Copilot / AI insights
Path: HR Analytics Overview → AI Digest ✨ / HR Insights Q&A; HR Analytics Setup → Suggest Setup with Copilot ✨ / Recommend Model; HR Analytics Charts → Suggest Charts with Copilot ✨
HR Analytics includes optional AI features that read your current indicators and explain them in natural language, plus two configuration assistants that turn a plain-language request into proposed analytics setup or ready-to-use charts. The insight features are a read-time lens — they never change any data, and they only ever see indicator results, never raw payroll ledgers or individual personnel records. The Setup and Chart copilots only ever propose: everything they create is disabled until you review and enable it.
AI Digest
From the Overview, AI Digest ✨ generates a short written summary of your current indicators:
- the three to five most important findings, most critical first (critical findings are marked 🔴);
- a brief overall assessment;
- what deserves action this week.
The summary is produced on demand (generate / regenerate / discard) in your language; you can Copy it or jump to the Overview. The same generator can prepend an AI summary to e-mail digests — see E-mail digests & alerts.
HR Insights Q&A
HR Insights Q&A lets you ask questions in plain language ("Which departments have the highest turnover?") and get an answer grounded only in your current indicators and the built-in catalog of enabled analytics. If the data does not contain the answer, the assistant says so and suggests which indicator to enable — it will not invent numbers, and it cites the indicator names behind every figure.
- Methodology questions ("How is turnover computed?", "Which analytic shows the reservation quota?") are answered from each indicator's built-in description — the same text you see in the Indicator Details FactBox.
- Follow-up questions keep context: the chat remembers the questions and answers of the current session (up to the last six turns), so you can ask "and what about last quarter?" without repeating yourself. Nothing is stored — closing the dialog forgets the conversation.
Explain this indicator / Explain anomaly
On the Overview, two per-row actions save you typing the question:
- Explain with Copilot ✨ (also on the Indicator Details FactBox) opens the Q&A chat with a preset question about the selected indicator — what it measures, its current value and trend, and whether the current severity needs action — and asks it immediately.
- Explain Anomaly ✨ is available on anomaly rows. Copilot compares the anomaly with the other indicators of the same period (±2 periods of series history) and formulates hypotheses about what could be related — clearly labeled as hypotheses, never as conclusions. If nothing correlates, it says so.
Setup Copilot — suggest analytics setup
From HR Analytics Setup → Suggest Setup with Copilot ✨ (also linked from the assisted setup guide), describe what you want to track in your own words — for example: "monthly turnover by department, a 6-month payroll fund forecast with a pessimistic scenario, reservation quota compliance". Copilot selects matching analytics types from the built-in catalog and proposes concrete instances:
- Propose — each proposal row shows the analytics type, a suggested code and name, grouping, key parameters, and the copilot's rationale. Every row is validated by the same rules as manual setup; rows that fail show the validation message and cannot be created.
- Review — untick Accept on anything you do not want. Nothing is created until you confirm.
- Create selected — accepted, valid proposals are created disabled and the setup list opens filtered to them. You review, adjust if needed, and enable — nothing calculates on its own.
The copilot never proposes person-level analytics (for example, the Bradford Factor): enabling those always remains a deliberate human decision.
Chart Copilot — suggest charts
From HR Analytics Charts → Suggest Charts with Copilot ✨ (also available on the HR Analytics Setup list, next to Suggest Setup with Copilot), describe the charts you want to see in your own words — for example: "turnover trend with a forecast, sick days by department, the age structure". Copilot proposes ready-to-use chart definitions over the indicators that are already calculated:
- Propose — each row shows the chart type, the source indicator, the group filter, an optional forecast overlay, the audience and the copilot's rationale. The chart type, layout, roll-up, forecast overlay and audience are decided by a deterministic rules engine, not by the AI: the model only chooses among the chart types the rules allow for each indicator, and any choice outside that set is corrected automatically with a note. A request for a chart over an indicator that is not calculated yet becomes an error row that points you at Suggest Setup with Copilot.
- Review — untick Accept on anything you do not want. A proposal that would duplicate an existing chart's data view (same indicator, layout and group filter) is blocked and names the chart that already covers it.
- Create selected — accepted charts are created disabled and the chart list opens filtered to them. Preview each one, then enable it — nothing appears on a Role Center until you do.
A forecast overlay is added only when the indicator has a time-series view with enough monthly history (at least a year) and an enabled forecast indicator paired to it; otherwise the chart is still proposed, without the overlay and with a note explaining which condition was missing. Because the rules engine works with the AI switched off, the proposed charts match the built-in recommended charts even when no Copilot is configured — the AI only interprets your request and explains the choice.
Forecast Model Advisor
On the setup card of a forecast indicator, Recommend Model analyzes the source series — history length, share of zero periods, demand intermittency, yearly seasonality, trend — and recommends a forecast model (for example STL for strongly seasonal series, the ETS+ARIMA ensemble as a robust default, or automatic selection for sparse series). The recommendation works without AI configured; when Copilot is available, it adds a short explanation of the reasoning. Choosing Apply writes the model onto the card — you still save the record yourself. With fewer than 12 history points the advisor warns that forecast quality will be degraded.
What the AI can and cannot see (privacy)
The AI is grounded strictly through the same read API as the screens, so the privacy rules apply automatically:
- Only indicator results the current user is allowed to read are sent to the model.
- Person-level rows are never sent — even to users who are allowed to see them on screen. The AI operates on aggregates only.
- Salary indicators are excluded for users without payroll-ledger permission.
- Raw ledger and personnel tables are never sent.
- The Setup Copilot and the Q&A catalog block see only the catalog of analytics types (descriptions, parameters, defaults) and the list of configured instance codes — person-level types are excluded from the catalog entirely.
- The Chart Copilot sees only the shapes of your calculated indicators (how many periods and groups they have, their allowed chart types) and your existing chart definitions — no indicator values, and never a person-level or salary-gated indicator the current user may not read.
Each indicator line the model receives carries the line's own insight text, whether it is the total or one group's figure, and the unit of the total (count, amount, percent…), so the model does not mistake, for example, one department's payroll fund for a number of anomalies.
Figure check. The AI Digest, the digest e-mail summary and the Q&A answers are checked after the model replies: every figure in the text must appear in the indicator data it was given (as written, or rounded, or abbreviated to thousands or millions). Figures that cannot be matched are listed on a closing line — ⚠ Check these figures, they could not be matched to the indicator data: … — so you know which numbers to verify. Small numbers (up to 12), years and dates are not checked.
Every AI output carries the standard disclaimer to verify AI-generated content before acting.
Turning it on
The AI features appear only when the Copilot capability ("HR Analytics") is active and an authorization is available:
- Cloud (SaaS) — works through the Microsoft-managed Copilot capability; an administrator can switch it on or off under Copilot & AI capabilities.
- Bring your own key — configure an Azure OpenAI endpoint, model and API key in the Copilot Setup (works both on-premises and in the cloud for customers who prefer their own subscription).
If neither the capability nor an authorization is available, the AI actions are simply hidden and the pages work normally.
Customizing the prompt
The system prompt used to generate the AI digest and to answer HR Insights questions can be overridden by a consultant under Communication setup (from the Subscriptions page) → AI Prompts. Leave a prompt blank to use the shipped default. Keep the %1/%2 placeholders so the company name and language are still inserted. A single prompt serves every language, because the model is always instructed to answer in the recipient's language.