a decision. Tickets that needed action still landed on human agents, preceded by a bot interaction that added two minutes of friction and zero resolution.

That team is not unusual. Zendesk markets a path to 80% automation, but public case studies cite actual resolution rates in the 39 to 66% range. The gap between the marketing number and the operational reality is where most CX teams live: somewhere between "this helps a little" and "we expected more."

This article covers what you can do to close that gap inside Zendesk, and how to know when the gap is structural rather than fixable.

Fix your knowledge base first.

Most Zendesk AI underperformance traces back to the knowledge base. The AI generates answers from what it finds in your help center, and Zendesk only indexes the first 100,000 bytes of each article. If your articles are long, the AI may never see the answer it needs.

There is a more fundamental problem. Zendesk's generative AI does not follow external links in help center articles. If an article says "see our refund policy" and links to a PDF or external page, the AI treats that information as nonexistent. Every piece of information the bot needs must live as plain text inside the help center itself.

The fix is labor-intensive but effective. Audit every article for external links, embedded PDFs, and references to information that lives outside Zendesk Guide. Pull that content in. Break long articles into focused pieces, each answering a single question. Zendesk's own documentation recommends creating one article per specific question so the AI can extract precise, accurate information. If you have the Copilot add-on, you can use the text expansion tool to turn bullet points into full paragraphs quickly.

One more thing. G2 reviewers report a noticeable lag from updating a knowledge base article to the AI actually incorporating it. Teams have noted delays of a week or more before new content appears in answer suggestions. Plan your content updates accordingly. Do not publish a new policy and expect the bot to know about it tomorrow.

Tune your intents.

Zendesk's intelligent triage automatically classifies tickets by intent, language, and sentiment. It is one of the more useful features in the Advanced AI add-on. But it has hard limits. Optimal performance sits at 30 to 40 intents. Mature models handle 60 to 80. Beyond 100, accuracy drops significantly.

If your business has more than 100 distinct customer intents, Zendesk's triage model starts to collapse. That is not a configuration problem. It is a model limitation.

The workaround: group related intents into broader categories and use triggers or automations downstream to subdivide. A single "billing" intent can route to a group where triggers separate refund requests from invoice questions from payment method updates. You lose some of the AI's upfront classification accuracy, but you keep the routing functional.

You also need a minimum of 1,000 resolved tickets in the past six months for intelligent triage to function at all. New Zendesk customers or teams that recently migrated cannot use intent-based routing until they build that ticket history. For custom intents, budget 50 to 200 training examples each.

Build answer flows, not just articles.

Zendesk offers two AI agent tiers: Essentials (included with Suite plans) and Advanced (the paid add-on from the Ultimate acquisition). Essentials generates answers from your help center. Advanced lets you build structured conversation flows with conditional logic, API calls, and backend integrations.

The difference matters. Zendesk's own progression model estimates that generative answers alone reach about 30% automation. Adding structured flows pushes that toward 40%. Adding backend data connections reaches 50%. Adding QA and continuous optimization reaches 60%. The 80% number requires all of these layers operating together, plus scale.

If you are stuck at 20 to 30% automation, you are probably running Essentials and expecting Advanced-tier results. The answer is not a better knowledge base. It is structured flows that walk customers through multi-step processes: checking order status by pulling from your OMS, processing a return by writing to your RMA system, updating a subscription by calling your billing API.

Building those flows requires the integration builder, which connects to external APIs. Zendesk's documentation stresses adding clear descriptions to every API action so the AI knows when and how to use them. Without those descriptions, the bot has integrations it does not understand.

Watch your resolution billing.

Zendesk's pricing model shifted to outcome-based billing with the Resolution Platform launch. Automated resolutions cost $1.50 per resolution on committed plans and $2.00 per resolution for pay-as-you-go overages. Each plan tier includes a baseline allocation: 5 resolutions per agent per month on Team, 10 on Professional, and 15 on Enterprise.

A 20-agent Professional team gets 200 included resolutions per month. If the AI resolves 700 conversations, the overage is 500 resolutions at $1.50 each: $750 per month on top of the $2,300 base plan and $1,000 Advanced AI add-on. That brings the total to $4,050 per month, or $48,600 annually.

The counterintuitive problem: the better your AI performs, the more it costs. Every additional resolution beyond your allocation is a new line item. Teams that successfully optimize their knowledge base and build effective flows can find themselves penalized by their own success.

Workaround: set confidence thresholds above 85% to filter out low-quality resolutions that would count against your allocation without actually satisfying customers. Zendesk documentation suggests this yields 35 to 40% cost optimization. You can also set your account to pause AI functionality when you hit your limit, routing requests to live agents instead of accruing overages. But pausing means losing automation entirely until the next billing cycle.

Plug the memory gap.

One of the most common complaints from teams evaluating Zendesk competitors is that Zendesk's AI has no conversational memory across interactions. Every conversation starts from zero. A customer who chatted yesterday about a billing issue and returns today to follow up gets treated as a stranger. The AI does not know what was discussed, what was promised, or what resolution was attempted.

