Athena on Every Dashboard
Ask what, where, why and how on any workspace or module screen; the question inherits the screen's context and follow-ups narrow it.
AI-Powered Analytics: Ask Any Dashboard What, Where, Why and How
Type a question on any workspace and get a scoped answer, a forecast with its confidence, an anomaly with its records, or an explanation with its sources — from three AI models that are built to read, rank and narrate, and built not to write.
AI-Powered Analytics is what administrators experience when the platform's intelligence models sit on the dashboards: Athena answers in place, the ML Engine forecasts and flags, BR Insights explains, BR Spotlight ranks. Every answer is scoped to the person asking, logged with the model that produced it, and computed on the same records for an SAP SuccessFactors entity and an entity on Workday, UKG, ADP, Dayforce or Paycom.
Ask in place · what, where, why, howForecasts with confidenceAnomalies with recordsExplanations with sourcesSame answers on any HRIS
Analytics You Can Talk To — Without an AI That Makes Things Up
A dashboard answers the question it was built for. Benefits questions rarely arrive that neatly: what is the number, where is it worst, why did it move, what happens next, and what do I do about it. AI-Powered Analytics puts that conversation on every workspace and splits the work three ways — one model plans the read, one forecasts and flags, one explains — so that the useful part of AI shows up and the dangerous part, a model inventing a figure or changing a record, cannot.
AskA question on any workspace, in contextPlanAI Query → a scoped analytics planForecast & FlagML Engine: predict, detect, segmentExplain & ActAI Insight narrates; a draft waitsAny HRISSame answers for every entity
4questions every workspace answers in place — what, where, why and how — plus what happens next
3models with one job each: AI Query plans, the ML Engine predicts and detects, AI Insight explains
1confidence stated on every forecast and anomaly, so a prediction is never mistaken for a fact
0records a model can write — a campaign, case or correction is drafted in the owning module for a person to commit
Working Flow — BR Analytics, the Read-Only Operations Workspace
The shipped analytics app behind this page, in ten steps as it runs: four workspaces opened from home tiles, two Ask boxes that never share a catalog, and one real send. The last column is the boundary the platform holds at that step.
| # | Who · where | What happens | The rule at this step |
|---|---|---|---|
| 1 | Admin opens BR Analytics Home tiles | Four workspaces open from the home tiles — Enrollment analytics, Costs & contributions, Athena chatbot analytics, Document analytics | Read-only: nothing here changes an election, a payroll line or the chatbot's configuration |
| 2 | Opens Enrollment analytics Decisions view (default) | Decision tiles show what needs a human — blocked enrollments, exceptions — with the people or counts behind each tile one click away | A tile is a count of records that fail a check, computed by the enrollment service, not a model's opinion |
| 3 | Moves through the E-views Pace · KPIs · exceptions · funnel · coverage · cost · contribution | Charts share one period and one set of filters — Access Control, plan, organisation; the default period is last month | Eligibility is read as of today from the benefits eligibility record; the open-enrollment coverage window is a different date and is never substituted |
| 4 | Opens the employee ledger Ledger view | One row per employee in the filtered population, with the person overview a click away | Role-based permissions scope the population before a row is drawn |
| 5 | Asks the enrollment Ask box Enrollment Athena Admin | Types a question about the open-enrollment population on screen; the answer follows a fixed order — session, intent, playbook, allow-listed read — and fails rather than guesses | This Ask box answers only the open-enrollment population; it cannot answer chatbot-quality, one employee's private chat or document questions |
| 6 | Sends a direct notification Person overview → send | The one write in the app: a notification to the people behind a tile, with copy drafted by a commercially available language model from the enrollment facts already loaded, and a fixed fallback if every model fails | The copy may not invent a payroll deduction or a deadline; the facts come from the loaded context |
| 7 | Opens Costs & contributions Landing · Overview · Category | Employee cost and employer cost for the period, and the same split inside a benefit category | Employee share and employer share stay in separate series — never a blended rate presented as either side; the plan cost-variant code is never charted as money |
| 8 | Opens Athena chatbot analytics A-1 Overview → A-8 Recommendations | Eight views of the employee chatbot's health, on aggregates only, default period last month | The employee chatbot widget is hidden on these screens so the analytics bot and the employee bot are never on screen together |
| 9 | Asks the chatbot Ask box Chatbot analytics Ask | Same fixed order as step 5, on the chatbot catalog only | Aggregates only — the bot must not return a person, message text or feedback text |
| 10 | Opens Document analytics DMS analytics → one document | Documents by category and their movement through the document center, with drill-down to a single document | Computed by the Document Center's own actions — the analytics app draws them |
One Dashboard, Five Stages of a Conversation
Follow an administrator asking the Benefits Participation workspace about a rising waive rate — what it is, where, why, what next and what to do — and then the same conversation on an entity running Paycom.
