Data visualization best practices: A practical guide to clear, honest, effective charts

Richard Tomlinson has spent the bulk of his career helping organizations work with their own data, including getting their own people to trust it. In both BI and product roles, he has watched the same failure pattern repeat itself: dashboards built with good intentions that people quietly stopped using. Now that AI can generate a chart in seconds, his experience tells him that "self-service" has to mean something more than fast.

In this exchange, Richard explains why dashboards lose trust long before they lose good design, what AI genuinely changes about that problem and what people have to bring to a chart that no AI assistance can supply on its own.

Why do people stop using BI dashboards?

Tell me about a dashboard you've seen fail. Not one that looked bad, one that people simply stopped opening.

Richard Tomlinson: The dashboards that fail are not necessarily ugly. Often they fail because someone sees a number they don't believe and can't quickly understand why it is there.

Picture an executive opening a revenue dashboard on Monday morning and seeing a figure that's different from the report Finance circulated on Friday. They don't know whether the dashboard is using a different revenue definition, a different refresh time, a different filter or simply the wrong data. If they have to ask an analyst to find out, the dashboard has already failed one of its most important jobs.

Trust tends to disappear much faster than it is built. Users rarely file a ticket that says they no longer trust the BI platform. They quietly stop opening it and go back to spreadsheets, analysts or manually prepared reports.

That's why I think good visualization starts before the visualization itself. A clear chart can't compensate for an unclear metric, a questionable source or an inconsistent business definition. In an AI-powered analytics environment, this matters even more, because users can generate many more analyses much faster. If the underlying data, semantics and governance are inconsistent, AI ends up just helping you produce inconsistent charts faster.

The best modern analytics experiences connect the visualization back to its data and its meaning, so there is one version of the truth across every surface instead of five. For example, Databricks AI/BI Dashboard datasets inherit Unity Catalog governance, and the platform maintains lineage between data assets. That means a user who doubts a number can trace it back to its source instead of writing off the whole system.

What should leaders prioritize when adopting AI dashboard tools?

As AI transforms how we use and interact with data, what should leaders prioritize when bringing on new AI dashboarding tools?

Richard Tomlinson: I like the idea of trust debt, because organizations accumulate it every time users encounter two versions of the same metric or an unexplained discrepancy, an outdated dashboard, or an answer they can't verify.

AI can make that debt substantially worse, because it changes the economics of analytics. Organizations used to build hundreds or even thousands of curated dashboards over years. With generative AI, anyone can create a new chart in seconds and with agentic authoring, AI can now quickly create an entire multi-page dashboard, including datasets, visualizations, filters and layout.

Databricks' Genie Code, for example, can take a natural-language objective, find the relevant data, build datasets, create visualizations, configure filters, organize pages and refine the result much faster than that used to take. That is a real productivity gain, but it changes what leaders need to govern, since you can't review every AI-generated chart by hand.

So, my answer is that I would prioritize four things:

1. Govern the data, so AI inherits the same permissions and controls that apply everywhere else.
2. Govern the meaning, so important measures and dimensions come from reusable business definitions instead of being reinvented every time someone generates a chart.
3. Preserve provenance, so a user can understand what data and calculation produced an important result.
4. Design for inspection, so that visualizations invite follow-up questions, filtering and drilling rather than presenting an unexplained answer as fact.

Good governance actually enables more self-service. With strong foundations, AI can safely generate far more analytical experiences without a central BI team handcrafting every one.

Where do AI-generated charts break down?

When an AI assistant generates the chart, it is making decisions based on the shape of the data, not the decision the person is trying to make. Where does that break down?

Richard Tomlinson: I would adjust that premise a bit. Modern AI assistants use more than the shape of the data. They can interpret the request and its context, too. If I say, "Show how revenue has changed over the last 12 months," that implies a time series. If I ask, "Which five regions generate the most revenue?," the same data might point toward a ranked bar chart. Ideally authors should be able to describe visualizations in natural language, which is why Databricks already enables Genie Code to plan and build multiple visualizations as part of a larger dashboard objective.

Where AI still struggles is purpose. A dataset tells you what values are available. A prompt tells you what someone asked. Neither one reliably tells you who is going to use the visualization, what decision they are trying to make, what comparison matters most, what should attract attention first, what level of precision is required, what context the audience already understands or what could be misinterpreted.

Let's say a metric has fallen from 94% to 91%. AI can chart that decline correctly, but whether it is catastrophic or irrelevant noise depends entirely on business context that lives nowhere in the dataset. That's why the human role is shifting from drawing charts to specifying intent and exercising judgment. AI can do the mechanical work of building a visualization, but someone still has to decide what it's supposed to communicate and whether it's accurate.

Why does legacy BI produce bad analytics even when the data is accurate?

What are the other challenges that creep up in legacy systems, beyond inaccurate or missing data?

Richard Tomlinson: The hardest problems show up not because of missing rows but because of missing context. A legacy BI environment might contain accurate data and still produce bad analytics because business logic has fragmented across dashboards, calculated fields, extracts, semantic models, spreadsheets and institutional knowledge.

You end up with five different definitions of the same KPI and nobody is sure which dashboard is canonical. Calculation logic lives inside individual reports, users have to know which tool and tab holds the answer and changing a definition means tracking down every place it was duplicated.

