---
title: "Semantic layer for HR and people analytics"
url: "https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-for-hr-data/"
description: "A semantic layer for HR data writes each people metric once: definition, as-of rule, access rule and minimum group size, read alike by dashboards and agents."
author: "Emily Winks"
author_role: "Data Governance Expert"
published: "2026-11-17"
updated: "2026-11-17"
---

## What is a semantic layer for HR data?

A semantic layer for HR data writes each people metric once: its definition and the population it counts, its as-of date rule, its access rule and its minimum group size. Every reader of the metric, whether a dashboard, a people analytics tool or an AI agent, then counts the same people for the same date under the same rules. Atlan's AI Context Platform records the definition as a certified metric with a named owner, and agents reach it through Atlan MCP, a Model Context Protocol server.

The 4 parts of a governed people metric:

- **A definition** that names the population it counts
- **An as-of rule** for history that changes
- **An access rule** that holds for people and agents
- **A minimum group size** for every drill-down

## Key takeaways

- "Attrition" can mean quits, layoffs and discharges, or other separations. Name the population on the metric.
- Headcount as of a past date needs rows that carry valid_from and valid_to dates, and a named reporting date.
- Row rules alone do not stop a small-team drill-down from identifying a person. Add a minimum group size.
- Score one people question before an agent answers it. The lowest dimension is the result, not the mean.

---

A semantic layer for HR data gives HR and analytics leads one answer to "how many left?" across every dashboard and AI agent. Without it, finance and HR each report "attrition" from a different population, and a manager drill-down on a small team can reveal who left. Atlan's AI Context Platform records each people metric as a glossary term with its formula in a dedicated field, the columns it is computed from, an accountable owner and a certification status of draft, verified or deprecated.

The trouble starts with the word itself. The [US Bureau of Labor Statistics JOLTS definitions (BLS, 2026)](https://www.bls.gov/jlt/jltdef.htm) split separations into quits, layoffs and discharges, and other separations. "Attrition" with no stated population can count any mix of the three.

### Score One People Question Before an Agent Answers

Give it one people question, such as voluntary attrition in Engineering in Q3, and what you can produce on its definition, its source systems, its access rules and its method. It scores five dimensions, gates on the weakest one and returns the work list. [Read the skill](/skills/ai-ready-question-score.md).

*Paste into a new chat*

```
Use the skill at https://atlan.com/skills/ai-ready-question-score.md to score whether one people analytics question is ready for an AI agent to answer. Ask me for whatever it needs.
```

*Run once in a terminal*

```
curl -fsSL --create-dirs \
  -o ~/.agents/skills/ai-ready-question-score/SKILL.md \
  https://atlan.com/skills/ai-ready-question-score.md
```

*For an agent*

```
curl -fsSL https://atlan.com/skills/ai-ready-question-score.md
```


    Human
    Agent


  ClaudeChatGPTGeminiCursorOther
  Copy

---

Atlan's [context layer](https://atlan.com/know/context-layer-for-ai-agents/), the shared set of definitions and rules that people and agents both read, puts each people metric under a named owner, so the definition of voluntary attrition belongs to a person instead of a wiki page. Inside Atlan, a purpose tied to a sensitivity tag can deny queries on pay and health columns or mask their values as `***`. A deny from a purpose wins over every persona grant, according to [Atlan's documentation on purposes (Atlan, 2026)](https://docs.atlan.com/product/capabilities/governance/access-control/concepts/what-are-purposes). Before an agent answers a people question, you can [score that one question on five dimensions](https://atlan.com/know/ai-agent/ai-ready-data/ai-ready-data-for-talk-to-data/) with the ai-ready-question-score skill and fix the weakest one first.

