AI Governance · Data Lineage

Where the AI’s numbers come from

Every AI-flagged surface, its grounding data, and refresh cadence · auditable provenance for every claim

AI surfaces

28

tracked across the prototype

Source systems

8

NetSuite, Ramp, Nexonia, Amazon Purchases, internal tracker, +3

Avg refresh cadence

Continuous

majority of surfaces · 5 min median for batch sources

Citations served / 24h

4,127

every AI-rendered number resolved to a source row

How grounding works

Every number rendered in an AI-flagged panel must come from a tool result — never inlined as prose. The Playbook’s renderer enforces this at the component level: numeric values are rendered through citation-grounded spans (tabular-nums) that link back to the source row in mock-data, NetSuite, Ramp, or another system of record. This page is the inventory of those linkages.

AI surface inventory

Each row = one AI-rendered claim or panel; each column = a piece of grounding metadata

18 of 28 surfaces
SurfaceClaim typeSource systemSource field / tableRefresh cadenceLast refreshConfidence
/clients/amazon-ads/retainer-healthAI summary card
Internal tracker
retainerEngagements.forecastTotalHoursContinuous2 min ago
High
/events/amazon-immersion-day-la/engage/statement-floralAI relationship intelligence
Internal tracker + Ramp
vendor.spendYTD, rampTransactionsContinuous5 min ago
High
/events/amazon-immersion-day-la/proposalsOutlier flag
Manual upload (PDF extract)
vendorProposals.extractedItems[].flaggedOn upload4 days ago
Medium
/admin/time-trackingBurn forecast
NetSuite
jobCodes.weeklyHours velocityWeekly2 days ago
Medium
/insights/portfolio-healthAI synthesis
Internal tracker
events.status, projectBurnContinuous2 min ago
High
/contractors/lin-zhao#feedbackPattern detection
Internal tracker
contractorFeedback.behaviorNotesOn feedback submit1 day ago
Medium
/insights/geo-costingCost recommendation
Manual entry
geoCostScenarios.optionsOn scenario refresh1 week ago
Low
/clients/amazon-ads/items-catalogCategorization
Internal tracker
catalogItems.categoryOn vendor catalog upload4 days ago
Medium
/ (Mission Control heads-up)Risk synthesis
Internal tracker
events.riskReasonContinuous1 min ago
High
/events/amazon-immersion-day-la/recommend-vendorsVendor ranking
Internal tracker + tariff data
vendor.rating, vendor.spendYTD, vendor.riskFlagsContinuous5 min ago
High
/contracts/statement-floral-decor-2026Anomaly flag
Internal tracker
contracts.anomalies[]On contract revision2 days ago
High
/insights/anomaliesPattern detection
Multi-source
globalActivities, rampTransactions, contractsHourly30 min ago
Medium
/askAsk the Playbook answers
Multi-source
All entity tablesOn queryReal-time
High
/standupMorning brief
Multi-source
events, approvals, globalActivitiesDaily 06:00 ET8 hours ago
High
/admin/ai-decisionsPast AI decisions
Multi-source
audit logContinuous2 min ago
High
/clients/amazon-ads/brand#viz-idReuse opportunity
Internal tracker
vizIDPackages.applicableTemplateSlugsOn template add21 days ago
Medium
/approvalsApproval recommendation
Internal tracker + NetSuite
approvals.aiSuggestion, aiConfidenceOn approval queue update12 min ago
High
/admin/ai-costsSpend rollup
Anthropic + OpenAI billing
aiCostLedger.tokenSpendHourly45 min ago
High
Sample citation trail

Below is the actual source chain for one AI claim — the Statement Floral $4.2M YTD figure

  1. 1

    UI component

    <EngagementPanel /> at /events/amazon-immersion-day-la/engage/statement-floral
  2. 2

    AI claim

    “Statement Floral — $4.2M YTD across 14 events”
  3. 3

    Renderer

    <span className="font-medium tabular-nums">{vendor.spendYTD}</span> — pure data, no AI inference
  4. 4

    Data source

    vendors.find(v => v.slug === 'statement-floral').spendYTD === 4_200_000
  5. 5

    System of record

    NVE internal vendor master · synced from NetSuite GL spend rollup nightly
  6. 6

    Last attestation

    2026-05-08 via finance-ops review
When grounding fails

What happens if a source system is down or a number can’t be cited

  • If the source row doesn’t exist, the AI surface renders an em-dash placeholder, not a fabricated number.
  • If the source system is unreachable (e.g. NetSuite outage), the panel shows a “Source unavailable” state with last-good timestamp.
  • If the AI confidence is below threshold (Low), the surface adds an “AI-tentative” Badge prompting human review.
  • Every claim has a click-through to its source row — anyone can verify.

See also

Lineage answers “where does the number come from?”. The decision log answers “what did the human do with it?”.

AI decision log

Inventory derived from observable code · 18 surfaces catalogued of 28 AI-rendered panels · AI flags, doesn’t fix