Zendesk's AI also lacks native access to past conversations, purchase history, or cross-channel behavior within the bot interaction itself. The agent workspace shows this context to human agents, but the AI operates from the current ticket only.

The workaround is imperfect but functional. Use custom ticket fields to store resolution outcomes and key context from previous interactions. Build triggers that populate these fields when tickets close. Then reference those fields in your answer flows so the AI can at least check whether a customer has an open or recently closed ticket on the same topic. It is not true memory. It is a structured workaround that reduces the worst repetition.

Address the channel gap.

Zendesk's AI capabilities differ significantly by channel. Advanced email agents cannot use generative procedures and must rely on dialogues instead. Agentic AI features are unavailable for email entirely. If email is a significant share of your support volume, you are running a less capable version of the AI on your highest-effort channel.

This gap is structural. Zendesk built its AI agent capabilities primarily for messaging and chat, then extended partial functionality to email. Teams that handle 40 to 60% of their volume over email often see dramatically lower automation rates than their chat-only metrics suggest.

The workaround: build dedicated macro libraries specifically for email workflows and use intelligent triage to route email tickets to specialized agent groups where manual efficiency is highest. Accept that email automation will lag chat automation in Zendesk and plan your staffing accordingly. Or evaluate platforms that treat email as a first-class AI channel rather than an afterthought.

When workarounds stop working.

Every workaround in this article addresses a real limitation with a reasonable hack. But workarounds compound. A team running intent grouping to dodge the 100-intent ceiling, custom field triggers to simulate memory, confidence thresholds to manage billing, and separate workflows for email versus chat is maintaining four parallel systems of complexity on top of the platform they are paying to simplify their operations.

There are signals that indicate you have moved beyond what Zendesk AI can structurally handle.

Your automation rate plateaus below 50% despite optimized content. Zendesk's own progression model shows that reaching 50% requires backend data connections and structured flows. If you have built both and are still stuck, the limitation is in how the AI reasons through multi-step problems, not in your configuration.

Your tickets require actions, not answers. Customers asking "where is my order" need a tracking number pulled from your OMS. Customers asking to reschedule an appointment need a write operation to your booking system. Customers disputing a charge need the AI to review transaction history and apply a policy. If the majority of your tickets require the AI to do something rather than say something, you need a platform built for action, not deflection. The difference between an AI agent and a chatbot is whether it can execute the resolution or only describe it.

Your cost per resolution is climbing, not falling. The per-resolution pricing model means costs scale linearly with automation volume. If your cost per ticket is not decreasing as automation increases, the pricing structure is working against you. Some Zendesk alternatives offer flat or per-agent pricing that rewards automation rather than penalizing it.

Your team spends more time maintaining the AI than it saves. When the hours spent building flows, tuning intents, updating knowledge bases, and managing billing thresholds exceed the hours the AI saves in ticket handling, the ROI has inverted. The platform is consuming more operational energy than it generates.

What an upgrade looks like.

The market for AI-native customer support platforms has changed substantially since Zendesk first bolted AI onto its ticketing system. Purpose-built AI platforms do not treat automation as an add-on layer. They build resolution capability into the core architecture.

The practical differences show up in three areas. First, action execution: AI-native platforms connect to backend systems and execute transactions mid-conversation, not just retrieve information. Processing a refund, rescheduling a delivery, upgrading a subscription. The AI completes the action rather than handing the customer to a human who completes it. Second, cross-channel consistency: the same AI capability operates identically across chat, email, voice, and SMS, without channel-specific limitations. Third, contextual memory: the AI maintains awareness of the customer's history across interactions and channels, so every conversation builds on the last rather than starting from scratch.

Purpose-built AI agents achieve containment rates of 50 to 70% on complex ticket types where legacy platforms with retrofitted AI reach 20 to 30%. That is not a marginal improvement. It is a structural difference in how the AI approaches resolution.

What is Lorikeet?

Lorikeet is an AI customer support platform that acts as a universal concierge across chat, email, voice, and SMS. Unlike legacy chatbots or bolt-on AI features, Lorikeet makes judgment calls and takes action: processing refunds, rescheduling appointments, managing billing, and executing complex multi-step workflows by integrating with existing systems like Zendesk, Stripe, and internal APIs. Lorikeet resolves issues end-to-end rather than deflecting customers to help articles or routing them to human agents for every action that requires a decision. See how Lorikeet handles the tickets Zendesk AI cannot.

Making the decision.

If your Zendesk AI automation rate is below 40% and you have not optimized your knowledge base, tuned your intents, or built structured flows, start there. The workarounds in this article can push you from 20% to 40% without changing platforms. That is real, meaningful progress.

If you have done all of that and you are still stuck, the ceiling is not your configuration. It is the platform. Zendesk AI was designed to answer questions from a knowledge base and route everything else to humans. When your customers need an AI that takes action, maintains context, and works consistently across every channel, you have outgrown what workarounds can fix.

Lorikeet exists for teams at exactly that inflection point. The same Zendesk integration that connects your current ticketing workflow means you do not have to abandon your existing infrastructure. You layer resolution capability on top of it, turning a system that routes tickets into one that resolves them.

Talk to Lorikeet about what your Zendesk AI is leaving on the table.