Early January Google released a white paper on AI Agents. It's a 40 page overview of AI Agents (or Agents in short) that explores how agents can extend the raw capabilities of large-language models to not only read, reason and react to input, but to perform actual actions on that data.

Where most of the world looks at AI and LLM through the window of ChatGPT, within the Customer Care and Zendesk world we also know AI from the point of view of chatbots and agent copilots within the Zendesk platform.

The white paper almost serves as a how-to and motivation to start building AI Agents and Copilots within Zendesk, being it not that Zendesk has already done this.

So for the focus of this article it seemed like a fun exercise to compare the approach of the white paper to Zendesk approach in building AI Agents both as a direct comparison on how Zendesk's approach differs, as well as from an alignment standpoint to see how close Zendesk's AI Agents adhere to Google's theoretical approach.

Intro

Reading through the white paper there's a few sentences I highlighted as being highly relevant for Zendesk.

Tools bridge this gap, empowering agents to interact with external data and services while unlocking a wider range of actions beyond that of the underlying model alone. (p7)

I read this in two ways: one way is that by providing the LLM with external sources like Help Center, website or tickets, it can give more contextual and better responses. Secondly, Zendesk's Flow Builder and Ultimate's Bot platform are tools we can use to turn a raw AI model into a tool and feature you can actually use without building the entire tech-stack ourselves.

I think this is key to Zendesk's AI Agents' success. Combining (in their case) OpenAI's model with their own Zendesk AI model for intents and the unique knowledge of each customers' tickets and knowledge base creates powerful solutions that give better responses to their end-users than a Google search or conversation with a world-model like ChatGPT would.

By integrating retrieval-augmented generation (RAG), agents can connect with fresh data sources to extract valuable insights and uncover hidden patterns. (p29)

Zendesk AI is not just about bots. Zendesk QA and its autoQA tool can surface data across all your tickets and detect trends in quality, churn, expertise. Similar Ultimate's reporting can detect knowledge gaps in your data sources or surface new use-cases.

At the heart of an agent’s operation is the orchestration layer, a cognitive architecture that structures reasoning, planning, decision-making and guides its actions. Various reasoning techniques such as ReAct, Chain-of-Thought, and Tree-of-Thoughts, provide a framework for the orchestration layer to take in information, perform internal reasoning, and generate informed decisions or responses. (p40)

For me this is the big difference between using a tool like Zendesk AI and Ultimate, or "just putting ChatGPT in front of your data". Zendesk has done all the work and thinking to make sure their AI Agents react in a way that is useful for CX and it forces those bots to use your data to respond. And it abstracts all that technical stuff behind a UI that makes it easy to manage your bot without worrying about the underlying models, bias-prevention, validation e.a.

Extensions provide a way for agents to perceive, interact, and influence the outside world in a myriad of ways. (p18)

Bot the Zendesk Bot as well as Agent Copilot offer ways to integrate with external tools. Zendesk Bots offer hybrid flows that combine generative replies with raw API data, and Agent Copilot offers actions that allow you to integrate with external tools to react to customers, or update those tools based on customer input.

Comparison

How do Zendesk AI agents compare against those in the white paper?

In this first section I'm going to compare the AI Agents as described by Google to Zendesk's AI Agents and Copilot. Where is the approach the same, where do they differ?

Even though Google talks about AI Agents, what they describe can apply to both Zendesk's Bots as well as Agent Copilot since both are features that leverage AI to automate processes and assist your team.

Core Functionality

The white paper takes a very high-level approach to AI Agents. It looks at them in a generic way approachable to a myriad of use cases.

• Focus on autonomous task completion using external tools, orchestrated reasoning, and decision-making frameworks.
• Leverage cognitive architectures (e.g., ReAct, Chain-of-Thought) for reasoning and iterative task planning.
• Use tools like Extensions, Functions, and Data Stores for real-time data interaction and multi-step task execution.
• Support advanced use cases such as retrieval-augmented generation (RAG) for dynamic, data-driven responses.

Zendesk's AI-powered self-service tools (Bots) and agent-assist tools (Copilot) are there solely to enhance customer experiences.