Ask in place — the question inherits the screen.
Athena sits on every analytics workspace and every dashboard in the platform — Benefits Participation, ACA Intelligence, the Workflow Control Tower, Carrier Billing, the Athena Quality Console. The question inherits the Analytics Context the screen already holds: the evaluation date, the filters, the selected row. “Why is this high?” means the number the administrator is looking at, not a number the model guesses. Follow-ups narrow the same context; a new filter on the screen changes the next answer.
Where you can ask
- Every Essential Insights workspace and the Benefits Admin Dashboard
- The Workflow Control Tower, Carrier Billing, 4-Way Reconciliation and ACA Intelligence screens
- The Athena Quality Console — analytics about the employee chatbot itself
- The BR mobile app for administrators on the move
Why it matters
A question asked on the screen is a question about the screen — no copying figures into a separate chat.
Employees ask their own Athena; administrators' Athena is a separate, permission-scoped assistant.
AI Query plans the read — it never writes SQL and never answers alone.
A natural-language question becomes a structured analytics plan: which analytic, which metrics, which dimensions, which date range, which comparison. The plan is checked against the administrator's role-based permissions and the cohort minimum before the orchestrator runs it, and the database aggregates at source. The model never produces raw SQL, never sees raw rows and never composes the answer from its own knowledge — the figures in the reply come from the same KPI engine the dashboard tiles use, so a spoken answer and a tile can never disagree.
What a plan contains
- The analytic and metrics — resolved to the orchestrator's single definitions
- Dimensions and filters, merged with the screen's context
- Date range and comparison set — prior period, prior year, peer band
- A permission check and cohort check before anything runs
Why it matters
The model decides what to ask; the engine decides the answer.
A question the catalog cannot map — an undefined metric, a dimension the role may not see — returns “I can't compute that” with the reason, not an estimate.
The ML Engine predicts and detects — with a confidence on every result.
Forecasting, anomaly detection, segmentation and risk scoring are Python models inside the analytics orchestrator, not a language model. They forecast plan migration before Open Enrollment, premium change under a renewal, the benefits impact of a wave of life events; they flag an invoice that does not match its enrollment, a waive rate that turned, an ACA member drifting toward the 95% threshold, a sync feed that stopped; they segment a population by the factors that actually move its choices. Every output carries a confidence band and a link to the records it was trained and scored on.
What the models do
- Forecast — plan migration, premium change, life-event impact, OE participation
- Detect — billing variances, enrollment drop-off, offer-rate drift, feed health
- Segment and score — populations by the factors that move them; risk per item for Spotlight
- Benchmark — peer comparison where a peer band exists for the plan type
Why it matters
A prediction with a confidence band is a planning input; a prediction without one is a guess dressed up.
Boundaries are explicit: a model may predict a full-time transition or a threshold breach; it never sets legal status, a form line, a safe-harbour figure or a filing.
AI Insight explains; the owning module drafts — and a person commits.
When the answer is a number, AI Insight writes the sentence that goes with it: what it is, what moved it, who is behind it, what it compares to, what to do — every clause linked to its source. When the answer is an action, Athena asks the module that owns it to prepare a draft: a plan-comparison campaign in Communication HUB, a variance case in Carrier Billing, a delegation in Workflow, a re-send in HRIS Sync. The administrator reads the reasoning, opens the draft, changes it and commits it inside that module's own audit trail. Nothing changes because the AI said so.
What comes back
- An explanation with linked sources, in the shape BR Insights uses everywhere
- A draft in the owning module — never sent, never posted, until approved
- The model, prompt version and context behind the answer, logged
- A “what I could not do” note when a comparison or a permission is missing
Why it matters
AI proposes; a named person commits — on every answer, without exception.
