Do You Still Need a Dashboard?

The dashboard has been declared dead roughly once a year since 2021.

This year the argument has better funding behind it. An analyst piece asked whether a major BI acquisition spelled the end of the dashboard. Two large data companies merged and described the result as infrastructure for trusted AI agents, not for reports. Every BI vendor blog I can find is now dominated by posts about agents.

So it is worth asking the question properly, with evidence rather than vibes. I went looking for the data behind the claim. What I found is more interesting than either answer.

Who This Is For

• Analytics leads deciding what to build next, and what to stop maintaining
• Founders who have been told their BI tool is obsolete by someone selling the replacement
• Data engineers tired of maintaining reports they suspect nobody opens
• Anyone who has sat in a meeting where two people had different numbers for the same thing

What You Will Need

Only a willingness to check a number before repeating it. That turns out to be the whole methodology here.

The Pattern

Three claims travel together in this argument. They are not equally true.

flowchart TD
  C[The dashboard is dead] --> A[Claim 1: nobody reads them]
  C --> B[Claim 2: chat replaces them]
  C --> D[Claim 3: the artefact changed]
  A --> E[No primary source exists]
  B --> F[Benchmarks say no]
  D --> G[Well evidenced]

Claim 1: Nobody Reads Them

Everyone knows the statistic. Some large percentage of dashboards are never viewed, or viewed once and abandoned. You have seen it in a deck.

I could not find it. Neither could a second, independent search that I ran specifically to check the first. Between them, those passes enumerated the published blogs of four BI vendors and grepped them for dashboard-usage telemetry, and opened the canonical essay in the genre.

No vendor publishes dashboard usage data. Not one. The companies that could measure this trivially, because they host the dashboards, do not report it.

And the essay that named the argument — the 2021 piece titled “Dashboards are Dead” — contains no statistics and no citations at all. Its concrete evidence is one anecdote about a 67-page dashboard at a former employer. It is a good essay. It is an argument, not a finding, and it has been cited for five years as though it were the latter.

There is a related claim I also went looking for: that some percentage of teams lost stakeholder trust because of a data quality issue on a dashboard. No survey asks that question. The nearest real figures are about who notices a problem first, which is a different thing and should not be quietly upgraded.

So claim one is unsupported. Not disproved. Unsupported, which is a different and more honest position. Dashboards may well go unread at scale. Nobody has published evidence that they do, and I am not going to be the next person to repeat a number I cannot source.

Claim 2: Chat Replaces Them

This one can be tested, and it has been.

Microsoft caps its own Fabric data agent at 25 rows by 25 columns per answer. That is not a criticism of the product; it is a sensible limit on a conversational surface. But a 25-cell answer is not a substitute for a view of your business.

On a benchmark of 112 real dashboards where agents had to filter, drill and navigate rather than just read, the best agent scored 38.69%. A separate vendor benchmark found agents answering direct questions at around 75% while dropping to 54% on questions containing a plausible trap. They answer well and exercise judgement badly.

And the capability gap is not where the marketing suggests. Tableau’s agent explicitly cannot build a dashboard. Looker’s conversational analytics cannot create Looks or dashboards. Two of the three largest vendors do not ship the thing the category is sold on.

There is a subtler problem. A controlled study found that conversational explanation produced better understanding and higher trust than a dashboard — along with clear overreliance that the conversational format amplified. People felt more informed and checked less. That is a poor trade for a number that drives a decision.

What AI can and cannot do to the dashboard itself, judged against the same benchmarks, is the subject of can AI build your dashboard yet.

Claim two is contradicted by the available evidence. Chat is an excellent way to ask a question. It is a poor way to keep several people looking at the same reality.

Claim 3: The Artefact Changed

Here the argument is right, and this is the part worth acting on.

The shift is not from dashboards to chat. It is from the dashboard being the deliverable to the metric definition being the deliverable, with the dashboard as one of several ways to render it.

The evidence for this is unusually consistent, because it comes from people building the tools rather than people writing think pieces:

• Databricks now argues in its own product positioning that the semantic layer, not the dashboard, is the primary artefact.
• Snowflake improved text-to-SQL accuracy from 57% to 78% by having agents refine the semantic model rather than the prompt or the model.
• One BI vendor’s benchmark found a coding agent given both a semantic layer and a detailed playbook still trailed a purpose-built harness by 21 points, concluding: “The semantic model isn’t enough; it’s the harness because it forces the agent to obey the model.”
• Cube, Lightdash, Rill and Evidence all pivoted to shipping MCP servers and agent skills, so the tool becomes a documented target rather than a chat window.

flowchart LR
  S[Semantic layer] --> D[Dashboard]
  S --> Q[Chat answer]
  S --> A[Alert]
  S --> G[Agent action]
  D --> T[Same number everywhere]
  Q --> T
  A --> T
  G --> T

This is why the “dead or alive” framing is the wrong question. The dashboard did not lose to chat. It lost its position as the first thing you build. Once the metric is defined once and governed properly, rendering it as a dashboard is cheap, and so is rendering it as an alert, a Slack message, or a tool an agent can call. That architecture is what semantic layers for trustworthy dashboards describes, and it is what makes conversational analytics safe enough to expose to non-technical people.

