Qlik Community Challenge: Put Agentic AI to Work

[SarahUrbiss]

We're challenging our Qlik Community: put agentic AI to work. 

We saw lots of great posts in our recent Make Your Data Work for AI forum. We’re building on that momentum with a new board focused on the next step: driving real impact with agentic AI.

This is the place to get started with agentic capabilities, try them out, and share your experience with the rest of the Community.

Here's the challenge:

1. Get started. Toggle on Qlik Answers, activate Qlik’s MCP Server, or get started with whichever capability fits your work, and see how easy it really is.
2. Build your first workflow. Identify a use case and put the capabilities to work on a real problem or task.
3. Share your experience. Tell the Community what you learned, what worked, and what surprised you.

Ask questions, share screenshots (obscuring any sensitive data), get ideas, and help your Qlik Community peers on their journeys.

You'll earn NEW Community badges along the way: Agentic AI Adopter when you share your first workflow or use case, and Agentic AI Expert when you share your learnings and results.

Plus, we’ll be awarding some fun surprises and new Qlik swag along the way!  

[rdm6532]

We have setup an embedded assistant which works alongside when of our most utilized dashboard resources. From the user's prospective, the assistant is called with a simple button and exists along side the dashboard.

[priscilarubim]

Hi everyone!

For this challenge, my colleague @marksouzacosta and I decided to go all in: instead of a single workflow, we built an entire governed analytics solution end to end and we used Claude as our AI teammate at every step along the way.

We are turning the whole build into a free YouTube series on the Data Voyagers channel, and we wanted to share the concept here first, because it is exactly the kind of "real problem, real workflow" this challenge is asking for.

The problem we picked

Education data is messy, sensitive, and full of governance requirements, which makes it a great stress test for agentic AI. We used the Ed-Fi ODS sample database (the Grand Bend dataset, a realistic K-12 data model maintained by the Ed-Fi Alliance) as our raw source, running on-prem in Docker.

The story we are telling with the data: does attendance drive performance? We combine course grades with school- and section-level attendance to build a Student Performance dashboard that answers that question. It is not a toy example but a real analytics product, with naming standards, documentation, and governance built in from day one.

The architecture, in one line

PostgreSQL (Docker) → Qlik Data Gateway - Data Movement → Snowflake → Qlik Cloud Script Apps → Parquet → Data Product (Glossary + Semantic Model) → Qlik App → Reports → Automation

Four governed spaces (landing, storage, development, production), one Qlik Cloud tenant, and Claude sitting alongside us at every stage. Not as a gimmick, but doing real work: drafting pipeline YAML, writing metadata, building glossary terms, reviewing master items, and helping design the dashboard.

Why this matters for the challenge

Agentic AI is most valuable when it is applied to the unglamorous, repetitive, "someone has to do it" work that governance requires: field naming, metadata, documentation, glossary terms, lineage checks. That is precisely where we pointed Claude, and where we found it earned its keep the fastest.

We are not claiming Claude replaced our judgment anywhere in this build. Every "Claude moment" in the series ends with a human review pass on screen. That is the model of responsible AI-assisted work we wanted to teach, not an afterthought.

What is coming

The series covers 10 episodes, from connecting to the source database all the way to a published, reportable, automated dashboard:

• EP0 Introduction and architecture
• EP1 Connecting to Ed-Fi
• EP2 Moving data with Declarative Pipelines
• EP3 Extracting data with Script Apps
• EP4 Enriching datasets
• EP5 Business Glossary
• EP6 Data Products and Semantic Models
• EP7a/7b Building the dashboard
• EP8 Reporting
• EP9 Qlik Automation

Everything will be reproducible: the Docker image, the scripts, and the actual prompts we used will be shared in a companion GitHub repo, so anyone following along can rebuild the whole thing themselves.

We will launch EP0 and EP1 together, then release weekly. If you build or govern Qlik solutions and have ever wondered where agentic AI actually pulls its weight (and where it does not), this series is for you.

