Dashboards are dead… or are they? The rise of Natural Language Insights

Natural language interfaces democratise data access for non-technical users. A business user who may be intimidated by complex filters, charts, and data models on a dashboard can simply ask a question in plain English. This eliminates the need for extensive training on a new BI tool and lowers the barrier to entry for data-driven decision-making across an entire organisation.

Dashboards often answer only narrow, repetitive questions. This leads to multiple versions for different teams, creating data silos and inconsistencies. The result: a potential data governance nightmare that could be avoided with strong underlying data models.

Natural language interfaces or conversational AI can help by allowing users to ask questions directly, retrieving insights from a single, trusted data source, and reducing the need for multiple, static dashboards.

Dashboards struggle to adapt to evolving questions. They’re fast to build for a specific scenario, but once leadership asks “what if?” or deeper questions, the dashboard often falls short, leaving decision-makers without the insights they need or leaving the dashboards useless (often leading to data/dashboard swamps). Conversational analytics is better equipped to deal with these more exploratory lines of thought.

Natural language insights and personalised data feeds can deliver insights directly to users based on their role, responsibilities, and past queries. Instead of a user having to “pull” information by going to a dashboard, the system can “push” relevant, personalised alerts and summaries to them. This shifts the paradigm from a reactive, “look-up” model to a proactive, “be-informed” model, which is often more valuable for busy executives and managers.

Traditional dashboards and business intelligence tools struggle with unstructured data like customer feedback emails, support tickets, or meeting notes. An LLM, with its natural language capabilities, can process this data and provide insights that would be impossible with a standard dashboard. For example, it could analyse thousands of support tickets and summarise the top five recurring issues, even if a key metric for that doesn’t exist.

Despite new advancements, for many business functions, a consistent, standardised view of key metrics is non-negotiable. Dashboards provide a single source of truth and a common framework for comparison, ensuring everyone is looking at the same numbers in the same way. This is critical for internal governance, regulatory compliance, and day-to-day operational monitoring.

We are also seeing a mismatch between expectations and reality with these tools. Business users will prompt AI for BI tools using business terminology and natural language, which is typically not how we present data in lower layers in our data warehouse. We cannot rely on the underlying LLM to infer how that maps to the data model. 

Instead of natural language, data needs to be modelled in business terminology. This means semantic layers will become more important than ever. Conversational AI exacerbates the problem of data silos and a lack of data governance. When different departments have their own definitions for key metrics (e.g., “customer” means something different to sales vs. marketing), an LLM has no way of knowing which definition to use. This leads to conflicting answers and a breakdown of trust in the system. While a traditional dashboard forces a single, standardised view, an LLM allows users to ask ad-hoc questions that expose these underlying inconsistencies.

In addition, we don’t want to overload LLMs / agents with too much data – I see a need for domain/function-aligned agents which specialise in query data from models purpose-built for specific use cases. For example, you don’t want to point an agent / LLM at an enterprise data model; that’s far too large and complex. The approach needs to be more targeted with building semantic models for specific groups of semantically related use cases (e.g. a semantic model for our HR data). Getting really good, high-quality business metadata (e.g. column descriptions, semantic meaning, example values, semantically related terminology) will be crucial for successful AI for BI implementations.

AI for BI can’t simply point the model at the data and shoot – we need to give agents / LLMs the necessary context to understand the data and make good, informed decisions to generate correct SQL. This means rich business metadata, well-named columns, semantic layers with business terminology embedded, and data modelled in a way to make querying (for the LLM) simple (more specifically to reduce the number of potential points of failure). 

The other really important point is that business users need to understand the SQL statements generated and check if they make sense. A core challenge with LLMs is their tendency to “hallucinate” or generate plausible-sounding but factually incorrect information. This can be especially dangerous when dealing with business-critical data. If a user asks a question the system can’t answer from the underlying data, the LLM might invent an answer based on its training data, which could be outdated, irrelevant, or simply wrong. This risk is mitigated but not eliminated by connecting the LLM to an internal knowledge base.

If we make these tools accessible to the broader business users and they lack the technical ability to validate the SQL generated and executed, then they will always assume the outputs are correct. Key skills that will become more important for the non-technical business user include prompt engineering, understanding data models and understanding SQL – these are non-negotiable (at least with the current AI maturity).

Dashboards excel at providing a quick, high-level overview of business performance. A user can scan a well-designed dashboard in seconds to see if key performance indicators (KPIs) are on track, identify anomalies, and understand the overall health of an operation. This is far more efficient for routine checks than a conversational interface, which requires multiple back-and-forth interactions to get a comprehensive view.

Many business users are accustomed to and prefer the visual, organised structure of a dashboard. They have built a mental model around where to find specific metrics and how to interpret the charts. Disrupting this established workflow can increase cognitive load and resistance to change. For a user who needs to check the same set of metrics every day, navigating a familiar dashboard is often the most efficient and comfortable process.

The ability to customise and limit what people can and can’t see on traditional dashboards can be achieved much more easily compared to doing it in conversational analytics tools. Since dashboards are often built for specific / narrow use cases, making adjustments – whether it’s managing user permissions, changing filters, or modifying data views – is generally more intuitive, flexible and straightforward than in AI-driven BI tools. Example. Tools like Power BI come with their own security model to handle this, I highly doubt any AI for BI tool right now provides the same security measures natively. 

Dashboards excel at providing a quick, high-level overview of business performance. A user can scan a well-designed dashboard in seconds to see if key performance indicators (KPIs) are on track, identify anomalies, and understand the overall health of an operation. This is far more efficient for routine checks than a conversational interface, which requires multiple back-and-forth interactions to get a comprehensive view.

Many business users are accustomed to and prefer the visual, organised structure of a dashboard. They have built a mental model around where to find specific metrics and how to interpret the charts. Disrupting this established workflow can increase cognitive load and resistance to change. For a user who needs to check the same set of metrics every day, navigating a familiar dashboard is often the most efficient and comfortable process.

The ability to customise and limit what people can and can’t see on traditional dashboards can be achieved much more easily compared to doing it in conversational analytics tools. Since dashboards are often built for specific / narrow use cases, making adjustments – whether it’s managing user permissions, changing filters, or modifying data views – is generally more intuitive, flexible and straightforward than in AI-driven BI tools. Example. Tools like Power BI come with their own security model to handle this, I highly doubt any AI for BI tool right now provides the same security measures natively.