How to choose an AI-native analytics solution
A buyer's guide to evaluating AI-native analytics, built on FERRARO's seven principles of agentic analytics.
How to choose an AI-native analytics solution
To choose an AI-native analytics solution, check seven things: people control how much the AI does, AI is built into the core, analytics is available to every AI application and agent, agents work across the whole lifecycle, chat is one way to work alongside visuals and dashboards, every answer can be verified, and the system keeps learning. Separately, check that business context doesn't require a semantic layer project. This guide turns each one into questions to ask and red flags to watch for.
Every BI vendor now says "AI." Here's how to tell them apart
Most analytics products now have an AI feature. Some were rebuilt around AI; many added a chatbot or copilot to a product designed before AI. FERRARO Consulting's 2026 report, The Seven Principles of Agentic Analytics by John Santaferraro, describes the market as "a mix of imposters, innovators, and true inventors" and sets out seven principles to help buyers tell them apart.
This guide uses those seven principles as an evaluation framework. If you're new to the term, start with What is AI-native BI?, which explains the difference between AI added to BI, chat-first tools and AI-native BI.
The seven principles at a glance
People at the center
The person decides how much or how little the AI does and stays in the loop wherever they want to confirm a step.
AI at the core
AI is part of the architecture, working across the whole lifecycle from data connection to sharing insight, not a copilot panel added to an existing builder.
Put agentic analytics at the center of every AI application
The platform's analytics engine is open to other AI agents and applications through standard interfaces such as MCP and open APIs, with secure, multi-tenant access.
Agents across the whole lifecycle
Agents start working the moment data connects: profiling and provisionally cleansing data, completing discovery, suggesting questions and visuals, and continuing in the background to surface anomalies, risks and opportunities.
Conversation as one modality
Chat is a way in, not the whole product.
Trust through governance, transparency and validation
Access policies and permissions apply to people and agents alike, every step is auditable, and every generated insight cites its source data, transformations, calculations and confidence.
Continuous learning
The system learns from how people use it, the data, and how insight is used across the organization, so it gets more useful over time, and people learn alongside it.
People at the center
What it means
The person decides how much or how little the AI does and stays in the loop wherever they want to confirm a step. Familiar analytics features stay available, so experienced analysts aren't forced into a chat box.
Questions to ask vendors
- Can each user choose how much the AI does, from fully manual to AI-built?
- Where can a user stop, review or override what the AI did?
- Can analysts still build by hand when they want to?
Red flags
- A chatbot tacked onto an old product
- AI that is either on or off, with nothing in between
- No way to step in mid-task without starting over
AI at the core
What it means
AI is part of the architecture, working across the whole lifecycle from data connection to sharing insight, not a copilot panel added to an existing builder.
Questions to ask vendors
- If you turned the AI off, would the product still work the same way?
- Can the AI make every change a user can, through the same logic?
- Which parts of the product does the AI not touch?
Red flags
- A copilot or chat panel bolted onto a pre-AI builder
- AI that only suggests and summarizes while people do all the work
- An AI roadmap that is mostly "coming soon"
Quick test: ask what happens if you remove the AI. If the product works the same, just less conveniently, the AI was added on top.
Put agentic analytics at the center of every AI application
What it means
Your AI agents and applications should be able to use the analytics platform as a governed source of curated data and insight, not just the people using its own interface. In its evaluation of Golden, FERRARO looks at support for MCP (Model Context Protocol), Open API and Search API access, protected-resource metadata (RFC 9728) for zero-touch authentication, and multi-tenant routing of requests.
Questions to ask vendors
- Can our own AI agents and applications query the platform through MCP or open APIs?
- How do external agents authenticate, and is it standards-based (for example RFC 9728 protected-resource metadata)?
- How are requests routed and isolated across tenants, and do the same governance rules apply to agent access as to people?
Red flags
- Insight is only reachable inside the vendor's own UI
- No MCP or documented API access for agents
- Agent access bypasses the permissions people are held to
Agents across the whole lifecycle
What it means
Agents start working the moment data connects: profiling and provisionally cleansing data, completing discovery, suggesting questions and visuals, and continuing in the background to surface anomalies, risks and opportunities.
Questions to ask vendors
- What happens automatically when we connect a new dataset?
- Which steps do agents handle: prep, discovery, analysis, dashboards, storytelling?
- Do agents keep working in the background, and how do they notify people?
Red flags
- AI that only appears after someone types a question
- Data prep and cleansing left entirely to people
- Agents limited to one step, such as writing SQL
Conversation as one modality
What it means
Chat is a way in, not the whole product. Conversation, visuals, dashboards, reports and stories work together, and you can move between typing and clicking on the same work.
Questions to ask vendors
- Can I start in chat and finish by editing the dashboard directly?
- Does the AI ask clarifying questions and recommend next steps?
- Can the output be a dashboard or story, not just an answer in a chat thread?