Prompts, models and providers are configuration, inventoried and evaluated under the NIST AI Risk Management Framework in the BR compliance engine.
Ask the same question of every entity — SAP or not.
The models read the canonical data model, so an entity on SAP SuccessFactors and an entity on Paycom answer the same question the same way. Forecasts are trained on the same record families, anomalies use the same thresholds, explanations use the same clause shape, and a group-level administrator can ask across entities in one breath — “compare the waive-rate forecast for our SuccessFactors plants and our Paycom distribution centres” — with cohort suppression and role scoping applied per entity. Where the HRIS source is the reason for a result, the answer says so and links the HRIS Sync record.
How entities appear in an answer
- One question across entities, or one entity at a time — same models, same definitions
- HRIS source named only when it explains the result — a late feed, an unconfirmed field
- Forecasts and anomalies per entity, with the group view aggregating them
- Owners for drafted actions resolved from each HRIS identity through SSO
Why it matters
A multi-HRIS group asks one question and gets one kind of answer.
Quebec French and Spanish administrators ask and read in their own language; the models and the logging are the same.
Ground Rules of BR Analytics — What the Platform Holds
Analytics that sit next to the records have to be trustworthy about what they read, what they compute and what they will not do. These are the rules the app enforces.
| Rule | How it works | Why it is there |
|---|---|---|
| Read-only | Every workspace reads; the only write is the direct notification send, which is a message, not a change to an election or a payroll line | An analyst can explore freely without any risk to the records |
| Three catalogs, three doors | Enrollment Athena Admin answers the open-enrollment population; the chatbot analytics Ask answers aggregated chatbot KPIs; the employee Athena answers one signed-in employee — a prompt, table list or session is never shared between them | A population question cannot leak a person, and an employee question cannot see the population |
| Ask order, never skipped | Session first; then intent routing by the language model when the intent is not already known; then playbook getters, with an optional rewrite of the wording from those facts; then allow-listed text-to-SQL against a written table contract; then fail | The model rewords facts the platform computed — it does not compute them |
| Fail closed | If neither a playbook nor an allow-listed read can answer, the Ask box says so; it does not invent a KPI | A wrong number with confidence is worse than no number |
| Employee and employer money stay apart | Cost and contribution charts carry employee share and employer share as separate series; no blended “total rate” is ever presented as either side | Renewal and budgeting conversations start from figures that mean one thing |
| Eligibility is as of today | Eligibility comes from the benefits eligibility record as of today plus the Access Control; the open-enrollment coverage window and the shopping window are different dates | “Eligible” never silently means “was covered in the OE window” |
| A cost code is not a premium | The plan cost-variant code on an election is a code, not an amount; money comes from the configured rule, the elected amount and the payroll actual | Nothing on a chart is a premium unless it came from a money field |
| Unique users are counted, not summed | Unique users are a distinct count of people on conversations; a daily-metric sum is a different number, and slices by domain, intent or channel are never added together | The same conversation sits in more than one slice; adding them double-counts |
| Satisfaction uses current feedback | Only the current feedback row per interaction counts; older rows are history | A changed rating replaces, rather than adds to, the score |
| Forecasts and recommendations are estimates | A forecast is a projection stored as analytics output; a recommendation is an estimate on the screen — neither writes an intent, a prompt, a tool, a payroll deduction or an enrollment deadline | Planning inputs never become production changes by themselves |
| Maintenance is not change | Backfill and recalculate rebuild analytics aggregates only | Rebuilding a dashboard never touches the live chatbot or the records |
Core Capabilities of AI-Powered Analytics
Built so that the useful part of AI shows up on every benefits dashboard — and the dangerous part cannot.
AI Query · Analytics Plans
A question becomes a structured, permission-checked plan the orchestrator runs — never raw SQL, never a figure from the model's own memory.
ML Forecasting
Plan migration, premium change, life-event impact and OE participation forecast with confidence bands and the reasoning behind the expected value.
Anomaly & Risk Detection
Billing variances, enrollment drop-off, offer-rate drift, feed health and Athena quality clusters flagged with their records — and ranked by BR Spotlight.
