← Back to the journal

DataFocus · Journal

BI Copilots vs an Ontology Platform: Who Answers Business Questions Correctly?

A BI copilot answers well inside one tool's semantic model. When a question crosses CRM, ERP and contracts, or must give the same figure to dashboards and external agents with the SQL attached, an ontology layer is the better home.

A BI copilot is enough when the question lives inside one well-prepared semantic model in one BI tool: "What were last quarter's sales by region in this report?" It is not enough when the question depends on definitions that span several systems, such as whether "Customer A" in the CRM, the ERP and a contract is the same company, or when the same answer must reach dashboards, search and external AI agents with the SQL and source tables attached. That second class of question needs a shared definition layer, which is what an ontology platform such as DataFocus OT provides. The useful comparison is traceability and shared definitions, not which chat window feels nicer.

What does a BI copilot actually ground its answers in?

Start with the vendor's own description. Microsoft's Copilot for Power BI overview (as of November 2026, page last updated 24 August 2026) says: "When a question is related to data in the semantic model, Copilot uses the semantic model to answer the question. Otherwise, it answers from the large language model's (LLM) general knowledge" (Microsoft Learn). The same page tells model owners to prepare the semantic model first, because "without preparation, Copilot can misinterpret the data and return generic or inaccurate results."

That is a sound design, and it states the boundary clearly: the copilot is as good as the one semantic model it sits on. The report pane answers about the open report; the app-scoped experience is "limited to the app's included content and the permissions users have within the app." Microsoft also offers verified answers prepared by app authors, which is a real governance aid inside that scope.

Tableau's AI features follow the same general pattern of grounding in the tool's own data sources; check Tableau's current documentation for the specifics, since these products change quickly.

The research literature agrees on where the hard part is. The Spider 2.0 benchmark paper (Lei et al., arXiv, 2024, revised 2025) reports that an o1-preview agent framework solved only 21.3% of 632 real enterprise text-to-SQL tasks, against 91.2% on the older Spider 1.0. And in a vendor-reported study, Collate wrote in April 2026 that adding its semantic layer raised one model's execution accuracy on a Spider 2.0-Snow subset from 10.8% to 76.5% (Collate, 2026); that is a single vendor's result with no question count or variance reported, but the direction matches the argument of its title: the text-to-SQL problem is mostly a context problem, not a model problem.

So the question is not "copilot or no copilot." It is: where does the context live, and how many tools can use it?

Where does one-tool semantics break?

DataFocus's product brief names three things enterprise AI usually lacks (OT product brief):

  1. Context — whether "Customer A" in the CRM, the ERP and a contract is the same company, what "delivered" means, and which rule version is in force.
  2. Boundaries — what AI may see or change.
  3. The ability to act with a person in control.

A semantic model in one BI tool handles the first point only for the tables that were loaded into it. Consider a hypothetical distributor whose analyst is asked, "Which customers are behind on payment for orders we have already delivered under their current contract terms?" That question needs:

  • the CRM account record,
  • the ERP order and delivery status,
  • the contract document with the payment terms,
  • an agreed meaning of "delivered" (shipped, received, or signed for).

If the contract terms live in PDFs and the delivery definition lives in a finance team's spreadsheet, a copilot on the sales semantic model has no place to find them. It may answer anyway, from the model's general knowledge, which is exactly the fallback Microsoft documents. The answer will read fluently and may be wrong in a way nobody can see.

The second failure is drift. If the sales team's report, the finance dashboard and an external agent each carry their own definition of "active customer," a well-behaved copilot on each will give three confident and different numbers.

How does an ontology platform answer instead?

In DataFocus OT, the ontology holds the objects and relationships, and AI follows them instead of guessing. Every query returns the SQL that ran, the tables it used and a data fingerprint (OT product brief). The analyst's job shifts from trusting the chat to checking the receipt.