• Bots: Automate customer interactions, resolve common issues, and escalate complex queries when necessary.
• Copilot: Provide agents with real-time suggestions and contextual data to improve productivity and accuracy.

Integrate seamlessly with Zendesk’s ecosystem (e.g., Support, Guide) and can index external resources like websites or use APIs to integrate with other platforms. They're uniquely tailored for customer service efficiency, enabling companies to deliver scalable and personalized experiences.

Target Audience and Use Cases

According to the white paper AI Agents have a broader scope beyond customer support, applicable to domains like travel planning, financial services, and research.

They can be targeted at developers and technical teams creating custom solutions for diverse workflows and are ideal for businesses seeking flexibility in designing multi-tool agents for varied, complex tasks.

Zendesk's AI offering however is focused on customer service teams and administrators looking to improve productivity and CX.

• Bots: Empower self-service for customers across multiple channels.
• Copilot: Assist human agents in resolving customer queries faster with intelligent recommendations.

They're suitable for businesses aiming to scale AI-driven customer support, particularly in multilingual, omni-channel environments but are not designed or focused on developers or highly technical teams.

Ease of Use and Implementation

Google assumes that for now AI Agents require significant technical expertise to implement and customize. There are pre-built frameworks to build personalized AI Agents for your use case but they require developer action and technical know how to deploy and build them.

Building one yourself provides high flexibility at the cost of simplicity.

Zendesk however chose to go for a highly user-friendly, offering no-code or low-code interfaces for configuration and deployment.

• Bots: Easy to set up workflows for self-service and escalation.
• Copilot: Instantly integrates with agent dashboards for seamless support.

They enable non-technical users to implement and scale solutions with minimal overhead but you're limited by the features enabled in the platform. (I struggled with this phrase since Zendesk's AI Agents are anything but limited in functionality, but you can only deploy and configure what Zendesk offers and can't build additional functionality within the platform, that's their prerogative, or, in other words that's what we pay them for).

AI Capabilities

AI Agents require advanced reasoning frameworks (e.g., Tree-of-Thoughts, Chain-of-Thought) for dynamic problem-solving. Using these bare-bones and low level frameworks offer extensive customization options for complex workflows across diverse domains if you provide them the required context and sources to do so.

Zendesk's bots however are solely optimized for customer and employee service applications:

• Bots: Excel in intent recognition, FAQ resolution, and escalation handling.
• Copilot: Enhances real-time agent support with contextual suggestions and knowledge retrieval.

They are focused on delivering efficient, AI-driven CX enhancements rather than general-purpose problem-solving. Which also means their weak spot is using them for anything outside of that CX and employee service niche.

Strengths and Weaknesses

Overall we can compare custom build and generic AI Agents against Zendesk's CX focused bots as follows:

Final Thoughts

While the white paper’s AI agents focus on broad applications with extensive customization, Zendesk Bots and Copilot are purpose-built for customer service automation. They excel in improving CX for Zendesk users, providing a fast-to-deploy and highly intuitive platform. The choice depends on whether the goal is advanced, multi-domain AI workflows or targeted customer service optimization.

Making it real

Putting the theory to the test

If we assume Zendesk Bots and Copilot are a purpose-built version of the AI Agents described in the white paper, we can use the white paper to evaluate the overlap, strengths, and potential limitations of Zendesk's offering.

Core Alignment

Zendesk Bots and Copilot align closely with the white paper’s principles:

• They are build on top of foundational AI models and offer external tools for action, such as customer-facing chatbots (Bots) and AI-driven co-pilots for agents.
• They combine multi-turn conversations, context, internal models and instructions that is very similar to the "cognitive architecture" as described in the white paper.
• Bots function as task-oriented agents, autonomously handling repetitive, structured queries, while Copilot enhances human agent performance with contextual support and insights.

Where the white paper’s agents utilize tools like Extensions, Functions, and Data Stores to solve diverse, domain-independent tasks (e.g., retrieving documents, performing complex computations), Zendesk’s focus remains tightly bound to customer service workflows, making adaptation to non-CX domains (e.g., supply chain management) challenging without major modifications.