Explanations with Sources
AI Insight writes the sentence for any number — what, why, who, compared to what, what next — every clause linked.
Governed, Logged, Bounded
Every call logged with model and prompt version; models inventoried under NIST AI RMF; explicit boundaries — no legal status, no filing figures, no writes.
BR Benefits Analytics — Design Standard v1.0
One Design Standard for Every Analytics Workspace
Every AI-powered analytics surface in Benefits Reimagined — Election Change, Enrollment, Participation, Demographics, ACA, ReVerge, Carrier Integration, Direct Billing, ISD, Communication and everything after them — is composed from one visual and interaction system. The interface's job is to get a person to a correct decision quickly, and to make the basis of that decision inspectable; everything else serves that or is removed. Seven principles govern it.
Work first, data second
Open on what needs a person, not on a wall of charts.
Rules out: Landing on a KPI grid and making the admin derive the task
One definition, everywhere
The tile, the chart, the export and Athena read one implementation.
Rules out: A number that differs between two places in the product
One context per workspace
Clicking a mark filters everything, including the assistant.
Rules out: Per-chart filters; a chart that disagrees with the table beside it
Show the basis
Metric ID, version, freshness, scope and lineage are always reachable.
Rules out: A figure you cannot trace
Deviation over magnitude
Where the question is “is this unusual”, rank by baseline shift.
Rules out: “Atlanta has the most changes” — Atlanta has the most people
Recommend, then let a person act
Python recommends with a rule version; the platform executes; the action is reversible and audited.
Rules out: Silent automation; irreversible bulk actions
Honest by construction
No colour-only meaning, no dual axes, no invented answers, no fake precision.
Rules out: Anything that reads well and misleads
Every Workspace Is Four Screens — and Opens on Decisions
A new workspace is a new metric catalogue and a new decision taxonomy, never a new visual language. It is composed from four fixed screens so administrators build muscle memory across the whole estate, and it opens on the one that needs a person.
The decision card: an instruction, its evidence, and undo.
The Decisions screen opens with one sentence of prose carrying certified metrics, then three buckets — Needs you now, Waiting on others, Cleared this week — and a flow chart of opened versus cleared per week. A queue that only ever shows what is broken reads as a product that never helps, so cleared work is first-class. Each card is one decision, not a pile of signals.
- The title is an instruction — “Post a payroll correction for 4 employees”, never “4 reconciliation mismatches”
- Why this is one decision, not four — generated from the data, naming the shared root cause
- Evidence rows — each affected person with plan, event, exposure and age; each row opens the full lineage
- Recommendation with a rule version and confidence — Python recommends, the platform executes, every action audited
- Reversibility is a design requirement — approving updates every count on the screen, appends to an “Actioned this session” strip and offers Undo; dismissal requires a reason
- A compliance judgement is never bulk-approvable — the card says so and shows no bulk action
A tile is a contract, not a link.
Every workspace is a full-width row tile on the BR Analytics home, and every tile answers the same seven questions: what is this, is it trustworthy, how fresh is it, is anything waiting on me, what are its headline numbers, what will I actually find when I open it, and who owns it. An administrator can triage the whole estate without opening anything — and is never surprised by what is behind a click.
- Four headline metrics ordered scale → quality → exposure → open work, each with a qualifier — “74.5 %” is unreadable without “target 85 %”
- The inside panel leads with the single most important thing waiting, in the same sentence the decision queue will use
- The attention badge counts decisions whose next step belongs to BR — never raw signals; zero renders nothing
- Em dash, not zero — an unbuilt metric shows “—”; rendering 0 claims a measurement that was never taken
- Certified metrics read from the workspace's own implementation, never tile-specific queries
✓ LiveCertified metrics · four real values · opens at Decisions
⚠ BetaReal values, caveated — “indicative, not to be quoted externally”; the caveat travels into Athena
◷ In designEm dashes · “Specification ready · build not started” · opens the specification sheet, never a dead end
🔒 RestrictedOutside your role — no numeric value rendered anywhere; names the required role and request path
Honest by Construction — the Hard Gates
If an output matches an entry below, it is wrong and does not ship. These are gates, not guidelines — they are what let an administrator trust a number enough to act on it.
