For twenty years, self-service BI promised that business users would answer their own questions. In practice, most people consumed dashboards someone else built, and the long tail of questions landed in the reporting team’s backlog. Conversational BI is the latest attempt to close that gap, and this time the technology is materially different. Large language models can translate a question into SQL or DAX, run it against governed data under the user’s permissions and explain the result in plain language.
In 2025 and 2026 every major platform made this a core feature. Microsoft has Copilot in Power BI and Fabric data agents. Databricks has Genie, now extended into Genie Agents and the Genie One workspace. Snowflake positions Snowflake Intelligence as a work agent for business users. Analyst firms have taken note: Gartner predicts that by 2028, 60 percent of self-service analytics users will use general-purpose LLMs for ad hoc and exploratory analysis, while production-grade reporting stays on traditional analytics and BI platforms, according to a summary of its Predicts 2026 research. This article explains what conversational BI can and cannot do today, and what reporting teams should change to make it trustworthy.
How Conversational BI Works Under the Hood
Despite different product names, the architectures are similar. Microsoft’s documentation for Fabric data agents offers a clear example of the pattern:
- Parse the question and check it against responsible AI policies and the user’s permissions.
- Select a data source. A Fabric data agent can be connected to up to five sources, such as a warehouse, lakehouse, Power BI semantic model or KQL database, and chooses the most relevant one.
- Generate a query in the right language: NL2SQL for lakehouses and warehouses, NL2DAX for Power BI semantic models and NL2KQL for KQL databases.
- Validate and execute the query with the requesting user’s credentials, so row-level and column-level security still apply.
- Return an answer as a summary, table or key insight.
Two design points deserve attention. First, the agent is only as good as the context it receives: table and column names, descriptions, relationships, measures, instructions and example queries. Second, the agent runs with the user’s identity, which makes it safer than a shared service account, but also means that inconsistent permissions or poorly designed security roles surface immediately.
The 2026 Landscape at a Glance
- Microsoft Power BI and Fabric. Copilot in Power BI helps authors build report pages and summarize visuals, and lets consumers ask questions of reports and semantic models. Fabric data agents are customizable, read-only artifacts that answer questions over governed OneLake data and can be surfaced in Microsoft 365 Copilot, Teams and Copilot Studio. Microsoft documents that data agents require a paid F2 or higher Fabric capacity, or Power BI Premium P1 or higher with Fabric enabled.
- Databricks. Genie spaces, renamed Genie Agents around the 2026 Data + AI Summit, answer questions over Unity Catalog data and, according to Aimpoint Digital’s summit summary, now support write actions, MCP, sub-agents and skills. Genie One provides a unified conversational entry point, and Metric Views supply a governed semantic layer.
- Snowflake. Snowflake Intelligence is positioned as a personal work agent for business users, with skills defined in plain language, MCP connectors to tools such as Slack, Jira and Salesforce, and a deep research capability that combines structured and unstructured data, according to Techzine’s coverage of the April 2026 announcements.
- BI-native assistants. Other BI vendors offer similar capabilities inside their tools, with varying depth of semantic modeling and governance.
What Conversational BI Is Good At, and Where It Falls Short
Strengths
- Long-tail questions. One-off questions that never justified a dashboard, such as “Which distributors in the Nordics grew fastest last quarter?”
- Follow-up exploration. Drilling into a number seen on a dashboard without waiting for a new report.
- Accessibility. Users who never learned to navigate complex reports can get answers in their own words, including inside Teams or Microsoft 365.
- Speed for analysts. Drafting queries and measures faster, even when analysts review the result.
Limitations to plan around
- Ambiguity. Business language is imprecise. “Revenue” might mean booked, billed or recognized; “last quarter” might mean fiscal or calendar. Without explicit definitions, agents guess.
- Result size and complexity. Microsoft documents that Fabric data agent responses are limited to 25 rows and 25 columns and are summarized. Conversational BI is for answers, not data extracts.
- Language and scope. At the time of writing, Fabric data agents support English only, are read-only and do not query unstructured files. Example queries are not yet supported for Power BI semantic model sources, so semantic model quality carries more weight.
