1even.lab

Whitepaper · EN · v1.0

Agentic Analytics

The CAPTURE Framework

A Practical Framework for Implementing Agentic Analytics

  1. CConnect
  2. AApply
  3. PProve
  4. TTeam Scale
  5. UUnify
  6. RRestrict & Govern
  7. EExpand

Connect. Play. Scale.

Introduction

Introduction

Analytics does not exist to produce reports, dashboards, SQL queries, or research. These are instruments. The ultimate product of analytics is a better management decision.

Over the past fifteen years, companies have steadily become more data-driven. They have collected more data, built warehouses, BI platforms and data products, hired analysts, data engineers and data scientists, adopted machine learning, and developed data governance practices. For digital businesses with large volumes of customers, transactions and events, data has become one of the primary ways in which an organization observes its own activity and makes decisions.

But this model has a fundamental characteristic: as the number of decisions a company wants to support with data grows, the analytical capacity around those decisions usually has to grow as well. Organizations need specialists who understand data structures, can write queries, know statistical and analytical methods, can interpret results, and can translate a business question into a technical investigation.

Some of this demand scales through dashboards, standardized reporting, data products and automation. But new, ambiguous and complex questions still tend to pass through a human analytical cycle: formulate the task, transfer context, wait for the investigation, clarify the question, validate the result and discuss the conclusions.

Modern LLMs and AI agents change the economics of this model. For the first time, broad analytical, statistical, programming and research expertise is accessible through natural language. Give an agent governed access to corporate data, business semantics and analytical methods, and it can do more than generate SQL. It can perform a multi-step investigation: understand the question, determine a plan, access data, test hypotheses, inspect intermediate results, change the direction of the investigation and formulate an evidence-based conclusion.

This is the model we call Agentic Analytics.

The CAPTURE Framework describes a practical path for implementing it: from connecting the first analyst to corporate data to creating a governed analytical capability that can be accessed directly by decision-makers.

Connect. Play. Scale.

The principle is deliberately pragmatic. Do not begin with months of designing a new AI platform. Connect a small group of specialists, start solving real problems, learn from failures, formalize what the organization learns — and then scale what works.

Part IWhy Analytics Is Changing

01

Why Analytics Is Changing

1.1The Ultimate Product of Analytics Is a Management Decision

Organizations are used to measuring analytics through production metrics: dashboards delivered, studies completed, requests closed, BI adoption, SLA performance and infrastructure cost. These measures are useful, but they describe the production of analytics rather than its ultimate value.

Analytics creates value when it helps a person understand what is happening, evaluate alternatives and make a better-informed management decision.

The two most important dimensions of analytical effectiveness are therefore Decision Velocity and Decision Quality: how quickly the organization can move from a signal or question to action, and how well-founded the resulting decision is.

The full decision cycle can be represented as:

Awareness
→
Insight
→
Decision
→
Action
→
Impact
Decision Velocity×Decision Quality
Figure 01 · The management decision cycle

Awareness begins when the organization detects an event or deviation. Insight is reached when there is sufficient understanding of what happened and why to choose an action. A Decision is then made, converted into Action, and ultimately evaluated through its Impact.

Sometimes this entire cycle takes minutes. For strategic, cross-functional or complex decisions, days or weeks may separate the stages.

Analytics has its most direct effect on the path from awareness to insight and from insight to decision. The fewer organizational and technical barriers between a question and a high-quality investigation, the more useful analytical work can fit inside the decision window.

1.2The Constraint in the Traditional Model Is Not the Absence of Knowledge — It Is the Cost of Applying It

Traditional analytics is capable of preserving knowledge. Mature organizations use data governance, data catalogs, documentation, Confluence, dbt, metric definitions, data contracts, best practices and knowledge-transfer processes. It would be inaccurate to claim that knowledge is inherently lost in the traditional model.

The constraint lies elsewhere.

Applying that knowledge to a new business question usually requires a specialist who can connect business context with the technical structure of the data and an appropriate analytical method. Analytical expertise therefore scales primarily through people: hiring, development, specialization and the organization of their work.