- Headcount by team as of a past quarter end that does not shift when someone changes teams later.
- A manager drill-down that withholds any result describing fewer people than the agreed group size.
- One voluntary attrition definition read by both the quarterly dashboard and the AI agent.
- A people question with no certified metric, labelled as uncertified and routed to a steward.

| Field | Detail |
|---|---|
| Persona | HR data leader, analytics leader |
| Key regulations | Vary by country and employer; set access and group-size rules with your privacy or legal reviewer |
| Primary stakeholders | HR analytics lead, HRIS owner, privacy or legal reviewer, metric owner |
| Typical data challenges | One word for several populations, history that changes, drill-downs that expose individuals |
| Maturity read | No published benchmark; score one people question at a time, and the lowest dimension decides |
| What a governed people metric holds | Definition, as-of rule, access rule, minimum group size, owner |

---

## Why does people data need its own semantic layer?

People data puts three demands on a semantic layer: a named population behind each word, history that holds for a past date, and privacy for anyone in a small group.

Most of what you know about a [semantic layer for analytics](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-for-analytics/) still applies. What changes is the cost of a mistake. Each of these failures can expose a person or a pay decision.

| People task | Where it breaks | What the layer adds | Source |
|---|---|---|---|
| Quarterly attrition report | Population: the word can count quits, layoffs and discharges, or other separations | A named population on each metric | BLS JOLTS definitions, 2026 |
| Headcount as of quarter end | History: teams and managers change after the date | Effective-dated rows with a validity window | dbt Developer Hub, 2026 |
| Manager drill-down | Privacy: row rules scope rows, yet a summary over a tiny group still points at a person | A minimum group size beside the row rule | Microsoft Learn (Fabric, 2026; Viva Insights, 2025) |
| Agent question on pay | Principal: the right answer depends on who is asking, a person or an agent | An access rule that follows the asset tag | Atlan docs, 2026 |

According to the [BLS JOLTS definitions (BLS, 2026)](https://www.bls.gov/jlt/jltdef.htm), quits are "Employees who left voluntarily," while layoffs and discharges are "Involuntary separations initiated by the employer." According to [Microsoft Learn's row-level security guidance for Power BI (Microsoft, 2026)](https://learn.microsoft.com/en-us/fabric/security/service-admin-row-level-security), a user who can see a row can see all of its columns, so hiding columns takes a separate object-level rule.

### Where a metrics layer stops

A [metrics layer](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-vs-metrics-layer/) standardizes a formula; for people data, the same metric also has to record whom it counts and who may read the result. Teams deciding [whether you need a semantic layer](https://atlan.com/know/ai-agent/semantic-layer/do-you-need-a-semantic-layer/) often start from [KPI definitions](https://atlan.com/know/ai-agent/how-to-manage-kpi-definitions-and-a-business-glossary-in-a-data-catalog/) in a spreadsheet or a [business glossary](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-vs-business-glossary/). People metrics need the glossary entry plus a date rule and an access rule.

The first people question that goes wrong is usually "how many left?", because the word has no single meaning. Write the population, the date rule and the access rule onto the metric, and a dashboard and an agent read the same thing.

---

## What does a governed attrition metric look like?

A governed attrition metric is one object that holds what it counts, whom it counts, how it treats time, who owns it and who may see it.

Picture finance and HR both reporting "attrition" each quarter. One counts every separation from the payroll; the other counts only people who chose to leave.

Both numbers are correct for their own population. Neither report says which one it used. The fix is one metric card per population, each with its own name, so every report resolves to exactly one of them.

### An example voluntary attrition card

Here is an example card that follows the JOLTS split; your company may classify retirements differently, and the card is where that choice gets written down.

| Field | Value |
|---|---|
| Name | Voluntary attrition |
| Definition | Employees who left voluntarily during the period |
| Counts | Quits |
| Does not count | Layoffs and discharges; other separations, such as retirements |
| Population | Employees on the payroll during the period (company-defined, shown as an example) |
| Grain | Employee, aggregated to department and month |
| Time rule | Department as of the separation date |
| Denominator | Average headcount for the period, here (start + end) / 2 (company-defined, shown as an example) |
| Example result | For example, 12 quits / average headcount of 240 = 5.0% for the quarter, not annualized (illustrative numbers) |
| Owner | A named HR analytics lead |
| Certified by and review date | A named certifier; reviewed each quarter |
| Access rule | Pay and health columns masked for readers outside the approved purpose |
| Minimum group size | Results withheld for groups below the agreed threshold |

