Try it yourself
Five claims. Verify each one with your own hands.
Everything below runs against the live serving layer for synthetic GlobalTech. Same RPCs, same policies, same 15 tools your own AI client would use.
1Governed AI access
Connect your AI
Point ChatGPT, Claude, Copilot, Cursor or any MCP client at the same governed tools this site uses. The model gets tools, not tables.
Endpoint
https://peopleanalyticsai.net/api/mcpPublic demo token
Maps to demo-external-viewer · aggregate only · min_cell 50 · 15 read-only tools.
people-mcp-demo-visitor-agg-min50-v1Client support · 2026-09
- CursorVerified
- ClaudePartial · see note
- ChatGPTNot available today
- Microsoft CopilotPartial · see note
- Other MCP clientsSupported
Where a client cannot send a static Bearer header we say so instead of pretending.
2Identity-aware access
Test permissions
Permissions in the AI era are data policy, not prompt wording. Ask the same question as two identities and compare what each one is allowed to see.
Same question
“What is the Sales compa-ratio?”
Confidential metric outside the bound org. Leader and HRBP are denied by org_scope; the visitor by sensitivity.
site visitor
org scope company-wide · identity min_cell 50 · confidential denied (sensitivity)
Pick a question and run it.
Policy matrix for this identity
| Scope | Metric | Policy |
|---|---|---|
| own org | headcount | allow |
| own org | compa_ratio_median | deny · sensitivity |
| other org | headcount | allow |
| other org | compa_ratio_median | deny · sensitivity |
| company level | headcount | allow |
| company level | compa_ratio_median | deny · sensitivity |
Engineering leader
org scope Engineering · identity min_cell 20 · company-level confidential denied
Pick a question and run it.
Policy matrix for this identity
| Scope | Metric | Policy |
|---|---|---|
| own org | headcount | allow |
| own org | compa_ratio_median | allow |
| other org | headcount | allow |
| other org | compa_ratio_median | deny · org_scope |
| company level | headcount | allow |
| company level | compa_ratio_median | deny · company_scope_confidential |
The prompt is identical. The difference comes from people_policy_demo_identity and FORCE RLS in the serving database, not from anything the model was told. Person-level data is never returned to any identity.
3Built-in data trust
Trace a number
Start from certified Headcount and open every layer that has to hold before it is allowed on a page. Then watch the same guardrails stop a bad feed.
Certified Headcount
49,823
as of 2026-08-31 · current_certified pointer · data-v1
One number. Open each layer to see what has to be true before it is allowed on a page.
- Pointer
- current_certified
- As of
- 2026-08-31
- Serving run
- data-v1
- Simulator SHA
- b28c00833782
The website reads whatever this pointer certifies. The incident replay lives behind a different pointer and never overwrites it.
4Explicit business logic
Govern a metric
A metric is a governed analytical asset, not a SQL snippet. Read the first layer, then open lineage and impact to see what a rule change would touch.
Governed analytical asset
Headcount
- Metric id
- headcount
- Business definition
- Certified workers as of month-end. Employment type and status follow BR-WF certified rules.
- Formula
- count of certified worker-month rows as-of
- Numerator
- count(is_certified) as-of month_end
- Denominator
- — (count metric)
- Inclusion / population
- — (not in data-v1 registry)
- Exclusion
- — (not in data-v1 registry)
- Grain table
- people_snap_worker_month
- Window
- month (as-of)
- Annualized
- no
- Effective date
- — (not in data-v1 registry)
- Version
- v1
- Owner
- People Analytics
- Status
- certified
- Sensitivity
- internal
- Metric min cell
- 5
- Health
- healthy
Approver, RACI roles, inclusion / exclusion rules and effective date are not stored in the data-v1 registry; they are marked as missing rather than filled in. Nothing on this card is typed by hand; it is read from people_get_metric_definition and people_metric.
5Reusable analytical intelligence
Run an analysis
Case 3 through the live agent: certified metric, breakdown, evidence, hypotheses, tool calls, next questions and the trace. Observed evidence and AI hypotheses never mix.
1 · Certified metric
Engineering voluntary attrition
16.0%
Engineering · trailing-12m (annualized) · as of 2026-08-31 · company 14.0%
Question
Reusable analytical skills
Target architectureWhere this goes next: methods that are versioned, tested and certified so the agent can reuse them. None of these are live on the public site today. Live today: 15 governed tools and the deterministic case skeleton you just ran.
Significance testing
Is the APAC-SIN gap larger than noise for cells this small?
Driver analysis
Which observed signals move with attrition after controlling for tenure and grade?
Cohort analysis
Do the 2025Q4 reorg cohorts separate from their peers over time?
Segmentation
Automated slicing across location × tenure × grade with suppression built in.
Causal inference
Design a quasi-experiment before anyone calls compa-ratio lag a cause.
Survey sentiment / NLP
Engagement free text is in the ontology but not in any public output.