As companies grow, analytical organizations often move through cycles of centralization and decentralization. A centralized team can concentrate expertise, standardize technology and manage quality. As demand grows, it can become a shared backlog and lose speed. Analysts are then distributed closer to Product, Marketing, Finance, Risk or Operations. Over time, technology and practices may fragment, definitions can diverge, and the organization develops a renewed need for standardization.

Centralization
→
Growing backlog
→
Decentralization
→
Fragmentation
↻ re-centralization
Figure 02 · Centralization–decentralization cycles

These cycles are not necessarily management failures. They reflect a real tension between speed, specialization and standardization. But each transition carries a cost: management overhead, coordination, onboarding, changes in responsibility, redistribution of people and sometimes the loss of tacit expertise.

Agentic Analytics does not replace data governance or make organizations independent of people. It changes how accumulated knowledge is applied. Users need to know less about the physical structure of the database, SQL syntax or the technical implementation of a calculation. They interact with the analytical capability through the language of the business.

1.3Access to Analytical Expertise Becomes Programmable

In the previous paradigm, natural language was primarily an interface to another person. The business formulated a question for an analyst, and the analyst translated it into the language of data and methods.

Natural language is now becoming an interface to analytical expertise itself.

Before
Business Question
→
Human analytical cycle request → queue → research → iterations
Now
Business Question
→
Agentic analytical capability the investigation starts immediately
Figure 03 · Natural language as an interface to expertise

Modern models can already write SQL and Python, apply statistical methods, create visualizations, decompose problems, propose hypotheses and interpret results. Agents add the ability to act: access sources, execute queries and code, inspect intermediate outputs and continue the investigation.

This is fundamentally different from Text-to-SQL.

Text-to-SQL asks: "How do I turn this text into a query?"

Agentic Analytics asks: "How do I conduct the investigation required to answer this business question?"

The unit of work is no longer the SQL query — or even the report. It is the analytical task.

1.4Analysts Are Augmented First

The practical transition to Agentic Analytics does not begin by replacing the analytics function or immediately giving the entire business direct access.

Analyst
→
+ Analyst + Agent
SQL Research Decomposition Modeling Visualization
Figure 04 · The analyst is augmented first

The first transformation happens inside the analytics team.

Each analyst gets an agent that can accelerate SQL, data exploration, decomposition, modelling, visualization and a large amount of repetitive technical work. Tasks that previously required hours or days can, in many cases, be reduced to minutes or hours.

The more important change, however, is in the analyst's role. As the agent takes on more execution, analysts increasingly focus on what makes the analytical capability reliable: validating conclusions, refining business semantics, formalizing methods, identifying edge cases and converting their expertise into reusable knowledge.

The analyst gradually stops being the "human runtime" through which every analytical request must pass. The analyst becomes one of the engineers and stewards of a shared analytical capability.

1.5The Democratization of Analytical Expertise

Once analysts have taught the system to understand corporate data and business context with sufficient reliability, the capability can be made directly available to business users.

This is the democratization of analytical expertise.

It does not mean that every executive must become a data scientist. Quite the opposite: a decision-maker needs to know less about the physical data model, SQL syntax or the implementation details of statistical methods. The essential skills become understanding the business, asking the right questions, critically evaluating conclusions and making decisions.

Expertise that previously had to be delivered by a person every time can increasingly be embedded in the system.

This changes the marginal economics of analytics. A new question no longer necessarily creates another human analytical cycle of comparable cost. Once business semantics, skills and methods have been formalized, they can be reused repeatedly by different users and agents.

It also changes the questions that get asked. In the traditional model, decision-makers often filter their own curiosity through the expected cost, delay or effort of requesting an analysis. Many hypotheses remain unexplored because they do not justify a separate analytical task. When analytical capability is available directly in the working interface, questions that previously stayed in a decision-maker's head can be explored immediately.

Agentic Analytics does not only answer existing analytical demand. It unlocks analytical questions that previously were never asked.
Analyst expertise
→
SemanticsSkills · Methods
→
Shared analytical capability
→
Decision-makers
Figure 05 · Expertise flow: from analyst to decision-makers
Part IIWhat Is Agentic Analytics

02

What Is Agentic Analytics?