The "counts" and "does not count" rows follow the [BLS separation categories (BLS, 2026)](https://www.bls.gov/jlt/jltdef.htm). The denominator is a company choice, so write it on the card. The grain row follows the [dbt Developer Hub documentation on entities (dbt Developer Hub, 2026)](https://docs.getdbt.com/docs/build/entities), where entities let you "aggregate a metric to the granularity of that entity."

### How Atlan records the card

In Atlan, a metric is a stricter tier of glossary term: it carries the owner and certification status every term has, plus a formula field and links to the exact columns behind it. [Atlan's guidance on structuring glossaries and metrics (Atlan, 2026)](https://docs.atlan.com/product/capabilities/governance/glossary/best-practices/structure-glossary-and-metrics) sets "One term per distinct formula" and a quarterly review cadence for critical metrics and KPIs, which is where the card's review date comes from.

The card sits on top of the [data model](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-vs-data-model-how-do-they-differ/) that stores employee rows, and does for one number what an [ontology](https://atlan.com/know/ontology-vs-semantic-layer/), a shared model of business entities and how they relate, does for those entities. Whether your team calls it a metric or a [semantic model](https://atlan.com/know/ai-agent/ontology/is-an-ontology-just-another-word-for-a-semantic-model/), the test is whether named [owners and stewards](https://atlan.com/know/ai-agent/ai-ready-data/ai-ready-data-owners-and-stewards/) stand behind it. Without a population and an owner on the card, nobody can settle which attrition number is right.

---

## How do you report headcount as of a past date?

Reporting "headcount as of March 31" needs employee rows that carry the dates they were true, and a query that names March 31.

Say an engineer moves from Engineering to Sales on April 15. If the org table is overwritten in place, a Q1 headcount by team run in May counts that engineer in Sales.

According to the [dbt Developer Hub documentation on dimensions (dbt Developer Hub, 2026)](https://docs.getdbt.com/docs/build/dimensions), slowly changing dimension (Type II) tables "typically have two time columns that indicate the validity period of a dimension": `valid_from` and `valid_to`. With an Engineering row valid until April 14 and a Sales row valid from April 15, a query for March 31 reads Engineering whenever it runs. Decide whether `valid_to` holds the last day in the old team or the first day in the new one, and write that on the card.

### Name the reporting date

The reporting date needs the same care. The same dbt documentation says that for a coarser window such as a month, MetricFlow by default uses the group valid at the start of that window. By that default, a quarterly report reads the team as it stood on the first day of the quarter, so a quarter-end headcount has to name its date. In dbt MetricFlow, time-based metrics and dimensions also need a time spine, which is joined for slowly changing dimensions, according to the [dbt Developer Hub documentation on the MetricFlow time spine (dbt Developer Hub, 2026)](https://docs.getdbt.com/docs/build/metricflow-time-spine).

### When a backdated correction lands

Effective dates answer what was true on March 31. A termination entered in May with a March effective date still changes that answer, so if a closed quarter must never move, keep a snapshot of the published figure as well. Whether you are [building a semantic layer](https://atlan.com/know/regovern-building-semantic-layer-guide/) on a commercial tool or an [open-source semantic layer](https://atlan.com/know/ai-agent/semantic-layer/open-source-semantic-layer/), the validity window plus a named reporting date is what makes HR history reproducible. A table that is [analytics-ready](https://atlan.com/know/ai-agent/ai-ready-data/ai-ready-data-vs-analytics-ready-data/) but holds only the current org chart cannot answer an "as of" question, for a person or for an agent.

---

## How do you keep a drill-down from identifying one employee?

Row rules, column masks and a minimum group size each decide a different thing, so a people metric has to declare all three.

