Skip to main content

Search

World's First 100% Zendesk BPO

Junior AI Agent Developer (Intern) — Candidate Profile & Role Requirements

1. Key Qualifications & Technical Skills

Non-Negotiables (Baseline Required to Start)

  • Zendesk Support Product Literacy: Must have completed Zendesk Support training and possess a solid working knowledge of core platform functionality: Triggers, Automations, Views, Macros, Ticket Statuses, Custom Fields, and SLAs. Understands how ticket lifecycle and event logic flow through the Admin Center.
  • Web Fundamentals (HTML & CSS): Writes clean markup, styles elements, and understands the box model. Needed for Zendesk Apps Framework (ZAF) app UIs, Web Widget customisations, and Help Center (Guide) theme edits.
  • JavaScript Fundamentals: Understands variables, functions, objects, arrays, conditionals, and loops. Can read a basic script, debug simple syntax errors, and explain what the code does.
  • JSON & REST API Literacy: Reads and writes valid JSON, understands nested objects/arrays, and knows how APIs work (GET vs. POST, headers, request bodies, status codes). Has tested at least one API endpoint using Postman or a browser console.
  • Strong Written English: Critical for writing clean generative persona prompts, formatting KB articles, and writing structured bug reports and documentation.
  • Basic Google Workspace Proficiency: Comfortable at a basic level with organising, filtering, and structuring data in Sheets (basic formulas, filters) and Docs.

Bonus (Accelerates Ramp-Up)

  • Zendesk Apps Framework (ZAF) Concepts: Familiarity with how ZAF uses standard HTML/CSS/JavaScript inside iFrames to build custom sidebar apps in the Agent Workspace using the ZAF Client API (client.get(), client.set()).
  • JSONata Basics: Understands JSONata as a query language for JSON. Can write simple expressions (e.g., $.customer.name or basic array filtering) to extract payload data in AI Agent webhook steps.
  • Generative AI & RAG Awareness: Conceptual understanding of Large Language Models (LLMs), why prompt formatting matters, and how Retrieval-Augmented Generation (RAG) works to answer user queries using KB articles.
  • Workflow Automation: Understanding of If/Then data routing in tools like Zendesk Integration Builder, Make, Zapier, or n8n.

2. Core Attributes & Soft Skills (The "DNA")

  • Reads Before Asking: The Zendesk Developer Portal and internal documentation are exceptional. Instinctively Googles, reads docs, and attempts a fix before escalating.
  • Attention to Detail: A missing bracket in JSON/JSONata, a wrong variable in a webhook, or a typo in a prompt silently breaks a bot. High precision is mandatory.
  • Analytical Decomposition: Instinctively breaks big problems into small ones. When a bot misbehaves, systematically checks: Is it the webhook, the JSONata parsing, the KB content, or the flow logic?
  • Genuinely Curious About AI: Plays with AI tools, follows industry updates, and asks "Why did the model generate that specific response?" rather than treating AI as a black box.
  • Structured Communicator: Writes bug reports and technical handovers in a clear pattern: Context → What happened → Expected behavior → Steps to reproduce.
  • Patient with Repetitive Testing: Bot QA requires executing the same flow repeatedly with slight variations. Willing to test thoroughly without cutting corners.
  • Comfortable in Ambiguity: Asks clarifying questions to define vague requirements rather than freezing or guessing.
  • Ownership & Empathy: Closes open loops without being chased and always evaluates solutions from the end user's perspective.

3. Core Responsibilities

What the intern will be trained to own independently after 4–8 weeks of onboarding:
 

AI Agent & RAG Optimization

  • Knowledge Base Structuring: Format, tag, and structure Help Center articles to optimize them for Zendesk AI Generative Replies and RAG.
  • Flow Building: Construct and edit action-oriented flows in the AI Agent Dialogue Builder using quick replies, variable capture, and conditional branching.
  • Knowledge Gap Analysis: Review conversation logs and failed resolutions to identify missing articles and draft content to fill those gaps.

Technical Development & Customizations

  • API & Webhook Wiring: Test API endpoints in Postman and configure webhooks inside AI Agent flows to pass and receive data.
  • Data Extraction: Write basic JSONata queries to pull dynamic values out of webhook JSON responses.
  • ZAF & Interface Customization: Build and maintain simple Zendesk Apps Framework (ZAF) apps using HTML, CSS, and JavaScript, and apply styling/template tweaks to the Help Center.