Agentic Analytics is an approach in which AI agents use governed access to corporate data, business context and analytical methods to perform multi-step analytical work in response to a business question.

The user provides a business question rather than a technical specification for the analysis.

For example:

Why did NGR decline last week despite an increase in active users?

To answer it, an agent may determine which metrics need to be decomposed, identify the appropriate sources and time periods, execute multiple queries, compare segments, test alternative explanations, identify a shift in player mix, build a visualization and formulate a supported conclusion.

A typical working cycle looks like:

Business Question
→
Understand
→
Plan
→
Query & Analyse
→
Test
→
Iterate
→
Explain
← Test / Iterate → Plan
Figure 06 · The agentic analytical workflow

The defining characteristic is not autonomy for its own sake. The agentic nature comes from the ability to perform a sequence of connected actions, interpret intermediate results and adjust the next step of the investigation.

The human remains in the loop. Depending on the maturity of the system, the human formulates the question, validates the result, determines what actions are permissible and makes the management decision.

Agentic Analytics is not Text-to-SQL. It is the delegation of analytical work to an AI agent.

03

The Agentic Analytics Architecture

For the user, mature Agentic Analytics can look almost trivial: open a familiar AI interface, ask a question and receive an investigation.

Security & Governance
Userasks questions in natural language
↓Authenticated & Secured Access
NATIVE AI INTERFACE Claude · Codex · Cursor · Slack / AI Bot
↓
ANALYTICS HARNESS skills · workflows · SQL recipes · instructions
↓
BUSINESS SEMANTICS metric definitions · business logic · golden sources
↓
CORPORATE DATA & ACCESS databases · marts · BI
Identity Permissions Data Protection Audit Versioning Quality Controls
Figure 07 · The layers of Agentic Analytics architecture

Underneath that simple interaction sits a layered analytical capability.

3.1Corporate Data & Access

The foundation remains the company's actual data: databases, warehouses, data marts, BI systems and other corporate sources.

Agentic Analytics does not automatically fix poor data and does not eliminate the need for sound data architecture. But the first implementation does not require a perfect platform. The capability can start with the data structures that already exist, while their weaknesses are progressively discovered through real analytical work.

Agent access to data must be governed and technically controlled.

3.2Business Semantics

A physical database structure is not enough. The agent needs to understand what the data means to the business.

Business Semantics includes metric definitions, business logic, authoritative or golden sources, relationships between entities, filtering rules, domain-specific conventions and other knowledge required for correct interpretation.

For a business user, the logical model of the business matters more than the physical storage model. The semantic layer becomes the bridge between the two.

3.3Analytics Harness

Above business semantics sits the Analytics Harness — a governed environment for analytical knowledge and methods.

It may include:

  • analytical skills;
  • research workflows;
  • SQL recipes;
  • domain instructions;
  • routing rules;
  • metric definitions;
  • reusable code;
  • tests and evaluations;
  • instructions for specific sources and classes of analytical tasks.

The Harness answers not only what the data means, but also how the data should be used to solve particular analytical problems.

As it matures, the Harness becomes a version-controlled engineering asset: changes are stored in Git, reviewed, tested and deployed into the agent environments used by the organization.

3.4Native AI Interface

Users do not necessarily need a separately developed analytics application.

The capability can be exposed through environments in which employees already work with AI: Claude, Codex, Cursor, an enterprise AI interface or, where appropriate, a Slack bot.

This is an important implementation principle:

Bring the analytical capability into the user's working environment instead of automatically building another analytics interface.

For many users, native proprietary AI interfaces can also be a cost-effective deployment model. They allow the organization to use the full capability of leading models through existing per-user enterprise or professional subscriptions, rather than routing every interaction through metered API usage and maintaining a custom platform. This is not universal — economics depend on provider terms, usage patterns, governance requirements and integrations — but it should be evaluated before committing to an API-heavy custom architecture.

3.5Security & Governance

Security & Governance is not simply another layer above or below the stack. It forms a control boundary around the entire architecture.

It includes identity, authentication, permissions, data protection, auditability, secrets management, versioning, quality controls and governance of the connectors and AI environments being used.