| Anti-pattern | Why it misleads | What ships instead |
|---|---|---|
| A dual-axis chart | The alignment of two scales is arbitrary and invents correlation. | Two charts, small multiples, or both series indexed to one base on one axis. |
| Recolour on filter | Colour assigned by current rank repaints the survivors when a series is removed. | Colour follows the entity, in a fixed eight-slot order validated for colour-vision deficiency; a ninth hue folds into “Other”. |
| A rate without a target | 74.5 % cannot be read on its own. | Every rate carries its target bar and label. |
| A percentage from a negligible baseline | Near-zero denominators produce wild, quotable numbers. | “No stable baseline” — stated, never ∞. |
| A number on every data point | A label on everything is chaos and goes unread. | Selective direct labels — the endpoint, the extreme — plus a table twin behind every chart. |
| A tooltip as the only way to read a value | Keyboard and screen-reader users never reach it. | Tooltips enhance, never gate — every value is also in a direct label or the table view. |
| A queue that only shows what is broken | It reads as a product that never helps. | “Cleared this week” is first-class, with median time to clear. |
| A bulk action that cannot be undone | An untrusted bulk action is worth less than no bulk action. | Undo on every action, dismissal with a reason, recorded against the rule version. |
| An assistant that answers a different question | Silently answering the nearest question is invisible and therefore the most damaging failure. | An unmapped question is answered as unmapped; out-of-domain questions are declined by name, pointing to the workspace that owns the domain. |
| A tile that renders 0 for a metric never measured | Zero claims a measurement that was never taken. | An em dash, and the gap is logged. |
The Athena answer card — required order, on every workspace
1Tool trace — the approved tools that ran, shown first2Headline — a number and a claim, no preamble3Narrative — two paragraphs, every figure bold4Inline figure — “where” answered visually5Citation strip — metric ID, version, rows, freshness, lineage6Permitted actions7Three follow-ups
Three honesty rules sit under it: every number comes from a tool result — the tool registry is an allowlist and there is no raw SQL; an unmapped question is answered as unmapped; out-of-domain questions are declined by name. An empty result is an answer, not an error, and an uncertified metric carries its caveat into the answer body.
One Estate, Many Workspaces, One Inheritance
Every workspace shares the same semantic layer, metric registry, tool contract and audit trail — that shared inheritance is the product claim. A new workspace earns its tile by naming the two metrics an administrator would triage on, defines its decision taxonomy before its charts, writes its situation line from certified metrics only, wires Athena's What · Where · Why · How starters and its out-of-domain refusal list, and passes the component, anti-pattern, accessibility and governance gates before it goes live.
ECElection Change IntelligenceWhat changed, why it was allowed, what it cost and what happened downstream — the reference workspace for the standard
ENREnrollment AnalyticsEnrolled, pending and waived as of any date; Open Enrollment progress; the legacy grid kept as a tab
BPIBenefits Participation IntelligenceEligibility-first participation, coverage gaps, waivers by reason, employer cost per capita
DEMEmployee Demographics & PopulationDemographic enrollment mix by configured bands, root-cause decomposition, population data quality
ACAACA IntelligenceReadiness before filing, employee-month facts, exposure as three separate figures, filing simulation
EOIEvidence of Insurability AnalyticsPending, approved and declined EOI by plan and carrier; grace dates; time to decision
D360Dependent 360Dependants across eligibility, verification, coverage and audit — one record, every module
RV · IC · DBReVerge · Carrier Integration · Direct BillingRetro depth and payroll alignment; carrier transmission and first-pass success; premium billing and delinquency — same screens, same gates
Workspace states — live, beta, in design, restricted — are shown on each tile, never assumed. Accessibility floor across the estate: WCAG 2.2 AA, keyboard-completable including approving a decision, colour never the only channel, motion collapsing under reduced-motion.
The Four Workspaces of BR Analytics
Four questions, four workspaces, four systems of record read and never written. The workspaces open from the home tiles; Athena sits on each with its own catalog.
Enrollment analytics — who is blocked, and how is this open enrollment moving?
The enrollment workspace opens on decision tiles: what needs a human, computed by the enrollment service from the eligibility, shopping-window and coverage records. Behind them sit the chart views — pace, KPIs, exceptions, funnel, coverage, cost and contribution — sharing one period and one set of filters, and an employee ledger for the filtered population with a person overview a click away. The Ask box on this workspace, Enrollment Athena Admin, answers questions about the open-enrollment population on screen and nothing else.