My colleague @marksouzacosta is posting right after me with a closer technical look at the Claude workflows we used: the pipeline YAML authoring, the MCP server tools, and a few surprises we ran into along the way. Go check that out too, and let us know in the comments what you would want to see us tackle with Claude next.

[marksouzacosta]

@priscilarubim covered the big picture of our "Building on Qlik with Claude" series - a full governed analytics build on the Ed-Fi ODS sample data. I wanted to follow up with the technical side: specifically where we pointed Claude, what tools it used, and what we learned.

The two Claude "surfaces" we used

• Claude Code, working against the repo and the Qlik MCP Server tools, for anything that touches Qlik Cloud directly: creating datasets, editing load scripts, building the business glossary, managing the data product, and configuring automations
• Claude Design, for the visual layout concept behind the dashboard sheets, before we built anything in the app

Neither one runs unsupervised. Every step ends with us reviewing the output on screen before it goes anywhere near a published space. That review step is not decoration - a few of our "learnings" below are things Claude got directionally right but wrong in the details, which is exactly why the review matters.

Where the agentic workflow showed up

Declarative Pipelines as YAML. Qlik's newer Declarative Pipelines let you define, configure, and deploy a Talend Data Integration pipeline entirely as YAML instead of clicking through a canvas table by table. That is a natural fit for an LLM: Claude Code drafts the pipeline definition (table selection, field mappings, schedule), we diff and review it like any other code change, then deploy it. Version-controlled, diffable, code-reviewable - the pipeline lives in git next to everything else.

Bulk load-script edits via Qlik-CLI. Once data lands in Snowflake, we extract it through one Script App per dataset. Applying a naming standard consistently across dozens of fields, across many scripts, is exactly the kind of repetitive, error-prone task we handed to Claude - it is fast, but you still spot-check the diffs.

Metadata and glossary, sourced not invented. This is the part I would flag as the real "aha": we did not let Claude invent field descriptions or business terms. We had it pull definitions from the official Ed-Fi data handbook and model documentation, and only then write that metadata into the Qlik Cloud datasets and the Business Glossary through the MCP Server. Grounding the AI's output in an authoritative source, instead of its own guess, made the difference between metadata we could trust and metadata we would have had to double-check line by line anyway.

Data Products and Semantic Models. Claude Code plus the MCP Server also builds the Data Product that bundles the governed datasets and the glossary, including the relationships, measures, and dimensions of its Semantic Model. The payoff shows up downstream: once the dashboard app consumes that Semantic Model, those measures and dimensions land as master items automatically on reload - real time saved, not hypothetical.

Automation, mostly for the boring-but-critical parts. Rather than an automation driving every reload, we wired the reload chain natively with Qlik Cloud task triggers (dataset -> script app -> parquet -> dashboard). Qlik Automation, built and iterated with Claude Code and the MCP automation tools, sits on top for what it is genuinely good at: monitoring the pipeline run and handling failures and notifications end to end.

What surprised us

• Claude was most valuable exactly where governance work tends to get skipped under deadline pressure - naming, metadata, documentation. That is the highest-leverage place to point agentic AI, more than the flashier steps.
• Treating the pipeline YAML and load scripts as regular code (diffed and reviewed, not just accepted) caught small mistakes early that would have been much more annoying to find after a reload chain was live.
• Grounding metadata generation in an authoritative source document turned out to matter more than the model itself - the quality ceiling was set by what we fed it, not by the AI's fluency.

Try it yourself

We are releasing the whole build as a YouTube series, and everything will be reproducible: Docker image, scripts, and the actual prompts we used, in a companion GitHub repo. If you want to see agentic AI applied to a full governed pipeline rather than a single isolated demo, keep an eye on Priscila's post above for the series link as episodes go live.

Happy to answer questions in the comments about the MCP Server tools, the Declarative Pipeline workflow, or anything else from the build.