The core principle is straightforward:

A user should never gain access through an agent to data they could not legitimately access without the agent.
Part IIIThe CAPTURE Framework

04

The CAPTURE Framework

CAPTURE is a practical sequence for implementing Agentic Analytics.

It emerged not as a theoretical model of an ideal architecture, but as a generalization of a real implementation process: first make the capability work within a constrained environment; then teach it through real questions; expand validation across the analytics team; structure the accumulated knowledge; and only then democratize access to business users.

Connect. Play. Scale.

The key difference between CAPTURE and a "build the platform first" approach is sequence. Semantics and the Harness are not designed completely in a vacuum. A significant part of them is grown empirically from real analytical work.

C
Section 5 · Stage 01 / 07

CONNECT

Connect a Small Trusted Data Team to Real Corporate Data

1–3 Data Specialists
↓
Claude · Codex · Cursor
↓
MCP / Connector
↓
Existing Corporate Data
No platform rebuild required
Figure 09 · The first connection

Start with one analyst or a small group of trusted data specialists. They may be analysts, analytics engineers or data engineers — people who already understand the company's data and have legitimate access to it.

Connect their existing AI environment — for example Claude, Codex or Cursor — to corporate data through MCP or another governed connector.

Do not begin by rebuilding the data platform.

If the company already has well-organized marts, dbt, a data catalog, documented metrics, data-quality controls and golden sources, that will make the process significantly easier. If it does not, imperfect data maturity should not become a reason to postpone the experiment for a year.

In some organizations, an existing BI tool can be the first source. Direct governed access to the database, however, generally gives the agent more investigative freedom: it can move beyond aggregated views and work with deeper layers of data.

The initial setup only needs enough bootstrap context to begin: descriptions of key tables, trusted sources, a few rules about where to look and what to avoid, and basic business semantics.

Security begins here as well, although the pilot threat model is different from enterprise rollout. Use corporate accounts, approved AI providers and connectors, controlled credentials, appropriate treatment of sensitive data and no unnecessary write access. A pilot should not rely on an employee's personal AI account with uncontrolled production access.

The objective of CONNECT is not to build the system. It is to get the first agent working safely with real company data.
A
Section 6 · Stage 02 / 07

APPLY

Start Playing with Real Analytical Work

Once the connection works, the PLAY phase begins.

There is little reason to invent synthetic benchmark questions when real analytical demand already exists. The analyst starts giving the agent their own tasks and genuine requests from the business.

Slack request
Jira task
Analyst question
Screenshot
→
PLAY
→
Agent Investigation
→
Answer
SQL
Evidence
Visualization
Real work = real test cases
Figure 10 · The flow of real demand

A request can often be passed in almost its original form: a Slack message, Jira ticket, email, screenshot or the analyst's own question.

The agent attempts to do the work: understand the problem, identify sources, write SQL or code, inspect the data, formulate hypotheses and return a result.

The first answers will almost inevitably expose problems. The agent may choose the wrong table, misunderstand a field, use an outdated source, make an incorrect join, miss an internal metric definition or misinterpret a business term.

This does not mean the experiment has failed. These errors reveal which pieces of corporate context the system is missing.

Lightweight instructions therefore begin to emerge naturally: table descriptions, source-selection rules, basic organizational skills, metric definitions and routing instructions.

The important principle is not to describe the entire organization in advance.

Do not design Agentic Analytics in theory. Let real analytical work teach you what the system needs to know.

Every real question becomes both useful analytical work and a test case.

P
Section 7 · Stage 03 / 07

PROVE

Verify, Correct and Teach

During the early stages, every material agent answer should be treated as a hypothesis that the analyst needs to verify.

Verification can take several forms. Compare the result with existing BI, a known figure, manually written SQL or another trusted reference. Sometimes it is enough to ask the agent to expose the generated SQL and inspect its logic.

When the result is wrong, the goal is not merely to obtain the correct number in the current conversation.

The goal is to find the cause of the error.

It may be the wrong source, join, filter, metric definition, business assumption or analytical method. Once the cause is identified, the relevant semantic definition, skill or instruction is corrected.

Then the original question is run again.

