People Analytics for the AI era

AI-ready People Analytics starts with trusted facts, not a smarter chatbot.

Not an LLM on top of HR data. A governed architecture where AI reasons over certified metrics, explicit business rules, traceable data quality, identity-aware access and reusable analytical capabilities.

Explore the Architecture Try It Yourself

Synthetic GlobalTech — not a real company. No customer or employer data. Frozen data-v1 snapshot as of 2026-08-31.

Five differences

People Analytics was built for dashboards. AI requires a different architecture.

Five places where a dashboard-era stack breaks the moment an AI starts asking it questions. Left: how most People Analytics teams still run. Right: what this site runs today; the one row that is still roadmap says so.

Dimension

Traditional People Analytics

Dashboards first, AI bolted on top

AI-Ready People Analytics

What this site runs today

1Layer L4

Governed AI Access

AI reasons over governed facts instead of inventing the facts.

Try it yourself Why is Engineering attrition increasing?

Traditional People Analytics

The LLM reads Excel, report extracts or raw tables
It may write its own SQL or do its own arithmetic
Nobody can say how the answer was computed

AI-Ready People Analytics

AI never touches raw data; it calls certified tools
Numbers come from SQL and deterministic templates
The model only selects tools, interprets, and writes hypotheses

2Layer L1

Built-In Data Trust

Data quality is monitored before bad data becomes a business story.

Try it yourself Can I trust this number?Why did Headcount suddenly drop?

Traditional People Analytics

Data problems surface in the dashboard, found by end users
Triage walks dashboard → report → SQL → ETL → source system
Locating one bad number takes hours or days

AI-Ready People Analytics

Contracts, mapping, pipeline, tests, lineage and snapshots are monitored
A failure is pinned to the layer that produced it
Failed checks become incidents, never a silent dashboard change

3Layer L2

Explicit Business Logic

Business rules become governed, versioned assets — not hidden logic in SQL or dashboards.

Try it yourself Can I trust this number?

Traditional People Analytics

Definitions live in analyst heads, SQL, dashboards and spreadsheets
The same metric has several versions across teams
Reporting conflicts are settled by argument

AI-Ready People Analytics

Each metric is a governed asset: business definition, formula, numerator, denominator, grain, window
Owner, version, status, sensitivity and min-cell rule travel with it
Upstream and downstream dependencies are visible before a rule changes

4Governance pillar

Identity-Aware Access

The agent inherits the user's permissions. It does not become a new authorization layer.

Try it yourself Can I trust this number?Why is Engineering attrition increasing?

Traditional People Analytics

Permissions are maintained separately in HRIS, ATS, warehouse and BI
Each layer can drift from the others
A chatbot on top adds one more place to get it wrong

AI-Ready People Analytics

Policy is enforced in the governed data layer with FORCE RLS
The agent inherits the user's identity; it is not an authorization point
Four identities asking the same question get four different answers, each within policy

5Layer L3

Reusable Analytical Intelligence

Analytical methods become reusable capabilities, not one-off projects.

Try it yourself Why is Engineering attrition increasing?

Target architecture

Traditional People Analytics

Every business question becomes a new project
Pull, clean, analyse, test, interpret, make slides — again
Days or weeks per question

AI-Ready People Analytics

Common methods become versioned, tested, certified skills
The agent reuses them; analysts do not start from zero
Target: detect → analyse → explain → narrate → present

Live today: 15 governed tools and deterministic case skeletons. The skills registry is roadmap.

How it works

Five layers. Two pillars that cut across all of them.

Governance and observability are not a layer you add at the end. They run through the data foundation, the semantic layer, the skills, the agent runtime and every page.

Proof points

Three questions the architecture has to answer.

Each case exercises specific layers. Trust the number, separate a feed failure from a workforce change, then analyse without confusing correlation with cause.

Demo environment

Synthetic data. Real enterprise complexity.

A seeded business-event simulator emits Frappe HR and Greenhouse documents for five years, then the same pipeline, tests and metric registry an enterprise would run turn them into certified aggregates.

Synthetic GlobalTech — not a real company. No customer or employer data.

49,823

Certified Headcount

people_get_metric_for(headcount) · as of 2026-08-31

Certified metrics

people_v2.people_metric · status certified

Automated quality tests

people_v2.people_quality_test · bronze / silver / gold

Frozen

Certified snapshot

Frozen at 2026-08-31 (data-v1)

Workforce history: 5 yrs; 2021-09-01 → 2026-08-31, month-end grain, 59 roll-forward months.
Cumulative workers: ~80K; Worker spells across five years; ~50K active at period end; rehire ~2%.
Trailing-12m hires: 10,489; Certified external hires via T1 in the last 12 months (~874 per month).
Historical openings: 55,351; Requisitions over five years; cancel rate 9.8%; time-to-fill p50 32 days.
Training participation rows: 990K; Trailing-12m training_hours_per_worker 20.4 (calibration band 8–60).
Departments / teams: 25 / 150; Simulator baseline org shape; ~120 O*NET-derived designations.
Regions · countries · cities: 3 · 10 · 30; AMER, EMEA, APAC. Location is a certified breakdown dimension.
Max reporting depth: 8; Certified BFS from a single CEO root; 0 orphan managers; span mean 8.5; 11.8% managers.

DomainsWorkforceRecruitingCompensationPerformanceLearningSkillsEngagement

Window 2021-09-01 → 2026-08-31. Every number here is from committed data-v1 documentation or the serving RPCs; nothing is estimated for effect. Dataset details

The boundary

Where AI adds value, and where it is deliberately constrained.

The model's entire remit fits in one short list. Everything else is SQL, policy and tests.

Try It Yourself Explore the Architecture

Deeper technical detail and the builder story: Perspective, About and all case studies.

The model may

Select which governed tool to call
Interpret certified results
Propose hypotheses, labelled as hypotheses
Suggest the next question

The model never

Produces a number
Rewrites a fact sentence
Writes its own SQL
Becomes an authorization layer