A manager filters attrition to a team of two and sees who left, and every row they saw was one they were allowed to see.

| Layer | What it decides | Documented by |
|---|---|---|
| Row access policy | Whether a given row in a table or view can be viewed | Snowflake docs |
| Dynamic data masking | Which column values are masked at query time | Snowflake docs |
| Row-level security | Which rows users with Viewer permissions see; column hiding is set separately with object-level security | Microsoft Learn (Fabric), 2026 |
| User-attribute filter | User-specific restrictions on an Explore, through `access_filter` | Google Cloud Looker docs (undated) |
| Minimum group size | The smallest group an insight may describe; in Viva Insights, at least five | Microsoft Learn (Viva Insights), 2025 |
| Tag-based purpose | Access inside Atlan scoped by asset tag, with values masked as `***` | Atlan docs, 2026 |

In [Snowflake's documentation on row access policies (Snowflake, 2026)](https://docs.snowflake.com/en/user-guide/security-row-intro), a row access policy "determines whether a given row in a table or view can be viewed," and [Snowflake's column-level security documentation (Snowflake, 2026)](https://docs.snowflake.com/en/user-guide/security-column-intro) describes masking column values at query time. The [Looker `access_filter` reference (Google Cloud, undated)](https://docs.cloud.google.com/looker/docs/reference/param-explore-access-filter) applies user-specific data restrictions per Explore.

### Where a minimum group size fits

None of those settings is about group size. [Microsoft Learn's Viva Insights privacy settings (Microsoft, 2025)](https://learn.microsoft.com/en-us/viva/insights/advanced/setup-maint/privacy-settings) define the minimum group size as "the minimum aggregation threshold for insights," and add: "You'll need to set this number to at least five." That floor is Microsoft's rule for Viva Insights; agree your own threshold with your privacy or legal reviewer.

A threshold stops a direct small-group result, yet other routes to one person stay open. If Engineering shows six people and "Engineering without contractors" shows five, both pass a threshold of five, and the difference describes one person. Apply the threshold to the count being reported, attrition numerators included, and limit overlapping filters on sensitive people metrics.

### Where enforcement happens

[Data privacy for AI agents](https://atlan.com/know/data-privacy-for-ai-agents/) adds one more reader to the same problem, and handling [personal data](https://atlan.com/know/ai-agent/data-for-ai/how-to-handle-pii-in-ai-pipelines/) covers pay and health fields. [Role-based access](https://atlan.com/know/ai-agent/context-layer/context-layer-role-based-access-control/) decides who gets in; [agent access control](https://atlan.com/know/ai-agent-access-control/) decides what an agent reaches under its own identity.

Each of these controls is enforced where its docs place it: in the warehouse, the BI tool or the analytics product. The semantic layer records which controls a people metric depends on, so [policy owners](https://atlan.com/know/ai-agent/context-layer-for-data-governance-teams/) write the requirement once and the team building the next dashboard or agent reads it before launch instead of rediscovering it.

---

## What changes when an AI agent asks the people question?

An AI agent needs more than a dashboard number: the definition, the grain, where it may and may not look, and a verified query.

Zach Michel, Atlan, said of analytics agents: "It actually needs the definition. It needs things like the grain, and it also needs to know where it can and can't look from a SQL sense."

Austin Kronz, Atlan, set the order an agent should follow: "you should always start with certified metrics where you have them." When it falls back to other data, the agent tells the user "Here's which route I took."

Joe DosSantos, Workday, speaking about analytics agents in general, described an agent that says "This is not a verified query" and a user whose agreement goes back to a governance framework that decides the go-forward definition. For people data, the steward makes that call, not the user. "Regrettable attrition" rests on a company judgment about which leavers it wanted to keep, so an agent asked for it without a certified metric should not compute an answer at all: it should say no certified definition exists and route the question to the HR analytics owner.