QA, Testing & Health Checks

  • End-to-End Bot QA: Test generative replies and flow actions across Web Widget, WhatsApp, and Email channels using varied user inputs.
  • Defect Logging: Record bugs with clear steps to reproduce, expected results, and actual API/payload logs.
  • Health Checks: Assist senior developers during pre-go-live and post-go-live bot health audits.

ASSESSMENT TRAINING:

Before attempting this assessment, please complete the training below. Each section is linked to a specific task in the assessment — read the articles and watch the videos for that section before answering the matching task. This should take you around 2 to 2.5 hours in total.

Everything you need to answer the assessment is in the material below. If you don't take the time to go through it, your answers will show that. Take it slow — this is how we learn on the job every day.

START HERE — Big-picture overview (watch first)

Watch this short video before anything else. They give you the context you need to understand the rest of the training.

Section 1 — For Task 1: Designing a Use Case and Procedure

What you'll learn: How AI Agents Advanced uses "use cases" and "customer request reasons" to understand what a customer wants, and how "generative procedures" and "dialogues" tell the AI Agent what to do about it.

Read these articles in order:

  1. About AI agents with agentic AI: https://support.zendesk.com/hc/en-us/articles/8966284087066
  2. Creating use cases for advanced AI agents to identify what customers are asking about: https://support.zendesk.com/hc/en-us/articles/9041901679130
  3. Best practices for creating use cases for AI agents: https://support.zendesk.com/hc/en-us/articles/8357733365402
  4. About generative procedures for AI agents: https://support.zendesk.com/hc/en-us/articles/10473649691418
  5. Creating generative procedures for AI agents with agentic AI: https://support.zendesk.com/hc/en-us/articles/8979864563610
  6. Best practices for creating and using generative procedures for AI agents with agentic AI: https://support.zendesk.com/hc/en-us/articles/9424547622298
  7. Examples of generative procedures for AI agents with agentic AI: https://support.zendesk.com/hc/en-us/articles/9424547984026

Watch this video:

Section 2 — For Task 2: Optimizing a Knowledge Base Article for AI

What you'll learn: How generative AI uses your Help Center content to answer customer questions (RAG), and how to write and structure articles so the AI Agent gives accurate answers.

Read these articles in order:

  1. Best practices: Preparing your help center for generative AI: https://support.zendesk.com/hc/en-us/articles/9067636151834
  2. Optimizing your knowledge content for generative AI: https://support.zendesk.com/hc/en-us/articles/4408845739162
  3. 5 strategies for building up your help center content for AI: https://support.zendesk.com/hc/en-us/articles/7849915550618

Watch this video:

Section 3 — For Task 3: API Integrations and JSONata

What you'll learn: How the Integration Builder lets AI Agents talk to external systems, and how JSONata is used to pull specific pieces of information out of an API response and save them as session parameters.

Read these articles in order:

  1. About the integration builder for advanced AI agents: https://support.zendesk.com/hc/en-us/articles/8357756844442
  2. Reformatting values in request parameters for advanced AI agents (JSONata basics): https://support.zendesk.com/hc/en-us/articles/8560203369242
  3. Integration builder FAQs: https://support.zendesk.com/hc/en-us/articles/8560379152282
  4. Adding custom functionality in AI agent conversation flows using fake integrations (practical JSONata examples): https://support.zendesk.com/hc/en-us/articles/8357749820570

Bonus resource (external, but the standard JSONata documentation): https://docs.jsonata.org/simple

Section 4 — For Task 4: Troubleshooting a Broken AI Agent

What you'll learn: How to test AI Agents, read conversation logs, use analytics to spot problems, and improve AI Agent performance based on what you find.

Read these articles in order:

  1. Testing conversation flows in advanced AI agents: https://support.zendesk.com/hc/en-us/articles/8357758879130
  2. Optimizing your advanced AI agent performance: https://support.zendesk.com/hc/en-us/articles/8357751836314
  3. About analytics for advanced AI agents: https://support.zendesk.com/hc/en-us/articles/8357756447898
  4. Analyzing advanced AI agent performance with the reporting dashboard: https://support.zendesk.com/hc/en-us/articles/9510024609178

Watch this video:

Section 5 — For Task 5: Written Reflection on RAG

The Task 5 question about RAG is fully covered by the "Best practices: Preparing your help center for generative AI" article you already read in Section 2 — no additional training required. Just make sure you can explain RAG in your own words after reading it.