- Consistency. The same question phrased differently can produce different queries. That is acceptable for exploration, but not for numbers that will be reported upward.
- Governance edge cases. Sensitivity labels and data loss prevention policies can block queries or produce incomplete answers, which confuses users unless explained.
Conversational BI does not replace the semantic model; it exposes it. Every ambiguous name, missing description and inconsistent measure that report designers used to hide behind a polished dashboard is now visible to every user who asks a question.
Why the Semantic Model Becomes the Product
In the dashboard era, a reporting team’s product was the report. In the conversational era, the product is increasingly the semantic model and its context, because that is what every agent, copilot and report consumes. Gartner has predicted that by 2030 universal semantic layers will be treated as critical infrastructure alongside data platforms and cybersecurity, and its Predicts 2026 research, as summarized by CData, describes a semantic layer working alongside MCP as necessary for consistency and governance in agentic analytics.
For Power BI teams, this means semantic models need to be designed for two audiences: report authors and language models. The practices overlap but are not identical.
Semantic model readiness checklist
- Star schema with clear fact and dimension tables and unambiguous relationships. Many-to-many relationships and bidirectional filters confuse both people and models.
- Business-friendly names for tables, columns and measures. “Net Revenue (Recognized)” is better than “NR_rec_amt.”
- Descriptions everywhere, especially on measures, explaining calculation logic, inclusions and exclusions.
- Explicit measures for every important metric, rather than relying on implicit aggregation of columns.
- Hidden technical fields, such as surrogate keys and staging columns, so agents do not select them.
- Date tables and fiscal calendars marked and documented, so time phrases resolve correctly.
- Synonyms and AI instructions where the platform supports them, capturing internal jargon and preferred interpretations.
- Security roles tested with real user personas, because agents surface whatever the user is allowed to see.
A Readiness Playbook for Reporting Teams
Step 1: Choose a domain, not the whole estate
Start with one well-understood domain, such as sales performance or service operations, where the semantic model is already mature and the audience asks many ad hoc questions. Avoid domains with contested definitions until those are resolved.
Step 2: Collect real questions
Gather 50 to 100 actual questions from users, help desk tickets and report request backlogs. These become your test set. Include ambiguous and tricky questions, not just easy ones.
Step 3: Harden the semantic model
Apply the checklist above. Add descriptions, rename fields, create missing measures, document fiscal calendars and write agent instructions that route questions to the right source. In Fabric, data agent instructions can specify, for example, that financial metrics should be answered from the certified semantic model rather than raw lakehouse tables.
Step 4: Evaluate systematically
Run the test questions and grade answers as correct, partially correct or wrong, recording the generated query. Fix the root cause for each failure, usually a missing description, an ambiguous name or an absent measure, and rerun. Keep this test set as a regression suite for every model change and every platform update.
Step 5: Launch with clear expectations
Tell users what the assistant is for (exploration and quick answers), what it is not for (official reported figures and large extracts) and how to report wrong answers. Provide links from answers back to certified reports where possible.
Step 6: Monitor and improve
Review question logs, failure reports and usage. Microsoft notes that data agent interactions can be logged and discovered through Purview Audit and eDiscovery. Questions that are asked frequently and answered well may justify a certified report or a verified answer; questions that fail repeatedly reveal gaps in the model.
Worked Example: Rolling Out a Sales Data Agent
Consider a manufacturing company with a mature Power BI sales semantic model used by regional sales managers. The reporting team receives dozens of ad hoc requests each week, most of them variations of “show me X by Y for Z.”
The team gathers 80 recent requests as a test set and creates a Fabric data agent connected to the certified sales semantic model and a lakehouse table of open opportunities. The first evaluation run answers a little over half of the questions correctly. Failure analysis shows three root causes: “revenue” is ambiguous because the model has both invoiced and shipped revenue measures without descriptions; time questions fail because the fiscal year starts in April and the date table is not documented; and regional questions fail because sales territories and legal entities share similar names.