### What the agent reads

The architecture of [HR agents and their compliance exposure](https://atlan.com/know/ai-agent/ai-agents-for-hr/) is a wider topic; the semantic-layer question is what the agent reads. Over [MCP](https://atlan.com/know/ai-agent/semantic-layer/mcp-semantic-layer-for-ai-agents/), the agent should read the certified definition and its linked columns first, the same card that serves [dashboards and agents](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-for-bi-vs-ai-agents/) alike, instead of composing [text-to-SQL](https://atlan.com/know/ai-agent/semantic-layer/text-to-sql-vs-semantic-layer-for-self-serve-analytics/) against raw HR tables from scratch. The number itself still comes from the warehouse or metric layer that computes it.

[Enterprise text-to-SQL](https://atlan.com/know/ai-agent/data-for-ai/text-to-sql-for-enterprise/) on people data raises the cost of a wrong guess, because the guess can disclose. [Context engineering](https://atlan.com/know/ai-agent/semantic-layer/context-engineering-vs-semantic-layer/) for an HR agent comes down to certified metrics first, the route always stated, and judgment calls left to a steward.

---

## How do you score one people question before an agent answers it?

You score one named people question on five dimensions, and the lowest score is the result. The ai-ready-question-score skill scores meaning, provenance (whether a figure traces back to its source systems), boundaries, reproducibility and procedure from 0 to 2: 0 means no artifact exists, 1 means it exists but is incomplete, unowned or out of date, and 2 means it exists, is owned and can be produced today.

Here is one HR question run through it: "What was voluntary attrition in Engineering in Q3?" The scores are an illustrative Atlan example, not customer data.

| Dimension | Example score | Why it is not a 2 yet |
|---|---|---|
| Meaning | 1 | The definition sits on an HR wiki page with no named certifier or review date |
| Provenance | 1 | The trace from report to HRIS extract is known to one analyst and not written down |
| Boundaries | 0 | The agent connects through a shared service account that can read pay columns, and the question has never been run as two users with different access |
| Reproducibility | 1 | Two runs have matched; the skill asks for ten runs with one distinct answer |
| Procedure | 0 | The method lives in one person's head |
| **Gated score** | **0** | Mean 0.6 of 2 |

The two numbers lead to opposite decisions. A mean of 0.6 reads as "partly ready," while the gated score of 0 says not ready, because the skill takes the weakest dimension instead of averaging. With boundaries at 0, an agent on that service account is a potential disclosure rather than a wrong number.

### The two-principal test

The boundaries test is concrete: run the identical question as two principals, meaning two users or identities with different entitlements, and confirm the results differ correctly, with an audit entry naming the agent under its own identity. The skill orders the work: boundary failures come before anything reaches production data, and meaning failures come first otherwise. Here that means a separate agent identity and that two-principal test, then a certified definition.

What makes data [AI-ready](https://atlan.com/know/ai-agent/data-for-ai/what-makes-data-ai-ready/) is decided per question, so a [readiness checklist](https://atlan.com/know/ai-agent/data-for-ai/ai-ready-data-checklist/) and the steps to [prepare data for natural-language analytics](https://atlan.com/know/ai-readiness/how-to-prepare-data-for-natural-language-analytics/) apply one question at a time. A question that fails the gate tells you which artifact to build next.

---

## What do your HR tools already do, and what does a shared layer add?

Your HR system, people analytics tool, warehouse and BI tool can each hold a definition and enforce their own controls, such as row rules, masking and user filters, for their own readers. A shared layer earns its place once a second tool or an agent reads the same metric.

At that point it adds one definition of voluntary attrition that every tool reads, and one place where the access rule and the minimum group size are written down for each tool to apply.

The number of readers a people metric has usually settles the choice between a [point solution](https://atlan.com/know/ai-agent/semantic-layer/point-solution-vs-context-layer-for-ai-data/) per tool and one shared layer, whether to [build or buy](https://atlan.com/know/ai-agent/semantic-layer/build-vs-buy-semantic-layer/), and which [semantic layer tools](https://atlan.com/know/best-semantic-layer-tools/) to shortlist. A definition kept inside one tool tends to get copied into the next one and drift.