ASSESSMENT:

Junior AI Agent Developer — Remote Assessment

Welcome, and thank you for applying. This assessment is designed to help us understand how you think, how you learn, and how you approach the kind of work we do every day at CX Experts. There are no trick questions — everything you need is either in the linked Zendesk articles or in your own reasoning.

Before You Begin — Please Read Carefully

What we're testing: Your ability to read documentation, understand it, and apply it. We are not testing memorization or how fast you can Google.

Format: Five written tasks. Submit your answers in a single Google Doc.

Time: 2 to 3 hours, self-paced. 

Ground rules:

  • You may use AI tools (ChatGPT, Claude, etc.) to help you learn concepts.
  • You may not copy-paste AI answers as your own. Every answer must be in your own words.
  • Complete the training material before starting each task.
  • If you don't know an answer, write what you do understand and where you got stuck. Honesty scores higher than guessing.
  • Please note: All assessment submissions must be received by 17:00 SAST on Monday, 28 September 2026. Late submissions will not be considered.

Task 1 — Designing a Use Case and Procedure (20 marks)

Why this matters on the job:
At CX Experts, most of the AI Agents we build use agentic AI. Instead of building rigid flowcharts, we tell the AI what to do in plain English through generative procedures, and the AI reasons through the conversation itself. Every client bot is a collection of well-designed use cases, each with a clear customer request reason and either a generative procedure or a scripted dialogue behind it. If a use case is written badly, the AI matches the wrong request, and everything downstream falls apart.

The scenario:
A fictional online bookstore, BrightBooks Online, wants a WhatsApp AI Agent to help their customers track orders. When a customer says something like "Where is my order?" or "Track my package", the AI Agent should:

  • Greet the customer
  • Ask for the order number
  • Ask for the email address
  • Look up the order using an API integration
  • Show the customer the order status
  • Ask if the customer needs anything else

Questions:

1.1 Write the use case for this scenario. Provide:

  • A short, descriptive use case name (3–5 words)
  • A clear customer request reason (around 120 characters, starting with "Customer wants to..." or "Customer is asking...")

1.2 Now write the generative procedure the AI Agent should follow to handle this use case. Write it as a numbered list of clear, direct instructions (for example: "Ask the customer for their order number", "Trigger the /Get-Order-Status API integration"). Include instructions for what to do if the order number is invalid or the API fails.

1.3 What session parameters would you save during this procedure, and why? Give at least two examples.

1.4 What should happen if the customer diverges from the expected path — for example, in the middle of asking for their order number, they ask "How long does shipping take?" How does an agentic AI handle this, and what does that mean for how you write the procedure?

1.5 Why is it important to end the procedure with a step that asks "Is there anything else I can help with?" instead of letting the conversation end silently?

Task 2 — Optimizing a Knowledge Base Article for AI (20 marks)

Why this matters on the job:
Zendesk AI Agents use a technique called Retrieval-Augmented Generation (RAG). When a customer asks a question, the AI splits your Help Center content into small chunks, finds the most relevant chunk based on meaning, and uses it to write an answer. If articles mix multiple topics, use vague language, or lack structure, the AI gives bad answers even when the information is technically there. Rewriting client Help Centers for AI performance is a huge part of your job at CX Experts, and it's often the difference between a bot that resolves 30% of tickets and one that resolves 70%.

The scenario:
Below is a real-world style article. It is poorly written for AI — it mixes several topics, has no structure, uses vague language, and would confuse generative replies.

"If you want to return a book you can do that but only within 14 days from delivery and you'll need your order number and the book must be in good condition otherwise we can't accept it. Refunds usually take 5 to 7 business days to reflect in your account. Damaged items are different, please email us at damages@brightbooks.co.za instead of using the normal return form. We don't refund shipping costs unless it's our fault, and international orders may have different rules depending on the country. Also if you paid with store credit you'll get store credit back not cash."

Questions:

2.1 Rewrite this article properly, following the Zendesk best practices you just read. Use clear headings, short sections, and simple language. You may split it into multiple articles.

2.2 Should this content be kept as one article or split into multiple articles? Justify your answer with reference to how RAG chunks content.

2.3 List five specific changes you made and, for each one, explain why that change will improve the AI Agent's generative replies.

2.4 What kind of article title would you give each of your rewritten articles, and why does the title matter for AI retrieval?

