Somewhere in your organization there is a dashboard that answers a question nobody has asked in fourteen months.
It still refreshes every morning. It still has an owner, nominally. It sits in a folder next to eleven other dashboards that were built for a board presentation, a funding renewal, or a privacy officer who has since moved on. The data underneath it is fine. The pipeline is healthy. The chart is accurate. And it is completely useless, because the question it was designed to answer is not the question anyone has today.
This is not a failure of execution. It is the dashboard model working exactly as designed, and the design has expired.
What a dashboard was actually for
Strip away the visual language and a dashboard is a story told in advance.
Someone sat down, decided which questions mattered, decided which slices of the underlying data answered them, and then froze that decision into a set of aggregations. Daily active users. Average time to triage. Referral completion rate by region. Each tile is a pre-computed narrative arc: here is the beginning, here is the trend, here is where you should feel concerned.
That was a rational design. When you are running nightly ETL against a warehouse that charges by the query, and the consumer is a human being who can hold about seven things in working memory, aggregation is not a limitation, it is the entire point. The dashboard compressed a firehose of raw events into something a person could read in ninety seconds before a meeting.
The trouble with compression is that it is lossy, and you have to decide what to throw away before you know what you will need.
Every dashboard is a bet that the future's questions will resemble the past's. It is a bet that gets worse every day it stays up. Forrester's research has put the abandonment rate at roughly a third of dashboards within two months of deployment, not because the data was wrong, but because the questions moved and the chart did not.
The constraint that disappeared
Here is what changed, and it is not subtle.
For twenty years, the reason you aggregated data before a human touched it was that a human could not read a million log lines. That was the binding constraint. Compute was expensive, sure, but the real bottleneck was the analyst.
A language model can read a million log lines. Not skim them, read them, correlate them, notice that seven of them share a session ID and an unusual sequence, and explain why that matters in a paragraph. The bottleneck moved. And when the bottleneck moves, the architecture built around it stops making sense.
This is why the interesting analytics work in 2026 has stopped happening in the BI layer and started happening at the log layer. The industry shorthand for it is agentic analytics, but the underlying shift is simpler than the term suggests:
Old model: pull information out of your logs, aggregate it into a visualization, hope someone looks at it. New model: leave the logs where they are and talk to them.
The move is from a pull-and-precompute architecture to a query-in-place, ask-anything one. You do not build a chart for "unusual after-hours access to psychiatric records." You ask the question, in English, at 4pm on a Thursday when your privacy officer gets a phone call, and you get an answer grounded in the actual event stream rather than a summary of it.
The infrastructure showed up faster than anyone expected
This would be a nice theory if the plumbing did not exist. It does, and it arrived in about two quarters.
Through the first half of 2026, essentially every major observability vendor shipped a Model Context Protocol server: Datadog, Grafana, Honeycomb, New Relic, Sentry, Splunk. The effect is that logs, metrics, traces, incidents, and deployments stopped being things you view in a vendor's UI and became tool calls an agent can invoke. Grafana's server covers Prometheus, Loki, and Tempo. Datadog's went generally available in March and is HIPAA-eligible. Honeycomb has gone as far as positioning itself as purpose-built for the agent era.
That reframing: every observability platform as a set of callable tools rather than a destination you visit, has done more for AI in operations than any assistant any of them launched in the same window.
The BI vendors are cannibalizing themselves in parallel. Tableau shipped Pulse and then Tableau Next. Databricks is running conference sessions titled some variation of "beyond legacy BI." Gartner's own framing has moved from augmented analytics to what it calls perceptive analytics, and it projects that augmented analytics capabilities will evolve into autonomous platforms managing a fifth of business processes by 2027. Their 2026 predictions push further: autonomous agents interpreting governance policy into machine-verifiable data contracts by 2030.
When the incumbents start eating their own dashboard business, the direction is settled. The only question is timing.
Why this matters disproportionately in health
Most industries have logs. Health has audit logs, and the difference is enormous.
An EHR audit trail is a legally mandated, timestamped, tamper-evident record of every interaction with protected health information: who accessed it, when, from where, what they did, whether they modified it. Under HIPAA, and under PHIPA for those of us operating in Ontario, you are already required to generate it, retain it, and be able to produce it for a regulator.
So health organizations are sitting on the richest behavioural dataset they will ever own, generated as a compliance byproduct, and overwhelmingly using it in the dumbest possible way: as a forensic archive you grep through after something has gone wrong.