The team renames measures, adds descriptions, documents the fiscal calendar, hides legacy columns and adds agent instructions that define default interpretations. The next run answers most questions correctly, and the remaining failures are genuinely ambiguous questions that the agent should clarify. The agent is published to Teams for the sales managers, with a note that official monthly figures remain in the certified report. Over the next quarter, ad hoc report requests fall noticeably, and the most popular question patterns feed into an updated certified report.
Decision Guide: Chat, Certified Report or Agent Workflow?
One of the most useful things a reporting team can do is publish a simple rule for where different kinds of questions should be answered. Without it, users will either distrust the assistant entirely or treat every chat answer as an official number. A practical rule of thumb uses three channels.
- Certified reports and dashboards for numbers that are reported upward, compared over time, used in performance management or disclosed externally. These need fixed definitions, version control, scheduled refresh and visual consistency. Monthly revenue, headcount and service levels belong here.
- Conversational BI for exploratory and follow-up questions, one-off slices and quick checks, where speed matters more than formatting and where the user can tolerate occasionally rephrasing a question. Answers should link back to the certified source where one exists.
- Agent workflows for recurring multi-step tasks that combine data with actions, such as monitoring a KPI daily, drafting a weekly commentary or opening a ticket when a threshold is breached. These need explicit design, testing and ownership, and usually draw on the same semantic model.
Two signals tell you when to move a question between channels. If the same conversational question is asked many times by many people and the answer feeds a decision, promote it to a certified report or a verified answer. If a certified dashboard is opened rarely and mainly to answer one-off questions, consider retiring it in favor of conversational access to the same model. Reviewing these signals quarterly keeps the estate lean and the channels clear.
Finally, be explicit about accountability. A certified report has a named owner who answers for its numbers. Conversational answers inherit trust from the semantic model, so the owner of that model is effectively accountable for what the assistant says about the domain. Making that ownership visible, for example in the agent’s description, encourages the right fixes when answers go wrong.
Pitfalls to Avoid
- Pointing agents at raw tables. Conversational BI over unmodeled data produces confident but inconsistent answers. Use certified semantic models and curated tables.
- Skipping evaluation. A good demo is not evidence. Use a test set of real questions and track accuracy over time.
- Overpromising. Positioning chat as a replacement for reports leads to disappointment and to users treating exploratory answers as official numbers.
- Ignoring capacity and cost. Conversational features consume capacity. Monitor usage and plan capacity accordingly.
- Fragmenting definitions across tools. If Power BI, Genie and Snowflake Intelligence each hold their own metric definitions, answers will diverge. Aim for one governed semantic layer per domain.
How Reporting Roles Evolve
Conversational BI shifts reporting work rather than eliminating it. Report developers spend less time building one-off views and more time on semantic modeling, documentation, testing and governance. A new responsibility emerges, sometimes called analytics engineering for AI: maintaining the context, instructions and evaluation suites that keep agents accurate. Reporting teams also become curators, deciding which questions deserve certified answers and which can stay exploratory.
Outlook
Expect conversational BI to move from a feature inside BI tools to a channel embedded across work tools, with agents in collaboration apps, email and mobile drawing on governed semantic layers through open protocols such as MCP. Expect also a split that Gartner already anticipates: exploration through general-purpose assistants, and production reporting on governed BI platforms. Organizations that invest in semantic models and evaluation now will be able to support both without multiplying definitions.
Conclusion
Conversational BI is finally useful, because modern agents can generate correct queries against well-modeled data under the user’s own permissions. Its quality, however, is bounded by the semantic model and context behind it. Reporting teams that treat the semantic model as their core product, evaluate agents with real questions and set clear expectations with users will turn chat from a demo into a dependable channel for answers.
How Hanoi Technology Can Help
Hanoi Technology helps reporting teams harden Power BI semantic models, configure Fabric data agents and set up evaluation suites for conversational BI.
Further Reading
- Fabric data agent concepts – Microsoft Learn
- Gartner Announces Top Predictions for Data and Analytics in 2026 – BigDATAwire
- Predicts 2026: AI Agents, MCP and Governance Transforming Analytics – CData
- Databricks Summit: What Business Leaders Need to Know About Agentic Analytics – Aimpoint Digital
- Snowflake Intelligence and Cortex Code become the agentic AI control layer – Techzine