Regards,

Mark Costa

[igoralcantara]

We have been using Qlik Answers, Qlik Predict and Qlik MCP in many scenarios, so it is hard to pick one. So, let me pick one that is recent and very insightful. Unfortunately, I cannot share screenshots since is a customer facing use case.

We were implementing some complex use cases for Qlik Answers. They were related to education.

A few findings:

- Really spend time designing your Master Dimensions and Master Measures. Make sure the descriptions on it are detailed enough. This improves Qlik Answers and MCP quality a lot.

- Make sure your master items have the proper tags. 

- Have the Master Items defined and reviewed BEFORE indexing the application.

- Review all items on the AI Optimization. After the first indexing process, Qlik will fill the semantic meaning using AI based on the field name only. You MUST review this carefully. Quite often the description is vague or wrong and that will make Qlik Answers produce inconsistent responses. We tested multiple users asking the same question on the same app and each one got a different number because of that reason.

- In AI Optimization, Synonyms don't seem to matter. At least not yet. We tested a few times and we did not see any change in result. Use it anyway, once this is fixed, your apps will be ready.

- If the same value is present in multiple fields, Qlik Answers will get confused. In one of our tests, for example, the values AM or PM were present in many fields. Even after we defined the synonym and the right description in Master Items, when we ask questions that mentioned "morning", each answer was different and most of them, wrong.

- In that case, a good work on the AI Optimization > Fields & Master Items was what fixed the problem. 
Uncheck Data Value Lookup for when a field should not be considered and create groups that reflect that. For example, in our case, we had a group for Arrival, one for Departure. So, when asking the questions about Morning or Afternnon, Qlik Answers were smarter to get the context.

Another finding: For complex use case scenarios, an assistant with a curated knowledge base + app works better.

One final though: people are using more LLMs now as their first interface. So, before designing an application, interview users and ask them: what questions would you like to ask this app? Start from there and use that as the center of gravity of the project.

[FOliv]

We have had Agentic AI enabled for several months and have gradually started incorporating its capabilities into our daily processes.

One of the key solutions we have been using is Qlik Answers, which allows business users to analyze application data more efficiently and obtain insights much faster. This has helped reduce the time required to find relevant information and supports quicker, data-driven decision-making.

We also make use of triggers and feeds to gain immediate visibility of unusual or unexpected values. This enables us to proactively identify anomalies, investigate their root causes, and determine whether they are expected business scenarios or issues that require intervention.

More recently, I have started exploring the Automation Agent to better understand how it can support some of the challenges and use cases we are currently trying to solve. We see significant potential in leveraging AI-driven automation to further improve efficiency and streamline repetitive processes.

In addition, we use AI-generated content to improve data discoverability and usability across our platform. For example, AI helps us automatically generate descriptions for datasets, data models, and other available items, making it easier for users to understand the purpose and context of the information available to them.

Overall, our goal is to leverage the different AI capabilities available within the platform to improve user experience, accelerate analysis, enhance data governance, and enable the business to extract value from data more effectively.

[hugo_andrade]

@kauenascimento and myself are working on Qlik Answers since the Early Access.

We are integrating the Agents into our Mashups! Simply AMAZING! We don't have to spend time tweaking the Agent code, we can simply focus on the Qlik Embeds and let this great agent do what is does best: analytics! 

[JandreKillianRIC]

Putting Qlik MCP to Work: From Prompt to Mashup and Board Pack

Hi Qlik Community,

I've been working with the Qlik MCP Server and Claude AI across several Qlik Cloud tenants, and I've put everything together in one place: MCP-Showcase 

Here's my contribution to the challenge in two parts.

---

Part 1 - Agentic AI Adopter: My First Workflow

Getting started
Activating the Qlik MCP Server and connecting it to Claude AI was the easy part. Once connected, Claude can see the apps, data model, master items and fields, and can query live data directly. No exports, no copy-paste, no middleman.

The use case: building a mashup from a single prompt
Building a Qlik mashup has always been a developer-heavy job: HTML, CSS, JavaScript, object IDs, layout, styling, testing, and repeat. I wanted to see how far agentic AI could take that.