What the workspace reads
- Elections — open-enrollment election rows only
- Coverage overlap from the plan coverage dates the administrator set
- Shopping window — when the employee may shop
- Eligibility as of today, plus the Access Control
Why it matters
A benefits manager sees who is stuck and why before the window closes, without a report request.
The one write here is a direct notification to the people behind a tile; its copy is drafted from the loaded facts and may not invent a deduction or a deadline.
Costs and contributions — what does this population cost the employee and the employer?
Three modes: a landing page into the cost workspace, an overview of employee cost and employer cost for the period, and the same split inside one benefit category. The charts come from the contribution and cost getters and keep employee share and employer share in separate series. The plan cost-variant code that lives on an election is never charted as money; money is the configured rule, the elected amount and the payroll actual, read in that order.
What the workspace shows
- Employee cost and employer cost for the period, side by side
- The same split inside a benefit category
- Contribution facts from the open-enrollment coverage records
- No blended rate, anywhere
Why it matters
Finance gets the employer side and HR gets the employee side from one screen that never mixes them.
The Costs & contributions report in BR Reports computes the same figures for scheduled distribution; this workspace is the interactive view.
Athena chatbot analytics — is the employee chatbot healthy?
Eight views — Overview, Usage, Intents, Quality, Knowledge, Performance, Predictions and Recommendations — on chatbot aggregates only, with last month as the default period. The Ask box on this workspace answers from the chatbot catalog and must not return a person, message text or feedback text. Forecasts and recommendation impact are estimates; a recommendation does not change an intent, a prompt or a tool. Backfill and recalculate rebuild the aggregates and nothing else.
What the workspace shows
- A-1 Overview through A-8 Recommendations
- Satisfaction from the current feedback row per interaction
- Unique users as a distinct count on conversations
- Forecasts and recommendations, labelled as estimates
Why it matters
Service-desk leaders see where the chatbot is failing employees without ever seeing an employee's chat.
The employee chatbot widget is hidden on these screens so the analytics bot and the employee bot are never on screen together — the Chatbot Analytics page covers the eight views in full.
Document analytics — how are documents moving through the document center?
Documents by category and their movement through the document center, drawn from the Document Center's own chart and drill-down actions, with a route to any single document. This workspace is operational document movement: it is not the enrollment Ask box and not the chatbot Ask box, and it computes nothing of its own.
What the workspace shows
- Documents by category, with drill-down
- Movement through the document center over the period
- A route to one document from any chart
- Computed by Document Center, drawn here
Why it matters
Dependent verification, QMCSO and case teams see the document backlog without leaving analytics.
Because the Document Center owns the actions, the same figures appear on its own home page — the analytics app never re-derives them.
“Ask Your Data” Add-Ons vs. BR AI-Powered Analytics
Most HR and benefits vendors now offer a chat box on a report. The difference is what sits behind it — one model doing everything, or three models each allowed to do one thing.
| The question | A chat box on a report | Benefits Reimagined |
|---|---|---|
| “What is our waive rate?” | The model writes SQL and hopes. | AI Query plans; the KPI engine computes — the same figure as the tile. |
| “Why did it move?” | A fluent paragraph, unsourced. | AI Insight explains with every clause linked to a record or a definition. |
| “What will OE look like?” | A confident-sounding number. | An ML forecast with a confidence band and the reasoning behind it. |
| “Anything unusual?” | Whatever the model notices. | Anomalies from the ML Engine, ranked by Spotlight, each with its records. |
| “Send a campaign to the waivers” | Done — and nobody approved it. | A draft in Communication HUB that a person edits and commits, audited there. |
| A metric the model can't compute | An estimate. | “I can't compute that” — with the reason: undefined metric, or a dimension the role may not see. |
| Our Paycom entity | A different data source, a different answer. | The same models on the same model, with the source named only when it matters. |
Athena for administrators
One Conversation, Four Kinds of Answer
The administrator types; the orchestrator routes. A “what” or “where” becomes an AI Query plan. A “what next” or “anything unusual” goes to the ML Engine. A “why” goes to AI Insight. A “do it” becomes a draft in the owning module. The answers arrive in one thread, each labelled with the model and the context that produced it.
