I have been thinking about what happens to Business Intelligence when people no longer need to know which dashboard to open, which filter to apply or even how to write a query.

For a long time, the BI function has provided a relatively clear interface between organisational data and the people trying to understand it. BI teams collect requirements and data, model the data, produce reports, maintain dashboards and help users interpret what they see.

That approach isn’t necessarily disappearing, but the interface is changing. Most major BI platforms are now adding natural-language querying (Conversational Analytics), automated visual creation, narrative explanations and agent-like capabilities. Microsoft Fabric, Tableau, Qlik, Looker, Oracle, IBM Cognos, ThoughtSpot and others all seem to be moving in this direction, although the maturity and depth of their capabilities seems to vary.

The practical question is not whether GenAI will be added to BI platforms - it already has been. What does that mean for the role of BI within an IT department.

What is it?

Traditional BI usually follows a recognisable sequence. Data is collected from operational systems, transformed into a usable structure, organised through models and then presented through reports, dashboards or analytical tools. A user selects the information they need, applies filters and interprets the result.

GenAI introduces another layer between the user and the data. Instead of opening a dashboard and navigating through it, someone may ask:

• “Why did operating costs increase last month?”
• “Which regions are falling behind forecast?”
• “Summarise the main causes of customer complaints.”
• “Create a chart comparing this quarter with the same period last year.”

The BI platform interprets the request, translates it into a query, retrieves the relevant data and presents an answer, visualisation or narrative.

Maybe the user wants to continue the conversation:

• “Exclude one-time expenses.”
• “Show me only the public-sector accounts.”
• “What appears to be driving the difference?”

This sounds like a better search interface, and initially that may be all it is. However, the direction of travel is toward something broader. An analytical agent may monitor information continuously, detect an unusual movement, explain what appears to have happened, identify who needs to know and distribute the finding through Teams, Slack, email or another business application. At that point, BI is no longer simply waiting for someone to open a report, it is beginning to participate in the flow of work.

There are seem to be several levels to this change.

• Conversational access: Users ask questions in ordinary language rather than navigating dashboards or writing queries.
• Assisted analysis: The system proposes visualisations, identifies patterns, summarises results or suggests follow-up questions.
• Proactive insight: An agent monitors selected measures and brings changes or anomalies to someone’s attention.
• Recommendation: The system combines data and business context to suggest a possible response.
• Action: The analytical agent passes information to another system, creates a task, initiates a workflow or takes some other operational step.

The final two levels are where the boundary between BI, workflow automation and agentic AI becomes less clear, a system that explains why inventory is below target is performing analysis. A system that creates a replenishment request is participating in an operational process - quite different levels of responsibility.

In spite of all this change, natural-language access does not remove the need for properly structured data, quite the opposite - it makes it more important. When a dashboard is designed by a BI professional, many decisions are built into it. The measures have been selected, calculations have been defined, relationships have been tested and the way information is presented has usually been reviewed.

A conversational interface allows users to ask a much wider range of questions. The quality of the answers therefore depends heavily on the semantic model beneath the interface: what terms mean, how measures are calculated, which data sources are trusted and which relationships are valid.

The easier it becomes to ask a question, the more important it becomes to govern the meaning of the answer.

What does it mean from a business perspective?

The conversational data analytics and agent approach to data analytics brings a range of benefits and concerns to be considered.