The conventional privacy monitoring workflow is a sampled report. Pull a percentage of break-the-glass events. Review same-surname access. Flag anything against a VIP record. These are dashboards by another name, pre-frozen questions, run monthly, catching exactly the categories of misuse someone anticipated when they configured the rules.
The categories nobody anticipated go unnoticed. Academic work on EHR anomaly detection has been circling this for over a decade. Fabbri and LeFevre's explanation-based auditing, Menon's collaborative filtering against peer behaviour, and more recent isolation-forest and transformer approaches. The persistent finding is that most of it is evaluated on single-provider data and optimized for model accuracy rather than whether a privacy team can actually deploy and act on it.
Conversational access to the audit log changes the shape of that problem. Instead of building a detector for a pattern, you interrogate the stream:
- "Show me every clinician who accessed a record for a patient they had no scheduled encounter with, in the last 30 days, and rank by how unusual that is for their peer group."
- "This account was compromised on the 14th. Reconstruct everything it touched and tell me which patients are notifiable."
- "Did anyone's access pattern change materially in the two weeks before they resigned?"
None of those are dashboard tiles. All of them are answerable today, against data you are already legally obligated to be keeping.
The economics are not marginal. Healthcare breaches remain the most expensive across all sectors, and the dominant cost variable is dwell time: the gap between compromise and detection. Cutting mean time to detect from months to hours is the difference between a contained incident and a notifiable one.
The honest part
Two things would make this article a sales pitch if I left them out.
Natural language querying is not as good as the demos suggest. On Spider 1.0, the long-standing academic text-to-SQL benchmark, frontier models score above 90%. On Spider 2.0. Same task, but enterprise-realistic schemas with thousands of columns, multiple dialects, and multi-step workflows, execution accuracy collapses to roughly 21%. Enterprise-adapted variants of BIRD land near 39%. The failure mode is the dangerous kind: the query does not crash, it returns a plausible number that is wrong.
The fix is not a better model. It is a semantic layer. The gap between 21% and something you would let a privacy officer act on is almost entirely made of modelled definitions, what "encounter" means, which timestamp is authoritative, what counts as an access event versus a page render. Organizations that skip this step and point an agent at raw tables get confident hallucination with a compliance workflow attached.
And conversation has its own failure mode. Researchers Ken Gu, Srishti Palani, and Vidya Setlur presented work at CHI 2026 on what they call conversational debt: people conducting extended analytical conversations cannot find their way back to insights buried in the history. There is no search, no navigation, no structure. The more you explore, the harder it becomes to recover what you found. Dashboards, for all their rigidity, are at least persistent and addressable.
So what do you actually do
The thesis is not "delete your dashboards." It is that the dashboard has been demoted from the primary interface to the durable artifact of a conversation that already happened.
Practically, for a health data organization:
- Instrument for questions, not for charts. Stop pre-aggregating at ingest. Keep event-level granularity in the log and let the aggregation happen at query time. Storage is cheaper than the questions you will not be able to answer.
- Build the semantic layer before the agent. This is the unglamorous 80% of the work and the entire difference between useful and dangerous.
- Expose logs as tool calls, under RBAC. Read-only, scoped credentials, with the agent's own access captured in the audit trail. An agent querying PHI is an actor in your access log like any other, and regulators will treat it that way.
- Promote the good conversations into dashboards. When a question gets asked three times, it has earned a chart. That is the correct order of operations, and the inverse of how every BI program has ever been run.
- Keep humans on the decisions. Detection and investigation can be autonomous. Notification, discipline, and clinical judgement cannot.
The dashboard's job was to tell you a story about your data. It did that job under a constraint that no longer exists that a person could not read the raw record.
Now they can. The story is already in the log. You just have to ask.
Sources & further reading
- Gartner, Top Predictions for Data and Analytics (March 2026) and 75% of Analytics Content to Use GenAI by 2027 (June 2025)
- Gartner, Top 10 Data and Analytics Technology Trends (2020) — the original decline-of-the-dashboard prediction
- Forrester Research on dashboard abandonment rates
- Lei et al., Spider 2.0: Towards Evaluating Text-to-SQL in Realistic Enterprise Settings (arXiv:2411.07763)
- Li et al., BIRD: A Big Bench for Large-Scale Database Grounded Text-to-SQL
- Gu, Palani & Setlur, "I Need to Find That One Chart" — CHI 2026, on conversational debt
- Fabbri & LeFevre, Explanation-Based Auditing (2011); Menon et al. on collaborative filtering for inappropriate EHR access (2014)
- Vendor MCP server documentation: Datadog (GA March 2026), Grafana, Honeycomb, New Relic, Sentry