What a Dashboard Is Still the Only Thing That Does

Strip away the argument and there is a residue of jobs nothing else does well.

Shared situational awareness. Research treats dashboards as a genre distinct from exploratory visualisation tools, precisely because their job is different: several people, the same picture, at the same time. A chat transcript is private by construction. Six people asking the same question get six answers and no shared object to disagree about.

Benn Stancil, who wrote “BI is dead” in October 2021, wrote this six weeks later:

“Data and the dashboards that display it create a shared sense of reality… Defining metrics and creating dashboards isn’t banal busywork. It’s one of the most important things that we do.”

That reversal is worth more than either post on its own.

Noticing what you were not looking for. Chat requires a question. A dashboard shows you the thing you did not think to ask about, which is the entire value of glancing at one every morning. Pair it with thresholds and it becomes the trigger surface described in dashboards that trigger actions.

Repeated questions on a schedule. If the same question is asked weekly, paying a model to re-derive the answer every time is worse than rendering it once.

Anything that must be auditable. A regulator, an auditor or a board wants a fixed artefact with a known provenance, not a conversation.

The Gap I Could Not Close

Honesty requires flagging what is missing, because it is the question underneath all of this.

I found no research answering whether dashboards actually improve decisions. Searching for it returns clinical decision-support studies, which are about a different kind of system in a different setting. The literature on dashboard design is rich. The literature on dashboard efficacy is, as far as I can find, absent.

Which means both sides of this argument are running on intuition. The people saying dashboards are dead have no usage data. The people defending them, including me in the section above, have no efficacy data. Anyone claiming certainty here is selling something.

Before and After

Before	After
The dashboard is the deliverable	The metric definition is the deliverable
A new question means a new dashboard	A new question means a new surface over the same metric
Metrics are defined inside each report	Metrics are defined once and governed
Chat is pitched as the replacement	Chat is one renderer among several, and a weak one for shared awareness
Dashboards accumulate and nobody deletes them	Unopened dashboards are retired on a schedule
”Are dashboards dead” is the question	”What is this metric, and who owns it” is the question

Failure Modes to Avoid

• Repeating the abandonment statistic. It has no source. Two independent passes failed to find one, and the vendors who could measure it do not publish it.
• Replacing dashboards with chat and calling it modernisation. You will have swapped a shared artefact for private conversations that agree with each other by coincidence.
• Building the semantic layer as a migration project. Define the twenty metrics that appear in more than one report. That is most of the value.
• Assuming the agent can operate what it can describe. Reading a chart and driving a dashboard are different skills, and the benchmark gap between them is large.
• Keeping a dashboard because deleting it feels risky. If nobody has opened it in a quarter, the risk is that someone trusts it while it quietly breaks.
• Waiting for certainty. There is no efficacy research coming to settle this. Decide from your own usage data, which you can collect this week.

What to Build First

1. Pull the actual view counts for every dashboard you maintain. Your BI tool has this even if your vendor does not publish aggregates. You are now better informed than anyone quoting the statistic.
2. Delete or archive everything unopened for a quarter. Announce it rather than doing it silently, and see who objects. The objections tell you what is load-bearing.
3. List the metrics that appear in more than one surviving report. That list is your semantic layer, and it is shorter than you fear.
4. Define those metrics once, in one place, with an owner’s name against each.
5. Render them wherever they are needed: the dashboard for the shared morning view, an alert for the threshold, a chat surface for the long tail of one-off questions.
6. Only then evaluate agents, and evaluate them against the questions your team actually asked last month.

Final Take

So, do you still need a dashboard?

Yes, for the jobs a dashboard is uniquely good at: several people seeing the same thing, noticing what you were not looking for, and producing an artefact somebody can audit. Those needs did not change because the tooling did.

What changed is that the dashboard stopped being the first thing you build. The metric definition took that position, and it deserves it, because everything downstream — the chart, the alert, the chat answer, the agent’s tool call — is only as trustworthy as the definition underneath it.

The teams doing well right now are not the ones who deleted their dashboards or the ones who defended them. They are the ones who can answer, in one sentence and without opening a file, what “active customer” means in their business.