I have spent years building dashboards and preparing the data behind them. I regularly saw the same paradox: organisations invest heavily in datasets, definitions and dashboards, while actual usage can remain limited.
Part of the problem is the human interface. A branch manager who wants to understand why commercial performance is falling behind needs to know the dashboard, trust it, know how to use it and interpret a whole set of two-dimensional charts in context.
Conversational analytics points towards a different future: ask the business question and let the technology determine what data and analysis are needed.
The technology is starting to become seriously capable of this. But Microsoft, Databricks and other vendors also show that a language model alone is not enough. Between the user's question and the data, a controlled layer of business context is required.
That controlled layer matters for a simple reason. The system cannot merely translate language into a query. It has to understand what the organisation means by revenue, margin, active customer, branch performance or budget variance, which source is authoritative, which access rights apply and which logic has already been validated. The future of conversational analytics therefore depends not only on better models, but on better semantic grounding.
How vendors are solving this problem
Microsoft builds on the semantic model in Power BI and Fabric. Best-practice guidance for Fabric data agents explicitly emphasises well-structured semantic models, AI-ready descriptions, trusted measures and clear business context to improve answer quality. That reflects a broader design principle: the agent should not infer critical business meaning from raw tables alone.
Databricks takes a conceptually similar approach from the data platform with Genie. A Genie Agent does not simply get access to tables. It can be configured with business instructions, curated data assets and benchmark questions so teams can test whether the system returns the expected answers consistently. The details differ, but the logic is comparable.
The architectures are different, but the underlying problem is the same.
The system cannot just understand language. It needs to know which revenue definition applies, which branch the user is responsible for, which time period is relevant, which data source is authoritative and which access rights apply.
Generative AI does not remove that work. Good definitions, metrics, metadata, business logic and data quality become prerequisites for having a reliable conversation with your data. Increasingly, that also means a governed semantic layer that carries business definitions, access policies and lineage consistently across dashboards, SQL, applications and AI agents.
What can't it do reliably yet?
The technology is not yet an infinitely reliable digital data analyst.
Ambiguous business concepts remain difficult. Complex business logic can produce incorrect results. Data quality problems do not disappear. Microsoft itself warns that poorly prepared semantic models can reduce the quality of answers from data agents, and practical implementation guidance repeatedly stresses the need for controlled scope, evaluation and iterative tuning.
This also introduces a new risk. An AI assistant can explain an incorrect number very convincingly. For management information, transparency, verification and systematic testing therefore become more important, not less.
That is why trustworthy conversational analytics should not only produce an answer, but also make visible what sits behind that answer: the source, the governing definition, the applied time frame, the user context and, where possible, the logic or query path that led to the result. If the system cannot make that chain sufficiently clear, it should be treated with caution in management use cases.
There is also a deeper analytical limitation. More fluent interaction does not automatically make a system methodologically mature. A strong conversational system should be able to indicate where evidence is weak, where sample sizes are too small, where comparisons are distorted, or where correlation should not be interpreted as causality. The more natural the interface becomes, the more important analytical discipline becomes behind the scenes.
Three years from now: analytics on demand
My working hypothesis is that conversational analytics will go much further than asking questions about existing dashboards.
Monitoring itself can become conversational.
A branch manager asks while travelling:
“How is my revenue developing?”
The agent understands the manager's context, analyses current data and responds, for example:
“Your revenue is 4% below budget this month, but 6% above the same period last year. Most of the gap to budget comes from two product categories.”
On the spot, the system generates the visualisations that best demonstrate that answer.
That is a fundamental change from traditional BI. Today, analysts try to predict in advance what information users will need. They build datasets, measures, dashboards, filters and visualisations around those anticipated needs.
In the future, the question itself can determine which analysis and visualisation are generated at that moment. A trend line for one question; a benchmark, waterfall, table or forecast for the next.
The agent can also suggest relevant follow-up questions:
“The gap is concentrated in two product categories. Would you like me to investigate what is driving it?”
Or:
“Your branch is growing more slowly than comparable branches. Shall I adjust the comparison for differences in customer mix and branch size?”
The manager can simply continue the conversation:
“Run that benchmark.”
“What differences stand out?”
“Which factors are most strongly associated with the better-performing branches?”
“Which of those can I influence?”
Behind that conversation, much more analytical computing power can become available than normally fits into a management dashboard. An agent can compare segments, detect anomalies, investigate large numbers of variables, identify correlations, generate forecasts and run scenarios.
