What does the data represent?
- A customer and a policyholder
- Written premium and earned premium
- A claim and a claim transaction
Insights · Data & AI
AI Needs More Than Data.
Why the semantic layer matters when organizations move from trusted data to usable intelligence.
01 - The Problem
AI can access data without necessarily understanding the business.
An insurer may have thousands of databases, millions of records, hundreds of reports, and years of accumulated business rules. TDWI's 2026 research finds that organizations cannot simply place a large language model on top of existing data and expect accurate, trusted, repeatable results. Generative BI requires an architecture designed for both people and AI.[1]
AI needs more than access.
Data provides the raw material. Business context, governed meaning, trusted metrics, and operational controls make that information usable by AI.
A simple business question
“What is our loss ratio?”
Which loss ratio? Is it written premium or earned premium? Are losses paid or incurred? Are allocated loss adjustment expenses included? What time period applies? Which business segment? Which policies are included?
Different systems may produce different answers, all of which appear technically correct. The problem isn't necessarily bad data. The problem is the absence of shared business meaning.
02 - What TDWI Research Shows
Generative BI is moving from experiments to enterprise capability.
TDWI's survey of 208 organizations shows that the business case is not primarily about replacing analysts. The leading drivers are productivity, faster insight generation, and better decision-making—and the research identifies business context as a strategic requirement for making those capabilities trustworthy.[1]
48%
cite productivity improvement as a primary driver for generative BI.
39%
cite faster insight generation as a primary driver.
37%
cite better decision-making as a primary driver.[1]
TDWI's central architectural point: the interface is not the foundation. Semantic layers, metadata, business glossaries, ontologies, knowledge graphs, governance, and trusted metrics provide the context required for meaningful and reliable generative BI.[1]
03 - Business Meaning
The semantic layer provides the bridge between technical data and business knowledge.[1]
What distinctions matter?
- An open claim and an incurred claim
- A policy cancellation and a policy non-renewal
- An operational metric and a financial metric
Why this matters
These distinctions may be obvious to an experienced insurance executive. They are not necessarily obvious to an AI system simply because the underlying data exists. TDWI describes the semantic layer as the mechanism that translates technical data into governed business concepts—metrics, dimensions, hierarchies, calculations, and relationships—and makes those definitions reusable across BI, AI applications, and agents.[1]
04 - From Data Architecture to Knowledge Architecture
Connect data to the language and logic of the business.
Traditional data architecture has focused on moving and storing information. AI introduces another requirement: access to the organization's understanding of its data.
The semantic layer should answer
- What does this data mean?
- Which definition should be used?
- How are business entities related?
- Which calculations are approved?
And establish
- Which data is authoritative?
- What business rules apply?
- Who owns the definition?
- How confident should we be in the result?
05 - The TDWI Blueprint
The semantic layer is critical—but TDWI places it inside a broader AI-enabled analytics architecture.
TDWI's Blueprint organizes generative BI around four interconnected layers. Business strategy and use-case selection guide the stack, while people and operating model, governance/security/trust, and evaluation/testing/operations apply across every layer.[1]
1. Data & Analytics Foundation
Data acquisition, integration, storage, quality, governed access, metadata, lineage, scalable performance, and integrated structured and unstructured data.
2. Business Context & Semantic Infrastructure
Agreed metrics and KPIs, semantic models, business glossaries, taxonomies, ontologies, knowledge graphs, and domain-specific context.
3. AI & Analytical Intelligence
LLMs, retrieval and grounding, semantic query translation, text-to-SQL, analytical tools, agent orchestration, and response generation.
4. Analytics Experience & Workflow
Conversational analytics, dashboards and reports, AI-generated narratives, recommendations, decision support, and increasingly AI-assisted and agentic analytical workflows. TDWI emphasizes that dashboards do not disappear; generative BI extends analytics beyond predefined views and into the workflow itself.[1]
Cross-cutting capabilities: business ownership and operating model; governance, security and trust; and continuous evaluation, testing and operations. TDWI's message is that generative BI is not a single-product deployment—it is a coordinated capability stack.[1]
06 - From Questions to Workflows
Generative BI is becoming a workflow capability, not just a question-and-answer interface.
TDWI describes a progression from natural-language questions toward AI-assisted analytical workflows: identifying the right metric, comparing periods, analyzing segments, investigating drivers, synthesizing evidence, and recommending next steps.[1]
Ask
Users can ask questions in business language rather than navigating every predefined report or writing SQL.
Investigate
AI can move beyond retrieval to compare results, examine segments, evaluate evidence, and explore likely drivers.
Act
Increasingly, AI can support recommendations and multi-step workflows—raising the importance of governed metrics, boundaries, testing, and human review.[1]
08 - Governance, Trust & Evaluation
As AI becomes more capable, governance becomes more operational.
TDWI identifies governance gaps, organizational silos, and lack of ownership/stewardship as the leading challenges to semantic consistency. The report also calls for representative question sets, “gold” answers, correctness and consistency testing, regression testing, monitoring, and incident management.[1]
Metric CertificationGovernance in the Query PathVersion ControlAuditabilityHuman ReviewRegression Testing
09 - The Executive Question
Does your organization have the foundation required for AI to produce reliable, repeatable business answers?
The harder question is not whether AI can answer a question.
Can the organization define the right metric, provide trusted data and business context, show where the answer came from, test whether it remains correct, and establish when AI should defer to a human? TDWI's research suggests that these complementary capabilities distinguish organizations moving from isolated experiments toward measurable business impact.[1]
10 - The Path to AI-Ready Data
Start with a business problem. Build the context. Then scale the intelligence.
Identify the critical decisions
What decisions do executives, underwriters, claims leaders, actuaries, and operations managers need better information to make?
Identify critical business concepts
Define the customers, policies, claims, premiums, exposures, expenses, losses, and other concepts that support those decisions.
Establish common definitions
Create agreement about what those concepts mean and how they should be measured.
Connect definitions to trusted data
Map business concepts to the underlying systems, data elements, calculations, and lineage.
Govern the semantic layer
Assign ownership, establish change processes, and continuously monitor quality.
Prepare enterprise information
Bring structured and unstructured information together through metadata, enrichment, retrieval, classification, and governance where the use case requires it.[1]
Evaluate, govern and monitor
Define representative questions and expected answers, test correctness and consistency, monitor drift and cost, and maintain human oversight for high-impact use cases.[1]
Make semantics available to AI
Expose approved business definitions, relationships, metrics, and rules to the AI systems and agents that need them.
11 - What the Research Says About Business Impact
Business impact appears to come from the combination of capabilities—not a single AI feature.
Trusted foundations
Higher-impact organizations are more likely to have trusted AI outputs, semantic and business-context technologies, prepared unstructured data, and governance and observability practices.[1]
Complementary capabilities
TDWI's analysis does not identify one product or technology as the answer. Business value is associated with the accumulation of complementary capabilities across the architecture.[1]
From pilots to production
The journey commonly expands from conversational access to trusted business context, prepared enterprise information, operational governance, and ultimately AI-native and agentic analytics.[1]
12 - AIA Perspective
AI readiness starts with the organization's trusted data, business context, governance, and decisions — not with an AI tool.[1]
Continue the Conversation
Start with a business problem. Build the context. Then scale the intelligence.
Then build the trusted data, business context, governance, and evaluation capabilities required to move from AI experimentation to reliable business use.
Research Sources
Source note: All statistics and research findings above are from TDWI's 2026 Blueprint Report based on a survey of 208 organizations, focus groups with data and analytics leaders, and industry-expert interviews. AIA's insurance examples and recommendations are its own framing of the research.