The working loop is:

Real Question
→
Agent Answer
→
Verify
→
Find the Gap
→
Correct Semantics / Skill
→
Retest
← Retest → Real Question
Figure 11 · The verification loop

A correction is complete not when the analyst manually tells the agent the right answer, but when the agent can reproduce the right analytical approach using the updated knowledge.

Individual mistakes gradually become improvements to the shared capability.

Every mistake should make the analytical capability better.

This is why:

Verification is not only quality control. It is how the system learns the organization.

The system is not "learning" here through retraining the weights of the foundation model. It learns the organization through accumulated governed corporate context: semantics, instructions, skills, examples and reusable analytical logic.

T
Section 8 · Stage 04 / 07

TEAM SCALE

Expand the Working Setup Across the Analytics Team

Once the initial setup performs reliably across a meaningful set of real questions, expand it across the analytics team.

This creates an immediate productivity effect: more analysts get an agent and begin using it in daily work.

+ Pilot
→
Team Scale
More analytical work·More real-world test cases·More captured knowledge
Figure 12 · Scaling to the team

But productivity is only half the value of this stage.

Every new analyst brings additional domains, question types, validation practices and accumulated business expertise. The number of real-world test cases increases dramatically. Residual inaccuracies appear that the original pilot group could never have encountered alone.

Analysts continue validating answers against real business requests and correcting source selection, definitions, business semantics and analytical logic. Their domain knowledge is progressively converted into shared instructions and reusable skills.

The analytics-team rollout therefore serves two purposes at once:

  1. it increases the productivity of the analytics function;
  2. it creates a distributed learning and refinement environment for the shared analytical capability.
The analytics team does not only use the system — it continues to teach the system how the business actually works.

More analysts → More real questions → More corrections → Better shared analytical capability

U
Section 9 · Stage 05 / 07

UNIFY

Turn Accumulated Knowledge into a Governed Analytics Harness

After several cycles of use, the amount of accumulated knowledge becomes substantial.

There are table descriptions, metric definitions, corrections, domain rules, analytical skills, SQL examples, reusable code and instructions created by different analysts.

Organic learning
corrections prompts table notes analyst knowledge skills SQL examples
Unify→
Analytics Harness Business Semantics Analytical Skills SQL Recipes Rules & Instructions Git / Versioning / Tests
Figure 13 · From organic knowledge to a structured Harness

At the beginning, some organic messiness is useful because it allows the team to move quickly. As usage scales, however, a lack of structure creates duplication, contradictions and maintenance problems.

This is the point of UNIFY.

Accumulated knowledge is structured into the Analytics Harness. Different knowledge types are separated:

  • what a field or table means;
  • where the authoritative source is;
  • how a metric is calculated;
  • how a particular type of investigation should be performed;
  • which executable logic is canonical;
  • which rules belong in every agent context;
  • which knowledge is relevant only to a particular domain or task.

The Harness moves into Git and begins to evolve as an engineering asset: version history, review, tests and controlled deployment.

The key architectural principle is:

Every fact has one owner. Never copy — reference the owner.

If a metric definition belongs to the metric-definition layer, it should not be independently copied into five skills and three prompts. If executable logic belongs to a SQL recipe or code asset, other components should reference it.

A second principle is progressive disclosure. The agent does not need the entire corporate knowledge base loaded into every context. A minimal always-on layer contains routing and invariants; domain semantics, recipes and skills are loaded when required.

What the analytics team learns through hundreds of real questions becomes reusable organizational analytical knowledge.
R
Section 10 · Stage 06 / 07

RESTRICT & GOVERN

Prepare the Analytical Capability for Secure Business Access

Until this point, the primary users have been trusted data specialists who often already have broad access to warehouses and analytical systems.

Before democratization, the threat model changes.

Audit · PII Protection · Secrets · Approved AI Environment
Marketing
Finance
Product
permitted data
→
Identity + Permissions+ Governed Connector
→
Analytics Harness
→
Corporate Data
Pilot security ≠ Enterprise-scale security
Figure 14 · Differentiated access

The capability will now be used by Product, Marketing, Finance, Risk, Operations and other business users with different access rights. The organization cannot simply give everyone a connector carrying an analyst's privileges.

Each user should operate through corporate identity and individual credentials. Access should follow least privilege and, where required, be enforced through database roles, row-level and column-level policies and other existing enterprise controls.