In OT Heavy, the mechanics are visible in the interface (OT Heavy manual):

  • Data Search shows a "Query SQL Statement" panel on the result's right-hand rail with the SQL the search actually ran and its computation time.
  • Typed input that is not standard column names and keywords is sent to FocusGPT to be parsed into a search statement; the user sees a prompt with a Revert action. Even for a standard statement, FocusGPT parses it in the background and may replace it.
  • Smart Q&A (FocusGPT) picks the data table, runs the query and replies with a line saying how the input was parsed. It usually takes a minute or two, the page warns "Content is generated by AI, please carefully identify it", and the manual advises checking important conclusions in Data Search.
  • My Knowledge Base lets users add business definitions and terminology for FocusGPT, one at a time or in batch, each item with content, keywords, type, category and an enabled flag. Administrators manage system-wide knowledge under Admin > Knowledge.

That last point is where "what does delivered mean" gets written down once, in a place the AI reads.

The product brief gives a worked example on built-in sample data (a demo dataset, in Chinese). A business user in Data Search and Claude Code connected through MCP both asked for totals by requesting department and got matching figures: Operations Support 7,842,000, Procurement 6,966,000 and Digital 5,786,000 (yuan). The agent's answer cited the SQL and a receipt field called evidenceRowsSha256. The point is not the numbers; it is that two very different clients produced the same figures and one of them could show a hash of the evidence rows.

How does one definition reach search, dashboards and agents?

The brief's rule is "define once, use everywhere": a business definition set once is used by search analysis, dashboards, control panels, the ontology agent and external agents, and they all get the same figures (OT product brief).

For business users, the surfaces on top of that layer include search-based analysis (type column names and keywords, get a chart, see the SQL, plus forecasting and attribution analysis), natural-language questions through FocusGPT, an ontology workbench, control panels, dashboards with scheduled delivery to WeCom, DingTalk or Feishu, data entry forms, and 21 interface languages.

Side-by-side: copilot vs ontology layer

CriterionBI copilot (e.g. Power BI Copilot, as of November 2026)Ontology platform (DataFocus OT)
What answers are grounded inThe tool's semantic model; otherwise LLM general knowledgeOntology objects and relationships across connected tables and documents
Scope of definitionsOne semantic model, report or appDefined once, reused by search, dashboards, control panels, ontology agent and external agents
Evidence returnedDepends on experience; verified answers in app scopeSQL that ran, tables used, data fingerprint on every query
Business glossary for the AIModel preparation and descriptions in the semantic modelKnowledge Base entries (Heavy) with type, category and enabled flag
External agentsWithin the vendor's ecosystemSame figures to external agents, e.g. Claude Code through MCP in the brief's example
Best fitReport-centric questions in an established Power BI estateCross-system questions that must be traceable and consistent

Checklist: which questions belong where?

Send the question to the BI copilot when:

  • [ ] all the data sits in one semantic model the copilot already uses,
  • [ ] the metric definitions are owned and documented in that model,
  • [ ] the answer is consumed inside that BI tool,
  • [ ] a wrong answer is cheap and easy to spot.

Send it to the ontology layer when:

  • [ ] the answer joins CRM, ERP and documents such as contracts,
  • [ ] a term like "delivered" or "active customer" has more than one meaning in the company,
  • [ ] the same figure must appear in a dashboard, a search result and an agent's answer,
  • [ ] someone will ask "show me the SQL and the tables" before acting on it,
  • [ ] external AI agents need the figures, not just people in the BI tool.

A simple decision rule: if two of the five ontology boxes are ticked, put the definition in the ontology first and let the copilot, dashboards and agents read from it.

Trade-offs and when an ontology is not the right fit

  • If your organisation is standardised on Power BI, your questions stay inside well-prepared semantic models and you rely on verified answers, the copilot is the simpler choice. Adding another layer creates work without much return.
  • An ontology only pays off if someone owns the definitions. A Knowledge Base nobody maintains will drift like any other glossary.
  • FocusGPT answers take time (the Heavy manual says a minute or two), and the product itself tells users to verify important conclusions in Data Search. Treat the natural-language answer as a draft and the SQL as the evidence.
  • Edition matters: OT Heavy continues the Hadoop, Hive, Trino and Iceberg architecture, while OT Flash is the lighter edition. Plan resources around data volume, concurrency and model requirements; system requirements and supported platforms are listed in each edition's release notes.

Next step

Read the OT product brief for the full definition model, then pick an edition on the download page. Both editions are free; installers and release notes are on GitHub for OT Flash and OT Heavy.