---

## How Atlan handles people metrics

Atlan's AI Context Platform holds the metric card, the access rule and the record of who approved what in one place that dashboards and agents both read. Four capabilities cover the definition and access side of a people metric:

- **Certified metric terms.** A people metric is a glossary term with an owner, a certification status, a formula field and column-level links, with a quarterly review cadence for critical metrics ([Atlan, 2026](https://docs.atlan.com/product/capabilities/governance/glossary/best-practices/structure-glossary-and-metrics)).
- **Policy that follows the tag.** Inside Atlan, a purpose scopes access by asset tag; its denies override persona grants, and it can mask pay or health values as `***` ([Atlan, 2026](https://docs.atlan.com/product/capabilities/governance/access-control/concepts/what-are-purposes)).
- **One read path for agents.** Atlan MCP is a hosted server on the open Model Context Protocol that lets agents search enterprise context, trace lineage (where data comes from and where it flows), read governed definitions and glossaries, and run SQL; every action respects the permissions already set in Atlan ([Atlan, 2026](https://docs.atlan.com/product/capabilities/atlan-ai/how-tos/atlan-mcp-overview)).
- **A score before access.** The ai-ready-question-score skill names the blocking dimension for one question before an agent is pointed at it.

Teams that [implement an enterprise context layer](https://atlan.com/know/how-to/implement-enterprise-context-layer-for-ai/) usually start with the metrics analysts argue over most. For HR, that is attrition and headcount, and the same certified terms then serve [context engineering for an AI analyst](https://atlan.com/know/context-engineering-for-ai-analyst/) and the [data analytics teams](https://atlan.com/know/ai-agent/context-layer-for-data-analytics-teams/) who already report them.

---

## How do you start with one people metric?

You start with one people question, write its metric card, and score it before any agent gets access. Five steps, lowest effort first:

1. **Pick one question.** Choose the people question a senior person chases by email every quarter, such as voluntary attrition by department.
2. **Write the metric card.** Fill every field, including what the metric does not count and the denominator. Name one owner, not a team alias, and set a review date.
3. **Add the as-of rule.** For any headcount-based number, store team and manager as effective-dated rows with valid_from and valid_to, and name the date the metric reads.
4. **Set access and group size.** With your privacy reviewer, set the access rule for pay and health columns and the minimum group size for drill-downs, numerators included.
5. **Score, then decide.** Run the five-dimension score. Keep the agent off the question while any dimension scores 0, and treat every 1 as the next item on the work list.

Four pitfalls show up early with people data:

- **One "attrition" for every audience.** Instead, publish one metric per population, each with its own name.
- **Masking pay columns but leaving small-group counts open.** Instead, declare the mask and the minimum group size together.
- **Modelling only the current org chart.** Instead, keep effective-dated rows so past quarters stay fixed.
- **Letting an agent answer an uncertified people question without saying so.** Instead, label the answer and route it to a steward.

Getting one metric through these steps also gives you a first consistency check toward [measuring what agent governance returns](https://atlan.com/know/ai-agent/roi-of-ai-agent-governance/): one people question answered the same way by every reader.

---

## One governed people metric before a hundred dashboards

The first people metric you govern becomes the template your team copies.

If voluntary attrition ships with a population, an as-of rule, an access rule and a minimum group size, the team has a pattern for headcount and time to hire. If it ships as a number on a dashboard, the next agent inherits the argument.

Before writing the card, it is worth asking a plain question: do your HR team and your data team currently agree on what "left the company" means?

---

## FAQs about semantic layer for HR and people analytics

Seven questions HR and analytics leads ask before they govern a people metric.

### 1. Why do two HR reports give different headcount numbers?

Two HR reports usually differ because they count different populations or read different dates. One may include contractors or people on leave, and the other may not. One may read the org chart as it stands today, the other as it stood at quarter end. Fix it with one governed headcount definition that names its population and states its as-of date rule.

### 2. What is a semantic layer for HR data?

A semantic layer for HR data is a shared, governed definition of each people metric that every dashboard and AI agent reads. Each metric carries four parts: a definition with its population, an as-of rule for changing history, an access rule, and a minimum group size. Atlan's AI Context Platform records the definition as a certified metric with a named owner and a formula field.