The organization also needs explicit policies for PII and confidential data, secrets management, auditability, user attribution, approved AI environments and connector governance.

It is particularly important to separate policy explanation from security enforcement.

A model can be told that a user should not see certain information. But a prompt is not a security boundary.

Enforce by configuration. Explain by prompt.

Where a restriction matters, it must be enforced by infrastructure, permissions and configuration.

The core test is simple:

The agent should never give a user access to data they could not legitimately access without the agent.

Enterprise-grade security appears in CAPTURE immediately before business rollout not because security was ignored earlier. Minimum guardrails exist from CONNECT. At this stage, however, the architecture must support a broad population of users with differentiated permissions and controlled scale.

E
Section 11 · Stage 07 / 07

EXPAND

Bring Proven Analytical Capability Directly to Decision-Makers

Identity · Permissions · Governance · Audit
Decision-makers
↓
Claude · Codex · Slack · AI Interfaces
↓
Analytics Harness
↓
Business Semantics
↓
Corporate Data
Figure 15 · The complete analytical capability

Once the capability has been validated by analysts, structured and prepared for differentiated access, democratization begins.

The first business users should not necessarily be everyone in the company. Start with people who combine high analytical demand with real decision responsibility — often leaders and specialists in Product, Marketing, Finance, Risk or Operations.

Give them access through familiar AI environments: Claude, Codex, an enterprise AI interface, a Slack bot or another approved environment.

A connector alone does not create adoption.

Users need basic AI literacy: how LLMs work, why a confident answer can still be wrong, how to ask follow-up questions, how to inspect evidence, when to involve an analyst and where human judgment remains essential.

The most effective enablement format is a hands-on workshop using the user's own questions and real company data.

The goal is not to "teach prompting". The goal is to establish a new working pattern: a business question can be investigated directly in dialogue with the analytical capability.

Democratization does not mean giving everyone access to everything. It means giving every decision-maker direct access to the analytical capability they are authorized to use.

EXPAND does not end the system's learning. New business users bring new questions and edge cases. These expose new gaps, which are closed through verification, analyst correction and Harness updates.

CAPTURE is therefore both a rollout framework and a continuous learning loop.

Part IVResults and the Path

12

Operating Principles

CAPTURE defines the implementation sequence. Several principles determine the quality of the system regardless of stage.

12.1Real Work First

Real business questions are the best source of requirements. They reveal which semantics are actually needed, which errors matter and which skills have recurring value.

12.2Verification Creates Organizational Learning

An error should become a reusable improvement rather than a one-off correction in a chat.

Question → Verification → Root Cause → Knowledge Change → Retest

12.3Every Fact Has One Owner

Each fact or rule should have an authoritative layer.

Knowledge type Authoritative owner —————————- ——————————————-- Field / table meaning Database metadata / semantic documentation Metric formula Metric definition Analytical procedure Skill Canonical executable logic SQL recipe / code Routing / invariants Agent instructions Access control Infrastructure / database configuration

Never copy. Reference the owner.

12.4Git Is the Source of Truth for the Harness

A mature Harness should be version-controlled. Claude, Codex, Cursor and other agent environments are deployment targets, not independent sources of truth.

12.5Progressive Disclosure

Do not load the entire organizational context into every conversation. Keep the always-on context minimal and load domain knowledge and skills as needed.

This reduces context noise and the risk of conflicting instructions.

12.6Enforce by Configuration

Security, permissions and critical restrictions should be enforced technically. A prompt may explain a rule, but it does not replace access control.

12.7Do Not Build the Platform Before Learning the Problem

Architecture should become more rigorous as real usage reveals what actually needs to be standardized.

CAPTURE deliberately starts with controlled experimentation and only then turns accumulated practice into a platform capability.

12.8Use Native Interfaces Where They Are the Better Economic Choice

A custom AI platform is not automatically the most efficient architecture. If approved native AI environments already provide strong models, enterprise controls and the interfaces employees use every day, they may offer a materially simpler and more cost-effective route to adoption.

A custom API-based layer is justified when the organization needs capabilities that native interfaces cannot provide — for example specialized workflows, centralized orchestration, bespoke UX, automation or integration requirements.

