Media

Goodbye dashboards, hello dialogue (probably)

What we told the delegates at DataFest Johannesburg in June 2026 about dashboards, chat and the work that sits between them, and what has changed in the four months since.

On 10 June 2026, one of our consultants presented at DataFest Johannesburg under the title “Goodbye Dashboards, Hello Dialogue: The Future of Analytics Is (probably) a Conversation”. The word “probably” was there deliberately, because the talk covered what we have learnt taking several clients from reports to chat-based analytics, and in a few cases partway back again.

We started with a claim we thought all would agree with: in any room of data professionals, at least one person has already exported something to Excel that day, and we are not exempt.

How we got here

For decades, business reporting has been moving information closer to the people who use it. Paper reports delivered the truth once a month. Excel let people edit the truth locally (and email three versions of it around). BI dashboards made it interactive, and chat tools such as Teams and WhatsApp now carry the discussion about it into the places where work happens.

Each of those steps moved information closer to the user. The step we focused on in Johannesburg moves the user closer to the decision they have to make, which is a harder problem and only partly a data one.

A very expensive way of starting a Teams conversation

One line from the day was “Help me make a decision, I don’t actually want another dashboard”, attributed to a CFO, somewhere, every Monday morning. Depending on whose research you read, anywhere from a quarter to more than half of corporate dashboards go largely unused, and most organisations make decisions using a small fraction of the data they hold. This is often referred to as dark data. The rest sits in storage adding cost, risk and lost opportunity.

Dashboards are good at answering the questions someone anticipated when they built them. They struggle with the follow-up that starts “yes, but why” and “ok, I see the data but guide me to a decision”. At that point the dashboard creates a second job for someone, who has to interpret it, reconcile it against another report, take a screenshot and forward it with a paragraph of explanation. A simple dashboard can take a few people a week or two to build and then be opened once a month, which makes it a very expensive way of starting a Teams conversation.

Decision support often stops at the dashboard, and the most useful step, working out what to do next, is left to whoever is holding the screenshot.

Chat helps, up to a point

The tempting response is to put a chatbot on top of everything. Most of us have tried to sort out a billing problem through a bank’s chat assistant and know how that can go. A chat interface can earn its place when it shortens the distance between a question, its context, the evidence and the action that follows. Without that, you have built a slower search box with a friendlier tone that still results in a frustrated user.

We don’t think chat should be the only interface either and the world of software development offers a useful comparison. Coding agents such as Claude Code have changed how a lot of software gets written, yet developers still read the diff before they accept it. The conversation sits next to the work (the code review) and points at the evidence (the pull request).

Our demo was built around one question: “Sales are down this month. What happened, who is affected, and what should I look at next?” The dashboard shows the decline, which is what dashboards do well. The conversation then breaks the drop down by region, product and customer segment, handles the follow-up question, cites the metric definition and source data it used, and suggests where to look next, such as distributor stockouts in a particular region or the ten largest accounts whose order frequency has fallen.

The interface is the easy part

If we could keep only one slide from the talk, it would be this one. A chat front end on your data can be built in days. Making its answers trustworthy takes a good deal longer, and the work usually includes

  • a semantic layer, or semantic views, so the model queries business concepts instead of guessing at table names
  • governed data sources, because retrieving something that looks relevant is a long way from retrieving the right number
  • agreed metric definitions and a shared business vocabulary across departments
  • role-based access, lineage and citations, so users can see where an answer came from
  • evaluations, A/B tests, guardrails, a kill switch, and proper instrumentation and logging
  • a human in the loop, and a simple way for users to flag answers that are wrong

If “revenue”, “active customer” or “churn” means one thing in finance and another in sales, a conversational interface will not settle the disagreement. It will help people reach wrong answers faster and phrase them more convincingly. Much of our project time goes into balancing governance and security against the freedom to experiment, and getting that balance wrong in either direction will stall the work and often leave the company fearful of trying the “next big thing”. The event’s platinum sponsors gave a whole session to semantic layers as the foundation for trusted AI, so we were not the only ones at the event making this point.

Measuring a measurement tool

Building metrics for a system whose job is to report metrics feels slightly circular, but it needs doing. Before anything else, a proposed conversational analytics project should pass a simple test of whether it will save money, make money or save time. If it does none of these, it is a vanity project with a chat window and likely won’t have the impact your business needs.

Projects that pass should track the following metrics:

  • time to answer a real business question
  • reduction in recurring ad hoc requests to the analyst team
  • fewer Excel exports and manual reconciliations
  • decision cycle time
  • adoption by the people who make decisions, as opposed to the people who built the system
  • answer accuracy and groundedness
  • cost per useful answer
  • user trust, measured with something as simple as an NPS question. NPS is still useful for internal users which is something we can all forgot too often.
  • the number of questions that turn into new governed metrics

We pay most attention to the last measure. Each question the system answers badly points to a definition missing from the semantic layer, so a well-run conversational layer improves the data underneath it over time.

What’s been happening in 2026 that supports this view

Two weeks after the talk, on 22 June, Snowflake permanently removed its legacy dashboards from Snowsight and pointed customers towards Streamlit apps or third-party BI tools.

Snowflake’s AI assistant, Cortex Code, became generally available in Snowsight in March 2026 and was renamed CoCo at Snowflake Summit in June. You can describe the dashboard you want and CoCo builds a Streamlit app against your data, with charts and filters, which you then refine through conversation. On the business user side, Snowflake Intelligence was renamed Cowork, and Snowflake describes its new automations as an agent that checks for you instead of a dashboard you check. Snowflake also announced Cortex Sense, which learns business definitions from query history and existing dashboards, and Semantic View Autopilot, which converts existing SQL and BI definitions into semantic views.

Microsoft introduced Fabric Apps at Build in June. You start from a Power BI semantic model, describe the app in plain language, preview it and publish it, and Fabric provides the database, authentication and hosting. At FabCon Europe in Barcelona in September, Microsoft made Fabric IQ in Copilot generally available, so a business user can ask a question in Copilot Chat and get an answer grounded in a governed semantic model. Apps in Power BI are moving into preview for Pro and Premium Per User licences, and the Fabric IQ MCP server lets agents such as Claude Code and GitHub Copilot query the same models.

Both vendors have built their conversational features on top of the semantic model, which is the conclusion we reached on client projects. Building a dashboard through chat now takes minutes. Whether anyone should trust what it shows still depends on the definitions, governance and testing underneath, and that part has not become any quicker.

In October, Anthropic released Claude Dashboards which, when combined with MCP servers, can produce live and interactive dashboards that surface data from your line-of-business systems from a prompt.

Where this leaves dashboards

Dashboards will stay, but their role is shifting from a destination people visit towards evidence that appears inside a conversation when someone needs it. For organisations already on Snowflake or Fabric, the conversational tooling is now part of the platform, so the effort goes into the semantic layer, the metric definitions and the governance.

If you are building something in this space, or would like to have a discussion about dashboards, semantic layers or why your favourite report should not be replaced by a chatbot, contact us.