### 3. What does "attrition" include?

Attrition includes whatever separations your metric says it counts. The US Bureau of Labor Statistics splits separations into quits, layoffs and discharges, and other separations such as retirements. Quits are the voluntary category in that split, and your company decides whether retirements also count as voluntary. Write the choice on the metric, along with the denominator, so every report that says attrition means the same population.

### 4. How do you report headcount as of a past date?

Store employee attributes as effective-dated rows, each with a valid_from and a valid_to date, so a team change creates a new row instead of overwriting the old one. Query the rows valid on the reporting date, and name that date, because a quarterly grain can default to the first day of the quarter. Keep a snapshot as well if a closed quarter must never change after a backdated correction.

### 5. How do you stop a drill-down from identifying one employee?

Combine row rules, column masking and a minimum group size. Row rules decide which rows a user may see, but a summary over a team of two can still reveal one person. A minimum group size withholds results for small groups; Microsoft Viva Insights requires at least five. Apply it to the count being reported and limit overlapping filters, since two groups that each pass can differ by one person.

### 6. What should an AI agent do when a people question has no certified metric?

The agent should say the answer is not from a certified metric, say which route it took, and never present the number as official. For a judgment-based metric such as regrettable attrition, it should not compute an answer at all and should route the question to the HR analytics owner. A user agreeing with an answer is a signal for the steward, who decides whether it becomes the certified definition.

### 7. Does a semantic layer replace the HR analytics tool you already use?

No. Your HR system, people analytics tool, warehouse and BI tool keep their jobs and their own access controls. The semantic layer holds the definition they all read and records the access rule and minimum group size each of them has to apply. The team building the next dashboard or AI agent then starts from that definition instead of rebuilding it.

---

## Sources

1. [Job Openings and Labor Turnover Survey definitions, US Bureau of Labor Statistics](https://www.bls.gov/jlt/jltdef.htm)
2. [Dimensions, dbt Developer Hub](https://docs.getdbt.com/docs/build/dimensions)
3. [Entities, dbt Developer Hub](https://docs.getdbt.com/docs/build/entities)
4. [MetricFlow time spine, dbt Developer Hub](https://docs.getdbt.com/docs/build/metricflow-time-spine)
5. [Row-level security (RLS) with Power BI, Microsoft Learn](https://learn.microsoft.com/en-us/fabric/security/service-admin-row-level-security)
6. [Customize Viva Insights privacy settings, Microsoft Learn](https://learn.microsoft.com/en-us/viva/insights/advanced/setup-maint/privacy-settings)
7. [Understanding row access policies, Snowflake Documentation](https://docs.snowflake.com/en/user-guide/security-row-intro)
8. [Understanding column-level security, Snowflake Documentation](https://docs.snowflake.com/en/user-guide/security-column-intro)
9. [access_filter, Google Cloud Looker Documentation](https://docs.cloud.google.com/looker/docs/reference/param-explore-access-filter)
10. [Atlan MCP, Atlan Documentation](https://docs.atlan.com/product/capabilities/atlan-ai/how-tos/atlan-mcp-overview)
11. [Purposes, Atlan Documentation](https://docs.atlan.com/product/capabilities/governance/access-control/concepts/what-are-purposes)
12. [Structure your glossary, taxonomy, and metrics catalog, Atlan Documentation](https://docs.atlan.com/product/capabilities/governance/glossary/best-practices/structure-glossary-and-metrics)
13. [Score one question's AI readiness (ai-ready-question-score), Atlan](https://atlan.com/skills/ai-ready-question-score.md)

## Next steps

- To see Atlan's context layer in action first, watch a recorded demo: [Atlan context layer demos](https://atlan.com/context-layer-demo/)
- To see Atlan working on your own data stack, book a demo: [Book a demo with Atlan](https://atlan.com/forms/talk-to-sales-contact/)