The workflow:

1. Point Claude at a Qlik Cloud app through MCP
2. Describe the mashup I want: audience, KPIs, layout and style
3. Claude explores the app, pulls the data it needs and builds the mashup
4. I review and refine with follow-up prompts

The full build is recorded on the showcase ("MCP Mashup Build - Claude AI"), along with a second recording of a data lineage build.

Try it yourself
I've published two prompt libraries you can copy straight into Claude:

• Output Prompts - generate HTML, PowerPoint, Word, Excel and multi-format exports from Qlik data
• Power Prompts - advanced patterns for getting more out of Qlik MCP

Swap in your own app name and go.

---

Part 2 - Agentic AI Expert: Learnings and Results

The result that matters most: time saved
Mashup development used to take a lot of time. With Qlik MCP and Claude, the time saved is unmatched: I now get a working, styled, data-connected mashup in a fraction of the time, and spend my effort refining instead of building from scratch.

The same goes for board packs. Monthly PowerPoint reports that used to mean exporting, formatting and writing commentary by hand now come out of a single prompt. The Board Pack on the showcase was produced entirely with Qlik MCP and Claude, with no templates and no post-processing.

Real outputs, not mock-ups
Everything in the "Actual Examples" section of the showcase was generated live from Qlik data:

• Vendor Defect Network
• Defects Dashboard
• Defective Products Dashboard
• Board Pack (PowerPoint)

What I learned

• Your data model is your AI's foundation. Clean field names, clear master items and good descriptions make a huge difference to output quality. If a human would struggle to understand your app, the AI will too.
• The prompt is the new template. A well-written prompt is reusable across apps and clients, which is why I've shared my prompt libraries.
• Iterate, don't over-specify. Start with a clear goal, let the AI build, then refine. That beats trying to write the perfect prompt up front.
• Output format is no longer a constraint. The same Qlik data can become a dashboard, a mashup, a board pack or a Word report, depending on who's asking.

What surprised me
How much of the "last mile" disappeared. The gap between insight in Qlik and a polished deliverable in front of a stakeholder used to be where most of the hours went. That's now where agentic AI adds the most value.

What's next
I'll keep adding examples and prompts to the showcase. If you're starting out with Qlik MCP, feel free to use anything there, and drop your questions below. Happy to help.

All data in the showcase is demo data. No real customer or business data is used.

Jandre Killian

[MatheusC]

I already mentioned our experience with Qlik Answers in the "Share Your Experience" thread, so here I’ll add to that and share a specific use case.
Here, a prompt was created asking Qlik Answers to generate a dashboard:

I provided the following instructions:

Act as a business intelligence analyst experienced in application development; based on the application data, propose and structure an executive dashboard.

Consider the following metrics, views, and data cross-references:

1. Executive View and Key KPIs:
- Total Sales (R$)
- Total Sales Quantity
- Total Current Inventory Value (R$)
- Inventory Quantity

2. Trend and Time-Series Analysis:
- Monthly trend comparing Sales (R$) vs. Inventory Level (R$)

3. Drill-down and Ranking Factors:
- Enable analysis by Product
- Present the Top 5 products with the greatest positive impact on revenue

And this was the result Qlik Answers delivered:

[mans]

I used Qlik’s agentic AI capabilities to build a workflow that detects unusual financial transactions and potential compliance risks. The AI agents analyzed 30 days of governed data, applied custom compliance rules, and surfaced anomalies with clear explanations. MCP Server made secure data access simple, and Qlik Answers provided transparent reasoning for every flagged item.
The workflow reduced manual review time and highlighted patterns I wouldn’t have spotted manually.
I was surprised by how quickly the agentic workflow came together and how explainable the results were.
Next, I plan to extend this into automated alerts and case‑management triggers.

[GTiscareno]

Using Qlik Agentic, has facilite several task one is organize customer source for help, for that we embed Qlik Anwser 

Also now that we Qlik Anwser can run automation customers had simplify several process.