Red flags
- A chat box is the only interface
- Fixing one detail means re-prompting from scratch
- Answers stay stuck in a chat thread instead of shared dashboards
Trust through governance, transparency and validation
What it means
Access policies and permissions apply to people and agents alike, every step is auditable, and every generated insight cites its source data, transformations, calculations and confidence. The most mature tools validate at three levels: built-in system checks, review by the person asking, and review by colleagues.
Questions to ask vendors
- Does the AI do the math, or does a deterministic engine?
- Can I see and edit the SQL behind every answer?
- Are permissions enforced for agents the same way as for people?
- How do colleagues review and validate an analysis?
Red flags
- AI does the math
- Black-box answers with no SQL
- No source, calculation or confidence shown with an answer
- Agents that bypass user permissions
Continuous learning
What it means
The system learns from how people use it, the data, and how insight is used across the organization, so it gets more useful over time, and people learn alongside it.
Questions to ask vendors
- What does the system learn from usage, and how do users correct it?
- Does that learning carry across projects and teams?
- Can admins see and control what is retained?
Red flags
- Every session starts from zero
- No way to give feedback on an AI answer
- Learning that isn't visible or controllable by admins
Context without semantic layer engineering
What it means
Separate from the seven principles, check how the platform builds business context. The platform infers business context from your data and metadata when it connects, keeps learning your vocabulary as people work, and lets people review and correct that context. FERRARO says context must be decoupled from the math and validated by both machines and people.
Questions to ask vendors
- How long from connecting data to a trustworthy first answer?
- Where does business context come from, and can we inspect and correct it?
- How do you detect when context drifts?
Red flags
- Requires a semantic layer project first
- Answers are only reliable on pre-modeled data
- Context lives in the model's prompt, mixed up with the calculations
How to run the evaluation
Shortlist by architecture
Ask each vendor whether AI was built into the core or added to an existing product, and apply the remove-the-AI test.
Test on your own data
Connect a real dataset and watch what happens before anyone types a question.
Check one number end to end
Pick an AI-generated answer and trace it to the source data, the calculation and the SQL.
Fix one detail
Ask for a small change to an AI-built dashboard and see whether you can edit it directly or have to re-prompt.
Score and compare
Fill in the scorecard for each vendor and weight the principles that matter most to your team.
How Golden scores on the seven principles
Disclosure: Golden sponsored FERRARO's report, and we're the vendor writing this guide. Here's how FERRARO rated Golden, so you can weigh it alongside your own evaluation:
- People at the center: Exceptional (the Slider of Autonomy lets users choose how much agents do; familiar analytics features stay available)
- AI at the core: Exceptional (AI profiles and cleanses data the moment it connects)
- Principle 3: Excellent (support for MCP plus Open API and Search API access)
- Agents across the lifecycle: Excellent (data, analyst and storyteller agents start when data arrives)
- Conversation as one modality: Excellent (chat works with drag-and-drop data prep, interactive canvases and optional storytelling)
- Trust and validation: Good (an agent-data firewall, so probabilistic agents never touch the data and a deterministic engine does all math; editable SQL; real-time collaboration for peer review). FERRARO notes Golden has more work to do on high-fidelity governance.
- Continuous learning: Base
FERRARO's verdict: for CDOs moving into agentic analytics, "Golden is worth adding to your top three." See how Golden is built to be AI-native.
Choosing AI-native analytics: common questions
What is an AI-native analytics solution?
An AI-native analytics solution is built around AI from the start, so AI and people work on the same analysis through the same logic and the user decides how much the AI does. In AI-added tools such as copilots in Power BI, Tableau or Looker, the AI assists a product designed before AI. See What is AI-native BI? for the full definition.
What should I look for when choosing AI-native analytics?
Check seven principles from FERRARO Consulting's 2026 report on agentic analytics: people control how much the AI does, AI is at the core, agentic analytics sits at the center of every AI application, agents work across the lifecycle, conversation is one modality, answers are governed and verifiable, and the system keeps learning. Separately, check that business context doesn't require a semantic layer project.
How can I tell if a BI tool's AI is just a chatbot added on top?
Ask what happens if you remove the AI. If the product works the same, just less conveniently, the AI was added on top. Also check whether you can edit an AI-built dashboard directly, or only re-prompt.
Should the AI do the calculations in an analytics tool?
No. Language models are probabilistic, so look for tools where a deterministic query and calculation engine does the math, and where every answer shows its source data, calculation and SQL so you can verify it.
Do I need a semantic layer before using AI-native analytics?
Not necessarily. FERRARO's report says agentic analytics should deliver business context "without forcing organizations into semantic layer engineering projects," inferring context from data and metadata and letting people validate it. Ask vendors how context is built, inspected and corrected.
Is conversational analytics enough on its own?
Usually not. Chat is a good way to start, but most decisions still need visuals, dashboards, reports or stories to share. Look for a tool where conversation and hands-on editing work on the same analysis.