AI and Business Intelligence: how Artificial Intelligence turns data into decisions

For years we asked Business Intelligence to tell us what had happened. Today we also ask it why, what might happen tomorrow and, in some cases, what we should do.
This is the point at which Business Intelligence and Artificial Intelligence begin to overlap. While BI structures data, metrics and access, AI adds forecasting, anomaly detection, generated explanations and agentic capabilities.

In 2026 the convergence has become visible across the main platforms. Microsoft integrates Copilot into Power BI and Fabric; Google is developing Looker around Conversational Analytics, Dashboard Agents and Agentic Workflows; Tableau has brought Tableau Agent into Pulse; Qlik has expanded its family of agents with Discovery Agent and Predict Agent (Microsoft, Google Cloud, Tableau, Qlik).

The most important change concerns the distance between question and action. When asking the data a question costs almost nothing, value shifts: knowing how to find the right chart matters less, while being able to trust the underlying definition, and knowing how far the system can go once it has found an answer, matters more.

There is, of course, a prior layer: business KPIs, which determine which signals truly deserve attention. Because if the metric is weak, AI simply provides faster access to a weak metric.
Let’s look at how to avoid this, how to make the most of AI in your company’s BI, and the costs and risks involved.

AI and Business Intelligence: what really changes

Although traditional BI remains useful, AI is radically changing its cost of access, the time needed for exploration and the ability to move from reactive reading to more proactive monitoring.
This chart sums it up best:

From dashboard to natural language

Imagine a CEO asking: “Why did margin in Northern Italy fall last quarter?”.
A plausible scenario, after all. In a classic system, answering would mean opening the dashboard, applying filters, comparing periods and then drilling down by product, channel or customer.
Today, conversational interfaces move part of this work into the question itself.

Looker Conversational Analytics lets users query data in natural language on top of the platform’s governed semantic modeling layer. Power BI Copilot offers chat-based experiences ranging from on-the-fly analysis to DAX generation and the creation or summarisation of reports (Google Cloud, Microsoft). All of this has significant organisational consequences: more people can query the data directly.

This reduces dependence on the analyst for simple questions, while also increasing the number of people who can use a wrong answer.

When the cost of asking falls, the value of definition rises

Before AI, an inconsistent definition could stay hidden behind a dashboard used by a handful of specialists. Today, with a conversational interface, dozens or hundreds of people can ask “what is the margin?” without knowing the underlying model.

The semantic layer thus becomes a kind of information constitution: it establishes what active customer, revenue, churn, margin, pipeline and conversion mean before AI starts reasoning about them.

Google places particular emphasis on the role of Looker’s semantic layer in keeping conversations anchored to governed metrics. Microsoft requires well-prepared semantic models for Copilot to produce useful, consistent results. The interface becomes freer; the structure beneath it must become more rigorous.

AI can look for anomalies before anyone asks

A traditional dashboard waits for someone to open it. Agentic systems, by contrast, can monitor metrics and look for significant changes periodically, on a schedule, continuously.

Qlik describes Discovery Agent as a capability that continuously analyses applications, identifies relevant changes and returns prioritised insights without requiring predefined static rules. Google has introduced Looker Agentic Workflows in preview to monitor metrics and carry out root-cause analysis in the background (Qlik, Google Cloud).

For management, the value lies above all in reduced latency: less time between when something changes and when someone notices.

From descriptive analysis to prediction

Classic BI describes the past and present well. Machine learning and predictive models add estimates of demand, churn, delays, probability of closing, inventory or risk.

A forecast, however, remains an estimate built on the relationships observed in the data. Its value declines when the context changes enough to make history less useful: a new competitor, a demand shock, regulation, supply chain disruption or a product change.

This is where AI increases analytical capacity without removing judgement. A forecast can be faster and more granular; management still has to recognise when the conditions that made it reliable no longer hold.

Generative AI can explain the data

A dashboard shows ‘conversion rate: -12%’. A generative system can add an initial reading: the drop is concentrated on mobile, originates in paid traffic and begins after a specific date.

Tableau Pulse combines metrics and Tableau Agent to enable natural-language questions; in July 2026 Salesforce updated the model underlying Agent in Pulse to handle more complex questions about metrics (Tableau). Microsoft and Google offer similar summarisation, conversational query and assisted generation features.