- “What is the waive rate, and where is it worst?” — the as-of figure and the location breakdown, cohorts under five hidden
- “Why did Plant 2 turn?” — the explanation with the 412 plan movers and the 98 new waivers linked
- “What will next OE look like if nothing changes?” — the HDHP forecast with its band and its reasoning
- “Is there anything unusual this week?” — the three anomalies Spotlight ranked, with their records
- “Draft the campaign, but only for hourly new hires” — the draft in Communication HUB, narrowed as asked, waiting for approval
What's our medical waive rate, where is it worst, and what will next OE look like if nothing changes?
Athena · AI Query + ML EngineAs of 09/01 the medical waive rate is 6.0% (pending elections shown apart). It is highest in Plant 2 hourly at 16.7%, up 6.1 points since June; every other location is within a point of last quarter. If nothing changes, the ML Engine expects Plant 2's HDHP share to reach 45% at Open Enrollment (band 43–47%, confidence 0.82) and the waive rate to stay near 16% — the 412 mid-year PPO→HDHP movers tend to stay, and the 98 new waivers are mostly new hires who have not yet seen a plan comparison.
Sources: Benefits Participation · as of 09/01 · Election Change Intelligence · plan moves · ML Engine · OE forecast v4
Show Plant 2 by hire dateExplain the 98
Draft a plan-comparison campaign for those 98 only, and tell me what you can't tell me.
AthenaDrafted in Communication HUB, not sent: a plan-comparison campaign to the 98 Plant 2 hourly waivers hired since June, with the HSA seed explained. What I can't tell you: there is no peer band for hourly manufacturing waive rates in this plan type, so I have not compared you to peers; and four of the 98 sit at a site under the cohort minimum, so they are included in the send but not shown separately.
Review campaignOpen the 94 shown
Ask this workspace anything…
Why Benefits, Finance and IT Leaders Choose BR AI-Powered Analytics
Ask Where You Are
Every dashboard takes a question; the question inherits the screen's date, filters and selection.
Plans, Not Guesses
AI Query produces a plan the engine runs — the answer is the same figure as the tile, never a model's estimate.
Forecasts With a Band
Every prediction and anomaly states its confidence and links its records; boundaries keep models away from legal and filing figures.
Explained and Sourced
AI Insight writes the sentence under the number, every clause clickable.
Nothing Writes Itself
Campaigns, cases and corrections are drafted in the owning module and committed by a person — logged under NIST AI RMF.
One Group, Any HRIS
SuccessFactors plants and Paycom, Workday or UKG entities answer the same question the same way.
Where the Intelligence Lives — One Orchestrator, Every Screen
AI-Powered Analytics is not a separate product; it is the three intelligence models of the analytics orchestrator, surfaced wherever a dashboard is. Essential Insights carries the workspaces, BR Spotlight carries the ranking, BR Insights carries the explanations, Chatbot Analytics carries the interaction log — and the Analytics Strategy page explains why the models are kept apart. This page is the administrator's view of all of it: one conversation on any screen.
- Essential Insights — the workspaces where most questions are asked — enrollment, participation, demographics, ACA
- BR Spotlight — the anomalies and risks the ML Engine flags, ranked into today's list
- BR Insights — the explanation shape every “why” answer uses, with sources
- Chatbot Analytics — the same conversation, turned on the employee chatbot itself
- Analytics Strategy — the model, the context and the reason three models are better than one
Bring Any Dashboard. Ask It Anything.
Pick the report your team uses most and the question it never quite answers. We'll load your data, put Athena on that screen, ask what, where, why and what next, flag what is unusual, and draft the action — then show you the log that proves nothing was written without a person.
Ask in placeAllow-listed reads onlyConfidence on every forecastSources on every explanationZero writes by AINIST AI RMF
AI-Powered Analytics FAQs
The ability to ask any benefits dashboard a question in place — what, where, why, what next and what to do — and get a scoped answer, a forecast with its confidence, an anomaly with its records or an explanation with its sources. Behind it are three separated models in the analytics orchestrator: AI Query plans the read, the ML Engine predicts and detects, AI Insight explains.