AI exposes that problem very quickly. An assistant may correctly understand the question and still have no reliable way to determine which of five revenue definitions the company considers authoritative. That is why the missing piece is not simply attaching a language model to a visualization layer. AI-powered dashboards need a foundation of governed data plus reusable business semantics and context.

Even Microsoft warns in its current Power BI guidance that Copilot output can be low-quality or misleading if the underlying data, semantic model and users have not been properly prepared. Databricks approaches the same problem by bringing dashboards, governed datasets, business semantics and conversational analytics onto one governed data foundation. AI/BI datasets can draw directly from tables, views and Metric Views, as well as inherit Unity Catalog permissions. That architecture matters more in an AI-first world because the AI needs to reuse business meaning rather than reconstruct it every time it builds a chart.

How do you write better AI prompts for charts and dashboards?

What are the habits of people who consistently get what they need from an AI assistant without three rounds of follow-up?

Richard Tomlinson: Successful users provide decision context, not just chart instructions. A weak prompt is, "Show sales by product." A much stronger one is, "I am preparing our weekly sales review. Show which product categories contributed most to the change in revenue versus last quarter. Make the biggest positive and negative contributors easy to identify." The second version tells the AI the audience, the comparison, the metric and the purpose of the visualization.

Providing this kind of context can dramatically improve analytical prompts. What am I trying to understand? State the business question rather than just naming a dataset. What should I compare? Time periods, regions, segments, plan versus actual or before versus after? Which metric definition matters? If there is any ambiguity, specify the business metric instead of letting the AI invent one. Who is this for? An executive dashboard and an analyst's exploration usually need completely different levels of detail. And what decision should this support? This is the piece people most often leave out.

There is another shift happening, too. Better AI systems should reduce how much prompt engineering you need to do in the first place. Agentic systems can increasingly clarify ambiguous requirements, inspect the available data, construct an analytical plan and iterate on the result. For instance, Genie Code can discover data, construct datasets and visualizations, and progressively refine a dashboard rather than requiring the author to spell out every implementation step. This doesn't mean in the future everyone has to become a world-class prompt engineer, it's about humans getting good at expressing intent while agents get better at handling implementation.

What are data visualization best practices in the age of AI?

As people start querying AI against dashboards, queries and pipelines directly, how do the characteristics of "good" visualization design change?

Richard Tomlinson: This might be the most interesting shift of all. A visualization used to be the destination. Someone opened a dashboard to look at charts predefined by an analyst. This would help them understand the data, but that is where the information transfer stopped.

With AI, interacting with visualizations is more like a conversation. A user asks a question, the system generates an answer and maybe a chart, the chart reveals something interesting, the user asks about that, then the system runs another analysis and produces another visualization. The experience is dynamic, not predefined.

AI/BI Dashboards already take this into account. A published dashboard can include a companion Genie Agent that lets viewers ask natural-language questions about the underlying data instead of being limited to something static.

This means the first chart doesn't need to anticipate every possible question. It just needs to communicate the key insight clearly and make the next question obvious and easy to ask.

This all leads to the four characteristics you want to see in your AI-era visualizations. They need to be:

1. Self-explanatory, because a generated chart may show up outside the dashboard where it originated, so titles, units, time periods and metrics need to stand on their own.
2. Traceable, so users can understand where an important number came from.
3. Interactive, so filtering, drilling and conversational follow-up enable you to move naturally from observation to investigation.
4. Context-aware, so the visualization reflects the user's question and business terminology rather than just plotting whatever fields happened to be returned.

There is a deeper implication here, too. Dashboards, queries, pipelines and documents are no longer only things humans consume. Analytics artifacts can now become part of an organization's machine-readable knowledge, not just its presentation layer. They can become context that AI systems use to understand how the organization works. For example, Genie Code already lets authors reference tables, pipelines, notebooks, queries and files as context when building an analytical experience.

Why do users stop trusting self-service analytics?

What is the pattern right before someone stops trusting self-service analytics and goes back to asking an analyst directly?

Richard Tomlinson: Usually, the issue isn't that the system gives someone nothing. Rather, it gives them something plausible that they cannot reconcile with what they already know. Maybe a number looks wrong, or they change a filter and get an unexpected result, or two dashboards disagree and they can't tell which definition is being used. They might ask a follow-up question and get another chart, but it doesn't explain the discrepancy.

At that point, someone has two choices: keep debugging the analytical system themselves or message someone and ask them. Once asking a person is faster than verifying the self-service answer, self-service has already failed.

That is why the critical metric for modern BI should not only be time to answer. It should also be time to trust. AI makes the first metric dramatically better. Queries, visualizations, and entire dashboards can be generated in seconds. But unless the second metric improves, too, all you've done is accelerate production of things people don't trust.

Getting this right means combining three things traditional BI has often treated separately: answers that are easy to understand, evidence that makes the answer credible and a path to investigate when something doesn't look right. A good self-service experience is not the one that removes the analyst from the loop at all costs. It's the one where people only need the analyst when a question genuinely requires human expertise, not because they can't understand or trust the system in front of them.

That is ultimately why visualization still matters in the agentic era. AI can generate an enormous amount of analysis. Good visualization means that analysis can be inspected, challenged and, most importantly, used to action.

See how AI-assisted authoring and governed data make dashboards people trust. Explore AI/BI Dashboards.