But I don't see this as a really lacking feature since their sole focus on CX is by design. Zendesk is a CX focused company so it would be illogical for them to build an AI Agent that manages my todo list and schedule for example.

Structure

Take a look at the diagram below.
We could easily imagine this to be a high-level overview of Zendesk's AI Agents.
The orchestration layer is how Zendesk instructs their bot. It contains your personae (tone of voice), its instructions to act as a self-service bot that replies to customers, intents, protection against biases or unwanted behavior,...

The Model is their Zero-Training AI model that underlies Ultimate's Use cases, the prebuilt intents for Zendesk AI and the custom entities we add to Intelligent Triage.

And Tools are the indexed websites and help centers, the actions that call external APIs and the custom flows we build out in the admin center.

Extensions and external sources

The white paper describes Data Stores that act as dynamic sources of knowledge, integrating external documents or unstructured data in real time.

Zendesk Bots and Copilot primarily rely on the Zendesk knowledge base and indexed websites. They don't really have a way to process other structured data, unless we make use of API integrations. Importing data into Zendesk is also done on a regular basis, and not in real-time as described in the white paper

However, the way they describe AI Agents, extensions and APIs is almost a 1:1 match with how we see Agent Copilot, Procedures and Actions being used in Zendesk.

Zendesk also kinda uses (as far as I know) the concept of functions. Google describes this as the bot reverting back to the end-user to execute certain actions. Zendesk has this for example as agents approving Agent Copilot suggestions, or tweaking responses. But it's limited to those instances.

Similarly Zendesk abstracts away a lot of the complexity that we should think about when building such AI Agents ourselves. The white papers' entire section on functions and use cases can almost be read as a view behind the curtain, instead of something we as users and admins of Zendesk Bots should worry about.

What Zendesk Improves Upon

Ease of Deployment

While the white paper’s agents offer flexibility, their proposed setup requires technical expertise and iterative design.

Zendesk Bots and Copilot however provide a turnkey solution with no-code/low-code interfaces, enabling businesses to deploy automation with minimal technical resources. It's Zendesk's Team that worries about technical expertise and iterative design, we just see the functional outcome.

Purpose-Built CX Features:

Zendesk Bots are optimized for common CX intents like FAQs, troubleshooting, and escalations. Agent Copilot enhances agent productivity with real-time recommendations, contextual data, and article suggestions. They seamlessly supports multi-channel CX (chat, email, tickets) with pre-configured workflows designed for customer service needs.

Google's concept of Agents is a more high-level approach where all these benefits Zendesk has need to be build from scratch.

Human-AI Collaboration:

Zendesk’s co-pilot model excels in AI-assisted human interactions, ensuring smooth handoffs and personalized support during escalations. This collaborative focus is more mature than the autonomous, standalone frameworks discussed in the white paper.

So, let's conclude this article.

If you've read the original 40 pager you might have noticed I skipped all the code examples and did not touch the concepts of Chain-of-Though, Tree-of-Thoughts or ReAct. There's a simple reason for this. While I have a limited understanding on how the Zendesk and Ultimate bots run, my knowledge on the underlying models and architecture is limited. I implement these tools, but do not build them.

So the focus of this article was mainly on comparing Zendesk's finished product to the concept of AI Agents as Google described them.

Zendesk Bots and Copilot represent a purpose-built implementation of the white paper’s agent concepts, confirming the benefits of combining AI models, external tools, and orchestration layers, and applying them real-world use cases.

Zendesk's AI Agents excel in customer service automation by offering pre-built models and instructions, ease of deployment, and human-AI collaboration. However, its scope is narrower, lacking the multi-domain flexibility, advanced reasoning capabilities, and dynamic data integration described in the white paper.

For businesses focused on customer service, Zendesk offers a refined, ready-to-deploy solution. For broader, multi-domain applications, the white paper’s generalized AI agent framework remains a more versatile foundation.

But for me, one thing is clear: you'd need to have very good reasons to build your own, and in my opinion you're better off spending that time on optimizing your Zendesk AI Agent, processes and knowledge base.