More complex causal questions can also be investigated within the same conversation, provided the available data and research design support those conclusions. A mature system should make clear where the evidence stops rather than presenting correlation as causality.
Analytics can therefore become simpler at the front end and statistically much deeper at the back end.
And every answer can be hyperspecific. Not the same dashboard with a different branch filter, but an analysis generated at that moment based on the user, the branch, its customer portfolio, historical development, relevant benchmarks and the specific question being asked.
That future also becomes increasingly multimodal. The interaction does not have to remain trapped in a chat box. The same analytical session can move between text, speech, chart interaction, alerts, briefings and embedded workflow actions, while preserving context throughout. In that sense, conversational analytics is evolving into a broader interaction layer rather than a single interface pattern.
From insight to action
The conversation does not have to stop at analysis.
A manager might ask:
“Where should my team focus this week?”
The agent could combine signals, predictions and expected value into a prioritised action list: customers with increased risk, commercial opportunities, underperforming product categories or other issues requiring attention.
The user can continue:
“Why is this customer at the top?”
“Only show me actions my team can realistically influence this week.”
Monitoring, analysis, forecasting, decisioning and operational action start to converge.
That convergence also raises an important design question: how far should the system go? There is a significant difference between a system that answers, a system that investigates, a system that recommends and a system that acts. As analytics becomes more agentic, those levels need clearer governance boundaries, explicit permissions and stronger auditability. Without that, the step from useful assistance to uncontrolled automation becomes too easy.
Analytics becomes less tied to a screen
I also expect speech to become a much more important interface.
A manager could request a spoken briefing while travelling. During a meeting, someone could verbally ask for a different breakdown and a new chart would appear immediately. An important deviation could arrive as a message. A more extensive analysis could temporarily open as an interactive visualisation.
Text, speech, tables and charts become different representations of the same analytical interaction.
Distribution becomes more fluid as well. Analytics no longer has to live exclusively inside Power BI, Tableau or another BI interface. It can reach users through the communication channels in which they are already working.
The future I see is therefore not simply chatting with a dashboard. It is a personal analytics agent that can monitor, investigate, visualise, benchmark, predict and prioritise potential actions — in real time and specifically for the question at hand.
To work well, that agent also needs to respect the same controls as the user it serves. Permissions, scope restrictions, lineage and trusted business meaning have to travel with the interaction rather than being bypassed by it. The interface may become lighter, but governance has to become stronger.
How do you start today?
I would not try to make the entire data warehouse conversational.
Start with one user group, such as branch managers, and systematically inventory the questions they actually ask of their data. Use interviews, existing dashboards and recurring requests to BI analysts.
Cluster those questions, for example, into:
- Monitoring: How is my revenue developing?
- Diagnosis: Why am I falling behind?
- Benchmarking: How am I performing compared with similar branches?
- Prediction: Where do we expect to end up this month?
- Action: Where should my team focus this week?
This tells you which definitions, metrics, dimensions, business rules and data sources need to be made reliable first.
For important and frequently asked questions, establish validated logic and answers. Those same questions then become a benchmark set for testing whether the system answers correctly when users phrase them in different ways.
That benchmark discipline is becoming a practical best practice. Databricks explicitly recommends benchmark question sets with known expected outcomes to evaluate Genie performance and build trust over time, rather than relying on anecdotal impressions of whether the agent “feels good.”
Methodologically, this resembles building a customer service chatbot. You do not try to answer every conceivable customer question from day one. You inventory common questions, organise the required knowledge, test different formulations and gradually expand the scope.
The difference is that conversational analytics does not simply retrieve existing answers. Each new question can trigger a new calculation, analysis and visualisation.
That also means that implementation should not stop at language enablement. Teams need semantic modelling, trusted definitions, evaluation routines, transparency mechanisms and clear decisions about where the system is allowed to advise and where it is allowed to act.
That is the real promise
Moving from a world in which we pre-build thousands of dashboards to one in which a large part of analytics is generated in real time, hyperspecific, statistically deep and increasingly driven by natural language and speech.
But the deeper promise is even broader than that. It is a shift from analytics as a destination to analytics as an interaction layer: less dependent on a fixed screen, more aware of context, more capable of dynamically generating analysis, and increasingly able to support the path from monitoring to explanation, prediction and prioritised action.
If that future is to be reliable, the foundations have to improve in parallel. Conversational analytics will only become truly valuable at scale when semantic grounding, validated business logic, transparency, governance and statistical discipline mature as fast as the interface does.