This capability can compress hours of manual reporting. Verification remains necessary: a fluent explanation can be built on a wrong definition, on incomplete data or on a correlation that does not support the conclusion.

From Generative BI to Agentic BI

In 2026 vendors increasingly use the term Agentic BI. What sets it apart from a conversational assistant is chiefly initiative and the ability to act.

Google has announced Looker BI Agents and Agentic Workflows to move from answers to downstream workflows. Qlik explicitly distinguishes between an assistant, which responds, and agents that monitor or act (Google Cloud, Qlik).

This is where the governance question changes shape. Traditional BI asked who can see a piece of data. Agentic BI must add: what can the system do after seeing it?

A scale of autonomy for AI within BI

To avoid treating every AI function the same way, it can help to classify autonomy into four levels.

This progression is consistent with the logic of Human Oversight: human control must be designed according to the impact of the action, not added generically at the end.

AI reduces the analyst’s monopoly on simple questions

A substantial share of requests to data teams is repetitive: sales by region, year-on-year comparisons, CAC by channel, customers above a threshold, variance against budget.

When the model is governed, many of these questions can go directly to business users. The analyst gains time for modelling, quality, causality, experimentation and problems that genuinely require analytical expertise.

AI democratises access to the answer. The ability to frame a useful question remains a rare managerial skill.

An example: AI applied to the sales pipeline

A B2B CRM holds deals, value, stage, salesperson, sector, expected close date and activity history. BI can show total pipeline, win rate, sales cycle and forecast.

AI can add another layer: “Enterprise opportunities open for more than 90 days have a lower probability of closing. Three accounts represent 41% of the forecast and show a slowdown in activity over the last two weeks”.

The system can then suggest a review of the three opportunities. At a higher level of autonomy, it could prepare the tasks in the CRM and submit them to the manager responsible.

The whole chain, however, depends on the quality of the CRM. If sales staff do not update the system properly, AI interprets an incomplete representation of reality.

Hallucinations and plausible answers: the risk changes when the question is about the business

Google explicitly warns that Conversational Analytics can generate plausible but factually incorrect output and recommends validating answers before use (Google Cloud Documentation).

Risk grows with the consequences of the decision. A flawed summary of an internal report can be corrected quickly; a misreading of liquidity, credit or production capacity can move capital.

This is where it meets AI Governance: tools, accountability, ROI, access and compliance must be part of the same system.

Five maturity levels of AI-enabled BI

Sequence matters.
Jumping from reporting to agentic autonomy without a semantic layer, quality and governance produces fragile automation behind a highly sophisticated interface.

How to implement AI and Business Intelligence in your organisation

The first question should be a managerial one: which decision do we want to make sooner, better or faster? The sequence is built from there.

This sequence avoids the “we bought Copilot, now let’s find something for it to do” project. Technology comes in once the problem is already defined.

A sensible first use case: an Executive BI Assistant

For many companies, I would start with a low-risk, high-volume case: an assistant for senior management that answers recurring questions using governed data.

• How far above or below budget are we?
• Which business units explain the variance?
• Where has margin deteriorated most?
• Which KPIs have crossed a threshold?
• Which changes warrant human review?

No automated actions in the first phase. First, measure whether the system improves access, speed and understanding. Autonomy can increase only after quality and errors have been observed.

How to measure whether AI in BI creates value

Microsoft also notes that Copilot in Fabric and Power BI consumes Fabric capacity and must be managed to avoid affecting other workloads. Computational cost therefore also becomes part of the Total Cost of Ownership (Microsoft).

BI becomes more accessible and more delicate

For years Business Intelligence had an access problem: the data existed, but it took skills, queries and dashboards to reach it. AI shortens that distance.

The result is a paradox: the easier it becomes to get an answer, the more important it becomes to govern the meaning of the data and the use of the answer. Scarcity shifts from access to information to the quality of the system that decides what that information means and what can happen next.

The real evolution of BI, then, is the shift from a system that shows to a system that observes, interprets and, with increasing permissions, prepares or takes actions. Management retains the hardest task: deciding how much autonomy to grant, and over which decisions.