• The most immediate benefit is access. Business users who previously depended on a BI specialist may be able to answer routine questions themselves. They do not necessarily need to understand the structure of the data, remember the name of a report or wait for someone to create another variation of an existing dashboard. They can wander through, and around the data looking for the answer to a question.
• BI backlog reduction. The conversational approach does not necessarily reduce the need for BI expertise. In fact, it may create a different backlog, instead of requests for new reports, the BI team may receive questions such as: Why did the assistant produce a different total from the finance report? Which definition of revenue is it using? Why can one manager see information that another manager cannot? Can this answer be trusted enough to use in an executive presentation? These are not report-development questions. They are questions about meaning, lineage, access, quality, behaviour and accountability.
• The BI role therefore begins to shift from report production toward stewardship of the organisation’s analytical environment (which arguably they have always had). The semantic layer is so important - terms such as customer, employee, active account, completed transaction or operating cost often appear straightforward until different parts of the organisation calculate them differently. A human analyst may recognise the ambiguity and ask a follow-up question. An AI assistant may confidently choose one available definition. The organisation needs to decide which definitions are authoritative, where alternatives are legitimate and how those distinctions are made visible to the user.
• BI specialists are well placed to lead this work because they already sit between raw data, business meaning and technical implementation. The difference is that the semantic model is no longer supporting only a set of known reports. It may be supporting an open-ended conversation.
• Quality control also changes. A conventional dashboard can be tested against expected values. Its filters, calculations and access controls can be checked before release. A natural-language interface introduces far more variation because two users may ask the same question in different ways, or ask questions the designers did not anticipate. Testing therefore needs to include realistic questions, ambiguous wording, incomplete requests and attempts to combine information incorrectly.
• Explainability helps. Some platforms can expose the generated SQL or DAX, cite the supporting data or show how an answer was constructed. But showing a generated query will not help every user. Someone still needs to decide how answers will be validated, which errors matter and what evidence is needed before an insight is used.
• The support model becomes more complicated as well. When an AI-generated answer is wrong, where does the incident go? It could be caused by poor source data, a failed data pipeline, an incorrect measure, an access-control problem, a generated query, the underlying model or the wording of the user’s request. Resolving that issue may involve the data engineering team, BI specialists, application support, security, the platform vendor or the business owner of the information.
• The relationship between BI and business teams will also change. Self-service BI was intended to give business users more control over analysis. In practice, it often created a mixture of centrally governed reporting and locally created spreadsheets, extracts and dashboards. GenAI makes self-service easier again - users may be able to generate analyses, measures and visualisations simply by describing what they want. That lowers the barrier to creation, but not necessarily the barrier to creating something reliable (vibe coding anyone?).
• Agentic capabilities push the issue further. There is a significant difference between an agent that tells a manager sales are below target and an agent that automatically changes a forecast, creates a purchasing request or reallocates resources. The first assists someone and the second affects a business process. As BI becomes connected to workflow platforms and operational applications, the organisation needs to separate: insight from recommendation; recommendation from approval; approval from execution, and automated action from accountable decision
• Licensing matters. GenAI capabilities may be tied to premium subscriptions, user licences or cloud capacity. The cost of making an assistant available to a small analytical team can be very different from making it available to thousands of employees. There may also be additional consumption costs when a platform calls an external or separately licensed model.
• Finally, there is the effect on the BI profession itself. Some work will become easier. Drafting a measure, creating an initial visual, documenting a model or summarising a dashboard may take much less time. That does not mean BI specialists become unnecessary - it means the valuable part of the role moves - understanding the organisation, challenging definitions, identifying misleading comparisons, recognising data-quality problems, designing useful measures and helping leaders interpret uncertainty remain difficult tasks. The role becomes less about operating the reporting machinery and more about ensuring the organisation asks useful questions and receives answers it can responsibly use.

What do I do with it?

The practical starting point is not to enable every available AI feature, start by deciding what role you want conversational and agentic BI to play.

• Strengthen the semantic foundation first. Review the measures, terminology, relationships and business definitions that will ground AI-generated answers (at least in the data set you need - not everything needs to be solved at once). If people already argue about which dashboard is correct, adding a conversational interface will not resolve the disagreement.
• Create a small set of trusted analytical domains. Rather than opening every dataset to natural-language querying, begin with an area where the data is understood, the business owner is engaged and the consequences of an incorrect answer can be managed.
• Separate exploration from authoritative reporting. Users should know whether they are looking at an approved corporate measure, an AI-generated interpretation or an exploratory analysis. These can all be useful, but they should not be presented as though they carry the same level of assurance.
• Test questions, not just features. Build a test set using the questions people actually ask, including vague requests, conflicting terminology, unusual periods, missing data and questions that should not be answered. Record the expected answer and the acceptable range of variation.
• Define the levels of autonomy. Decide whether the capability may answer, explain, recommend, alert or act. Apply more oversight as the system moves closer to operational action or consequential decisions - manage the risk.
• Design human review properly. “A person will check it” is not enough. Identify who reviews the output, what evidence they receive, whether they can detect an error, what authority they have and how their decision is recorded.
• Establish a support path. Users need somewhere to report questionable answers. Support teams need enough logging and lineage information to determine whether the issue originated in the data, model, semantic layer, prompt, permissions or integration.
• Review the BI operating model. The team may need clearer responsibilities for AI evaluation, adoption, training and agent oversight. These responsibilities may sit within BI, a central data office, an AI team or a combination of functions.
• Measure business outcomes. Do not judge success only by the number of questions asked or the number of users enabled. Look at whether people find information faster, make fewer reporting errors, reduce duplicated analysis, improve decision turnaround or rely more consistently on governed data.
• Treat integration standards as part of the roadmap. MCP and A2A may become important ways for analytical agents to interact with other systems and agents. For now, assess actual vendor functionality, security controls and support commitments rather than treating support for an emerging protocol as a simple procurement checkbox.

The point here is not that dashboards are about to disappear. Many business activities still require stable, repeatable and carefully designed reporting. Executives will continue to need agreed measures, regulators will continue to require consistent evidence and operational teams will continue to depend on reliable monitoring.

What changes is the way people reach that information and what the analytical platform can do after finding it. BI is moving from being primarily a destination, somewhere users go to look at data, toward becoming an interface through which people and systems ask questions, receive explanations and potentially initiate action, which gives the BI function a broader role within IT.

It also gives BI a more important responsibility: ensuring that easier access to an answer does not get mistaken for greater certainty that the answer is correct.