The architectural decision should therefore be based on total economics and control requirements, not on an assumption that Agentic Analytics requires a custom front end.

13

What Changes After CAPTURE

For Decision-Makers

Decision-makers gain direct access to a governed analytical capability. Investigation can begin when a question arises rather than after creating a request and waiting for analyst capacity.

More hypotheses can be explored within the same decision window, and dependence on analytical queues is reduced.

Just as importantly, the volume of analytical demand itself can expand. Questions that previously remained unasked because they seemed too small, too exploratory or too expensive to send through a formal analytical process can now be investigated directly.

For Analysts

Agents become part of daily analytical work. Analysts gain leverage, while their attention shifts from repetitive execution toward complex research, methodology, quality, semantics and development of the Harness.

Their role does not disappear. It becomes more engineering-oriented and more scalable.

For the Organization

The most important change is in the economics of analytical expertise.

Previously, much of that expertise was available to the organization at the moment when a specific specialist applied it to a specific question. Increasingly, that expertise can be formalized and reused through a shared capability.

This lowers the marginal human effort required to support an additional decision and allows the organization's analytical capacity to grow faster than analytical headcount.

In the target state, both core dimensions improve:

Decision Velocity ↑
Decision Quality ↑

Analytics moves from a service delivered by a function towards a capability embedded across the organization.

This does not mean the analytics function disappears. It means the way its expertise is distributed across the organization changes.

14

Agentic Analytics and the AI-Native Transformation

Agentic Analytics sits within a larger shift.

Companies are moving from isolated AI use cases toward a model in which employees continuously work alongside agents, while AI receives governed access to corporate context, knowledge, data and working tools.

If starting now
Agentic Analytics
→
Governed human + agent workflows
→
AI-native transformation
If already underway
Existing AI initiatives + Agentic Analytics
→
Shared AI-native foundations
Corporate Context·Machine-accessible Knowledge·Identity & Permissions·Agent Governance
Figure 16 · Two paths to a shared foundation

For companies that have not yet begun a meaningful AI-native transformation, Agentic Analytics is one of the most practical entry points.

The reason is simple: analytics already contains many of the necessary ingredients — corporate data, experts, digital workflows, measurable outcomes and a clear connection to management decisions. It allows the organization to move from an abstract conversation about AI to real operational practice.

For companies already applying AI in engineering, customer support, operations or other functions, Agentic Analytics is not "the first step". It becomes one of the foundational capabilities connecting AI to governed corporate context and real decision-making.

The deeper objective in both cases is the same: not to add AI to a collection of isolated tasks, but to progressively establish the foundations for human + agent workflows across the organization.

Corporate knowledge needs to become more structured and machine-accessible. Identity and permissions need to support agent-mediated access. The organization needs to learn what can be delegated, what must be verified and where human judgment remains essential.

Agentic Analytics provides a way to learn these patterns within one of the most understandable and measurable functions of the company, and then apply the lessons to Finance, Risk, Product, Operations and other knowledge-intensive areas.

What matters now is not merely whether a company has AI initiatives, but how quickly it accumulates this organizational learning.

If the AI-native transformation has not started, Agentic Analytics is an optimal place to begin. If it is already underway, Agentic Analytics is a foundational capability worth building alongside it.

This is the time to establish the foundations. While new operating patterns are still taking shape, organizations that learn earlier how to build governed human + agent workflows will be positioned to scale them faster as models and tools continue to improve — and to move faster than the market.

15

Where to Start

Despite the scale of the potential transformation, the first step can be very small.

You do not need to begin with a new platform.

You do not need to document the entire company's semantics in advance.

You do not need to connect hundreds of users on day one.

  1. Start with someone who knows your data well.
  2. Give them an approved AI environment.
  3. Connect it to real corporate data.
  4. Take a real business question.
  5. See what happens.
  6. Verify the answer.
  7. Find the gap.
  8. Teach the system.
  9. Repeat.

Then scale what has proven to work.

Start with one analyst.

Analytics is no longer just a function.
It becomes a capability available across the organization.

Connect. Play. Scale.

The CAPTURE Framework