Task 3 — API Integrations and JSONata (25 marks)

Why this matters on the job:
Most real AI Agents don't just answer questions — they do things. Checking an order, looking up a policy, verifying a membership number, creating a ticket. All of this happens through API integrations built in the Integration Builder. The API returns data in a format called JSON, and to pull specific pieces of that data out and save them as session parameters, we use a query language called JSONata. Every day you'll be reading JSON responses, writing JSONata expressions, and handling what happens when an API fails. If you can't do this, you can't build a working AI Agent.

The scenario:
BrightBooks Online has an internal order-lookup API. When the AI Agent calls it with a customer's order number, the API returns the following JSON response:

{
"status": "success",
"customer": {
"name": "Thabo Nkosi",
"email": "thabo@example.com",
"orders": [
{
"id": "1001",
"status": "shipped",
"total": 450,
"tracking": "ZA12345",
"items": ["Long Walk to Freedom", "Born a Crime"]
},
{
"id": "1002",
"status": "pending",
"total": 220,
"tracking": null,
"items": ["Things Fall Apart"]
}
]
}
}

Questions:

3.1 Write the JSONata expression you would use to save the customer's name (Thabo Nkosi) as a session parameter.

3.2 Write the JSONata expression to save the tracking number of the first order (ZA12345) as a session parameter.

3.3 Write the JSONata expression to return a list of all order IDs.

3.4 In your own words, explain what an API integration does in an AI Agent, and why we use JSONata to extract values from the response. Write this as if you're explaining it to someone who has never used Zendesk.

3.5 Notice the top-level status field in the response. If the API returned status equal to "error" instead of "success" (for example, because the order number doesn't exist), how would you handle that inside your generative procedure? Describe:

  • What the AI Agent should say to the customer
  • How you would instruct the procedure to check the status (remember: procedures are written in natural language)
  • Why it's important never to let a failed API integration silently break the conversation

Task 4 — Troubleshooting a Broken AI Agent (20 marks)

Why this matters on the job:
AI Agents misbehave. A customer asks something reasonable and the AI matches the wrong use case, or the procedure runs the wrong step. When this happens, our clients want two things from us: a clear explanation of what went wrong, and a fix. Being able to look at a broken conversation, form a hypothesis, check the right places in Zendesk, and write a clean bug report is one of the most valuable skills in this role.

The scenario:
A customer messages the BrightBooks WhatsApp AI Agent:

"Hi, I need to return my book, it arrived damaged."

The AI Agent replies with generic shipping information about when parcels are delivered, instead of the returns process. The customer gets frustrated and asks for a human agent.

You investigate the AI Agent and find the following:

  • Use case: "Order Delivery"
    • Customer request reason: "Customer wants to know when their order will arrive or has arrived"
    • Has a generative procedure that gives shipping and delivery information
  • Use case: "Returns"
    • Customer request reason: "Customer wants to return an item"
    • Has an API integration attached to look up the order
  • There is a Help Center article titled "Damaged Item Policy" that clearly explains what to do.
  • The conversation log shows the AI Agent matched the "Order Delivery" use case, not "Returns".
  • The API integration attached to "Returns" did not fire.
  • There is no separate use case for damaged items.

Questions:

4.1 Write a structured bug report for this issue using this exact format:

  • Context: what the scenario is, what channel, what feature
  • What happened: the actual behavior
  • Expected behavior: what should have happened
  • Steps to reproduce: numbered
  • Suspected root cause: your best guess and why

4.2 List three specific things you would check inside AI Agents Advanced to diagnose this issue. For each one, say where you would look (for example, which section of the AI Agent workspace, which log, which setting).

4.3 What do you think is the most likely root cause? Explain your reasoning based on the customer request reasons and the customer's message.

4.4 Suggest one specific improvement — to the use cases, the customer request reasons, the procedure, or the KB — that would prevent this exact problem from happening again. Be concrete: don't say "improve the use cases," say what you would add or change.

Task 5 — Written Reflection (15 marks)

Why this matters on the job:
Two of the most important skills in this role are (1) being able to explain complex things simply to non-technical stakeholders, and (2) being a self-driven learner. This section tests both.

Questions (max 150 words each):

5.1 Explain how Retrieval-Augmented Generation (RAG) works in a Zendesk AI Agent, as if you're explaining it to someone who has never heard of AI. Use plain language. Avoid jargon unless you define it.

5.2 Describe a time you had to teach yourself something technical. What was your process? How did you know when you had actually learned it, versus just memorized it? What would you do differently next time?

2 Day Workshop Itinerary:

(TBD)
 

Powered by Zendesk