As corporate analytics evolve, so do agentic BI systems—solutions that can take an active role in investigating business questions. Deloitte Hungary's Data & AI team warns that while such systems can run multi-step analyses to uncover relationships and root causes, their autonomy varies and should be matched to acceptable risk and clear accountability.
Natural-language data access is not the same as agentic behaviour
"Chat-with-data" tools make data more accessible: users can ask about KPIs, request explanations for anomalies, or generate visualizations in natural language. In these cases, the user typically directs the analysis. Agentic BI, by contrast, begins when a system autonomously creates an investigative plan for a business objective: it decomposes the problem, chooses datasets and tools, tests alternative explanations, and adapts its approach based on intermediate results. An agentic system leads the investigation, recognizes uncertainty, contradictions and missing evidence, and flags when no clear conclusion can be drawn from the data.
Five capability levels for Agentic BI
Deloitte Hungary's Data & AI team uses a practical framework that distinguishes five capability patterns to help assess how much analytical autonomy and execution authority can create business value at acceptable risk:
- Talk-to-your-data: natural-language questions produce queries, charts or summaries. This improves data access, but the user plans the analysis.
- Guided analytics copilot: relies on certified metrics, business definitions and enterprise semantic models. It more reliably supports financial reporting and self-service analytics but still operates under human direction.
- Investigation-performing analytical agent: designs multi-step analyses, tests hypotheses and can change the analysis direction based on intermediate findings (typical for root-cause analysis or variance investigation).
- Decision-preparing agent: recommends actions based on analysis, for example suggesting forecast adjustments, prioritizing collections cases, or creating an operational task for human approval.
- Reversible execution: under predefined rules, permissions and control points the system can autonomously perform low-risk, well-bounded and reversible operations.
Fejes Péter, senior consultant in the Deloitte Hungary Data & AI team, noted that these levels should not be treated as a strict maturity hierarchy: an organisation might operate at level two for financial reporting, level three in commercial analytics, and level five in certain IT operations. Higher autonomy does not always mean a more advanced or valuable solution.
Analysis autonomy is not the same as execution authority
A system can carry out complex analyses and only issue recommendations, while another may run simpler analyses and automatically perform a narrowly defined operational task. Thus organisations must answer two questions: how autonomously can the system analyse, and what can it do with the analysis results? This distinction matters most for financial, regulated or hard-to-reverse decisions.
Excessive recommendations, insufficient evidence, or unclear responsibilities can produce automation bias and "approval fatigue". Human involvement only provides genuine control if it enables substantive judgement rather than becoming a mere formality.
CDO and CFO perspectives differ
The relevance of Agentic BI varies by leadership role. Chief Data Officers typically prioritise the reliability and operability of the information backbone, while Chief Financial Officers emphasise reconcilability, control and measurable financial return.
CDOs focus on the trustworthiness of the business context used by the agentic system. Agentic BI increases the value of certified metrics, semantic models, data lineage, access controls and clear data ownership. However, a semantic layer alone does not guarantee reliability: a well-structured model may still encode outdated business logic, erroneous data or undocumented exceptions. For CDOs, proving safe, repeatable and cost-effective operation in production is the key challenge; a successful demo is not sufficient evidence.
From the CFO perspective, reconcilability, auditability and financial control can be more important than maximising autonomy. Financial value comes from reliably handling official numbers, manual corrections and associated accountability. Agentic capability alone does not imply financial value; the relevant question is whether the benefits in a given process justify the total cost of technology, integration, operation, human review and error-related risk.
How to measure success
Deloitte stresses that success should not be measured by the number of activated features or deployed agents. More meaningful metrics include:
- repeated usage and the share of successfully resolved cases,
- cost per outcome,
- frequency of overrulings,
- measurable business impact (operational efficiency, decision effectiveness, financial performance and risk reduction).
Takács István, senior manager in the Deloitte Hungary Data & AI team, said the future of BI goes beyond reports: users will interact with data naturally while analytical agents independently investigate business situations, prepare decisions and proactively initiate actions. Organisations that are likely to benefit are those that do not aim for maximal autonomy but instead define, for each decision, the level of autonomy that produces measurable business value with acceptable risk and clear accountability.
Conclusion
Agentic BI offers the potential to uncover insights and prepare decisions autonomously. Deloitte's view is that autonomy must be calibrated to the decision's characteristics, data quality, reversibility and regulatory constraints. Real business advantage comes from measurable results, not technological spectacle.
