FRENContact us

· By HeyData team

AI assistants and Qlik: why they get your numbers wrong, and how to prepare your apps

Connecting an AI assistant to Qlik Cloud now takes a day: Qlik’s MCP server gives Claude or ChatGPT access to your apps, data products and glossary, with each user’s own permissions. Qlik Answers does the same inside the platform.

The first questions come quickly, and with them a common disappointment: the assistant answers with confidence, but not always correctly. The cause is rarely the language model. It is almost always in the app.

Why the assistant gets it wrong

An assistant only knows what your app says about itself: field names, tables, and perhaps descriptions. Faced with a field called CA_HT_N1_cumul, three date tables or a measure that every sheet calculates its own way, it does what a newcomer without documentation would do: it guesses.

It picks a field that “looks like” the right answer, writes its own revenue formula, and ignores an exclusion everyone on the team knows about. The result is plausible, well worded, and wrong.

So the MCP server does not create value: it reveals the value already in your model. Two apps connected to the same assistant will give answers of very different quality depending on how they were prepared.

First lever: Master Items

Qlik documents it clearly: Qlik Answers relies on master dimensions and measures in preference to raw fields. A “Revenue” master measure with its official formula is a formula the assistant no longer has to invent.

  • Create a master measure for every indicator your users actually ask about, with a business name.
  • Create the matching master dimensions, especially for time axes and hierarchies.
  • Organise them with tags: they also help the people who maintain the app.

Second lever: descriptions and the glossary

A name is not enough. For Qlik Answers, a Master Item’s description is its primary meaning. A good description says what the indicator measures, its scope and what it excludes, in one or two sentences.

The business glossary goes further: a term linked to a master measure brings its definition and its synonyms, which the assistant uses to understand your users’ vocabulary (“revenue”, “sales”, “turnover”). Two rules worth knowing:

  • if the description and the term definition disagree, Qlik Answers does not blend them and keeps the Master Item description: keep them consistent;
  • Master Items and their terms are read at indexing time: a glossary change needs a reindex to take effect.

Writing a glossary from scratch has sunk many governance projects. Today, AI can draft a first version from the app: your experts only have to validate it, which takes five to ten times less time than writing it.

Third lever: a readable data model

The rest is modelling hygiene that every Qlik developer knows, but here it directly affects answer quality:

  • explicit field names, with different fields renamed so their purpose is obvious;
  • technical fields (keys, identifiers, load timestamps) hidden, for example with the % prefix and HidePrefix;
  • no synthetic keys, which usually signal a modelling error;
  • dates loaded as real dates, not as text;
  • no unnecessary personal data in the model.

Finally, the app must be made available in Qlik Answers to be indexed; it is then reindexed at every reload.

Measure before you fix

Where do you start when you have dozens of apps? That is why we built the AI Readiness Monitor: a Qlik app that reads your tenant’s APIs and scores each app out of 100, based on 17 checks covering the levers in this article. The action plan sorts fixes by points to gain, and the next assessment confirms each improvement.

The score translates into a maturity level, from “Readable” (the assistant explores the model on its own) to “Governed” (it qualifies source, freshness and reliability). Our full approach, on-premise or with Qlik Cloud MCP, is presented on our Services page.

Key takeaways

  • The language model is not the problem. Answer accuracy is decided in the app, not in the choice of assistant.
  • Master items first. Every indicator users ask about needs its master measure, with a business name.
  • Consistent descriptions and glossary. The description gives the meaning, the glossary brings synonyms; in case of conflict, the description wins.
  • A clean model. Explicit fields, hidden technical fields, no synthetic keys, real dates.
  • AI drafts, a human validates. Descriptions and glossary carry business meaning: an expert validates them before publication.

Sources