Skip to content
ESEN
Let’s talk about your data
Menu

Data governance · AI with context

The semantic layer.
Your data, under your control.

Unify your sources and the meaning of your metrics. Decide who can access what and which information reaches each LLM. A shared foundation for your business, your teams and your AI.

Let’s talk about your dataSee how it works

For CTOs, CIOs and CEOs bringing AI to their business data.

From your sources to authorised context
ERP and banks
CRM and sales
Internal data
Semantic layer
Shared metricsPermissionsMinimisation
Only the context neededDashboards · Teams · LLMs
Reference architecture. Scope is defined with your team.

01A shared business language

02Access according to the task

03Traceability from source to answer

The foundation of your decisions

The same definition.
In your dashboard and your AI.

NextScenario connects banks, ERPs and sales channels to understand cash, profitability and growth. The semantic layer brings that shared view to your analysis and assistants: what each data point means, where it comes from and how it can be used.

One business definition

Net revenue

Sales − returns − discounts

Source, period, currency and calculation rules defined before querying the model.

DashboardForecastLLM

01 / Architecture

From scattered data
to governed context.

  1. 01Connect

    Bring relevant sources together

    Banks, ERP, CRM, ecommerce and internal databases. We start with your stack and validate the connections each use case needs.

  2. 02Define

    Give every data point meaning

    Shared entities, relationships and metrics. Agree how to calculate margin, revenue or churn and resolve differences between sources.

  3. 03Govern

    Apply access rules

    Define permissions by role, company and field. Filter records and transform sensitive data before building query context.

  4. 04Deliver

    Prepare what the AI needs

    Select metrics, aggregate results and limit detail. The model receives authorised context for the task and its response is validated before use.

02 / What reaches the model

Explaining margin
does not need an email.

Explore two examples of how context changes with the task. Values are fictional and illustrate a possible policy, not a connection to your systems.

02 / What reaches the model
01 / Data in your environment
{
  "customer_name": "Cliente de ejemplo",
  "email": "cliente@example.com",
  "customer_id": "customer_042",
  "channel": "ecommerce",
  "revenue_eur": 1200,
  "cost_eur": 780
}
Excluded: nameExcluded: email
02 / Context for the LLM
{
  "metric": "gross_margin",
  "channel": "ecommerce",
  "period": "2026-08",
  "currency": "EUR",
  "revenue": 1200,
  "cost": 780,
  "margin_pct": 35
}
Aggregatedscope: read

Comparing profitability by channel needs aggregated amounts. Names, emails and customer identifiers stay out of the context.

01 / Data in your environment
{
  "customer_name": "Cliente de ejemplo",
  "email": "cliente@example.com",
  "customer_id": "customer_042",
  "orders": 3,
  "revenue_eur": 1200
}
Excluded: nameExcluded: email
02 / Context for the LLM
{
  "customer_ref": "hmac:7b4e…9c2a",
  "period": "2026-08",
  "orders": 3,
  "revenue_eur": 1200
}
Pseudonymisedscope: read

Linking purchases from the same customer may require a pseudonymous identifier. The model does not need to receive their identity.

Illustrative example · fictional data

03 / Governance by design

You define the rules.
AI works within them.

We design controls with you and validate how they apply to each source, user and use case.

01 /

Permissions before the prompt

Access by role, company, record and field. Authorisation is checked by the application before retrieving data or invoking tools.

02 /

Only the data needed

Field allowlists, filters and aggregation. A task that can be answered with totals does not need complete records or entire documents.

03 /

Hashing and pseudonymisation

Remove or mask identifiers. When relationships must be preserved, consider tokens or HMAC with keys managed outside the model.

04 /

Provenance and traceability

Link metrics to sources and calculation rules. Define access and transformation logs, including retention and permissions for the logs themselves.

05 /

Control over destinations

Agree which models may receive each data category and review the provider’s residency, retention and usage terms.

06 /

Bounded actions

Separate queries from execution. Define allowed tools and human approval for sensitive operations; validate outputs before acting.

04 / Technology and leadership

One shared foundation.
Three clear priorities.

CTO

An architecture you can govern.

Centralise definitions and policies to avoid a different integration for every assistant.

  • Reusable data contracts and metrics
  • Permissions outside the model
  • Integration with your existing stack
CIO

Visibility into the data journey.

Define owners, access and destinations to introduce AI with a shared policy.

  • Source inventory and classification
  • Traceability and retention criteria
  • Control over vendors and access
CEO

Decisions with shared context.

Connect AI initiatives to the metrics that drive cash, margin and growth.

  • Consistent numbers across teams
  • Use cases with a defined scope
  • Oversight of significant decisions

Technical questions

Before connecting
your first LLM.

What is a semantic data layer?

A semantic layer defines what data means: shared entities, relationships and calculation rules. Integrated with access controls, it lets teams and AI assistants query consistent metrics using only the information authorised for each task.

Is this another database?

The semantic layer defines entities, metrics, relationships and rules over your sources. It can build on your existing data infrastructure. Whether information needs to be copied or stored depends on the architecture and use case.

Does the LLM have access to all my data?

The design starts with minimum access: first authorise the query, then select and transform the data it needs. A permission in a prompt does not replace the controls that must be enforced in infrastructure and the application.

Does hashing make data anonymous?

Not necessarily. A deterministic hash can enable comparisons or dictionary attacks. Depending on the use case, consider removal, aggregation, tokenisation or HMAC with key management. Keeping pseudonymous identifiers does not by itself eliminate reidentification risk.

Is my data used to train models?

That depends on the provider, service and contract. Before enabling a destination, its training and retention terms must be verified and configured. The same policy should not be assumed for every LLM.

How do we get started?

Choose a business question, identify its sources and define the metric, permissions and allowed context. Validate the flow with your team before expanding to new data or automations. Specific connections and controls are agreed in the technical scope.

Your next scenario

Give your AI context.
Keep control of your data.

Let’s review your stack, one business question and the information the model actually needs to answer it.

Let’s talk about your data

Sources · Metrics · Permissions · Context