Not as a first step, and never freely. In the shipped BR Analytics app an Ask follows a fixed order: the session, then intent routing, then playbook getters that compute the facts, and only then an allow-listed text-to-SQL read against a written table contract — the columns a bot may touch are listed in that contract and nothing else may be queried. If none of those can answer, the Ask fails rather than guesses. The model never sees a raw employee row on the analytics paths, and the design standard on this page moves every new metric onto the plan-based orchestrator with the database aggregating at source.
Forecasts and anomaly detection are Python models, not a language model, and every result carries a confidence band and the reasoning behind the expected value. They are planning inputs. They may predict a full-time transition or a threshold breach; they never set legal status, a form line, a safe-harbour figure or a filing — those stay deterministic in the engines that own them.
No. When an answer is an action, Athena asks the owning module — Communication HUB, Carrier Billing, Workflow, BR Resolve, HRIS Sync — to prepare a draft. A named person reads it, edits it and commits it there, inside that module's audit trail. Every answer is logged with the model, prompt version and context that produced it.
It says so, with the reason: the metric is not defined in the catalog, the dimension is one your role may not see, the cohort is below the minimum, or no peer band exists for that plan type. It does not estimate.
They are the same models seen from different places. BR Insights is the explanation shape used under every number; BR Spotlight is the ranked list the ML Engine's anomalies and risks feed. AI-Powered Analytics is the administrator's conversation on any dashboard that reaches all of them.
Role-based permissions scope every plan, cohorts under a configurable minimum are suppressed, every call is logged with model and prompt version, and every prompt, model and provider is inventoried and evaluated in the BR compliance engine under the NIST AI Risk Management Framework. Data stays on the platform's own database on SAP BTP or AWS.
Yes. The models read the canonical data model, so entities on Workday, UKG, ADP, Dayforce or Paycom through HRIS Sync get the same plans, forecasts, anomalies and explanations as a SuccessFactors entity. A group administrator can ask across entities in one question; the HRIS source is named only when it is the reason for a result.
Yes. The BR Benefits Analytics Design Standard governs every workspace: seven principles (work first; one definition everywhere; one context per workspace; show the basis; deviation over magnitude; recommend, then let a person act; honest by construction), four fixed screens — Decisions, Overview, Exceptions, Ledger — a tile contract, a decision card that reads as an instruction with evidence and undo, a validated colour system, chart and table rules, the Athena answer-card order, a WCAG 2.2 AA floor and a set of hard anti-pattern gates. A new workspace is a new metric catalogue and decision taxonomy, never a new visual language.
A metric with one implementation in the analytics orchestrator, a metric ID, a version, a definition, a source dataset and a freshness stamp — and the same value wherever it appears: the tile, the chart, the export and the Athena answer. Uncertified metrics are shown as beta, with a caveat that travels with the number, and unmeasured metrics render an em dash rather than zero.
The read-only administrator analytics workspace inside Benefits Reimagined: four workspaces — Enrollment analytics (decision tiles, pace, KPIs, exceptions, funnel, coverage, cost, contribution and an employee ledger), Costs & contributions, Athena chatbot analytics (eight views) and Document analytics — opened from home tiles. Its only write is a direct notification send; forecasts, recommendations and maintenance actions never change a record or the chatbot's configuration.
Because the questions have different populations. Enrollment Athena Admin answers about the open-enrollment population on screen; the chatbot analytics Ask answers aggregated chatbot KPIs; the employee Athena answers one signed-in employee. Sharing a prompt, a table list or a session between them would let a population question return a person or an employee question see the population — so they never share any of those.
It says so. The order is session, intent, playbook, allow-listed read; if none of those produces an answer the Ask fails closed and does not invent a KPI. A refusal is logged like any other answer.
One thing: a direct notification to the people behind an enrollment decision tile. The copy is drafted by a commercially available language model from the enrollment facts already loaded — with fallbacks and a fixed hard-coded message if every model fails — and it may not invent a payroll deduction or a deadline. It is a message, not an election change.
Ask the Dashboard. Trust the Answer.
Plans instead of guesses, forecasts with a band, explanations with sources, drafts that wait for a person — on SAP SuccessFactors dashboards and every entity on another HRIS.