RetailOS: Agentic AI for Retail

Command CenterLive ยท 847 SKUs ยท 12 Stores
Revenue Today
ยฃ284K
โ†‘ 18% vs last Tue
Gross Margin
47.2%
โ†‘ 3.1pts AI pricing
Stockout Alerts
3
POs auto-raised
Basket Uplift (AI)
+ยฃ8.40
Personalisation engine
๐Ÿค– AI Agent Status
14 retail AI agents across inventory, pricing, customer, and operations
Demand Forecasting847 SKUs ยท 92% accuracy
Dynamic Pricing EngineLive ยท 2,847 price changes
Personalisation AI+ยฃ8.40 basket uplift
Inventory Intelligence3 stockouts ยท POs raised
Churn Prevention12 high-risk customers
Returns IntelligenceReturns โˆ’23% YoY
๐Ÿ“ก Live Retail Intelligence Feed
Real-time AI across all 12 stores and ecommerce channels
Top Inventory Alerts
โš  SKU-0847 ยท Nike Air Max 90 UK8 ยท 0 units ยท 47 sold/day โ†’ PO auto-raised: 200 unitsSTOCKOUT
โš  SKU-1203 ยท Patagonia Fleece M ยท 4 units ยท predicted 0 in 2 days โ†’ Reorder NOWLOW STOCK
โ†“ SKU-2847 ยท Winter Boots ยท 84 units ยท predicted demand โˆ’62% โ†’ Markdown 30% recommendedOVERSTOCK
Why RetailOS
๐Ÿ“ฆ Inventory: The Hidden P&L Killer
Retailers lose 4% of revenue to stockouts and tie up 30% of working capital in overstock. RetailOS predicts demand at SKU-store-day level, raises POs automatically, and recommends markdowns before overstock becomes deadstock.
๐Ÿ’ฐ Pricing: Margin Left on the Table
Static pricing leaves 8โ€“12% of gross margin uncaptured. Dynamic Pricing Engine adjusts prices in real time based on demand signals, competitor pricing, inventory levels, and elasticity models โ€” 3.1pts GM improvement.
๐ŸŽฏ Personalisation: Every Customer Different
Generic promotions convert at 1โ€“2%. Personalised recommendations convert at 8โ€“14%. The Personalisation AI serves unique offers to every customer based on purchase history, browse behaviour, and predicted intent โ€” +ยฃ8.40 basket uplift.
Total Agents
14
Decisions/Hour
12,847
Revenue Impact
+ยฃ284K
GM Uplift
+3.1pts
Inventory & Merchandising
๐Ÿ“ˆ
Demand Forecasting
SKU-store-day demand prediction from sales history, seasonality, weather, local events, and competitor promotions. 92% accuracy. Drives auto-replenishment and markdown decisions.
Running ยท 847 SKUs
Reflection + Time Series
๐Ÿ“ฆ
Inventory Intelligence
Real-time stock monitoring across all stores and DCs. Detects stockouts 48โ€“72h before they occur. Raises POs automatically. Optimises safety stock by SKU and location.
Running ยท 3 alerts
Sequential + Rules
๐Ÿ’ฐ
Dynamic Pricing Engine
Adjusts prices in real time from demand elasticity, competitor pricing, inventory levels, and margin targets. 2,847 price changes today. +3.1pts gross margin. Human approval for category-level rules.
Running ยท 2,847 changes
ReAct + Elasticity
Customer Intelligence
๐ŸŽฏ
Personalisation AI
Serves unique product recommendations, promotions, and content to every customer across email, app, and in-store. +ยฃ8.40 basket uplift. Conversion rate: 11.4% vs 1.8% generic.
Running ยท 47K customers
ReAct + Collaborative
โค๏ธ
Churn Prevention
Detects lapsing customers 90 days before churn using purchase recency, engagement decline, and competitor signals. Triggers personalised win-back offers. 67% retention success rate.
Running ยท 12 at risk
ReAct + Signals
๐Ÿ”ฎ
Next Best Action
Determines the optimal next action for each customer touchpoint: promotion, recommendation, loyalty reward, or service intervention. Maximises lifetime value across all channels.
Running ยท Live
Planning + LTV Model
Operations Intelligence
๐Ÿค
Supplier Intelligence
Monitors supplier lead times, quality scores, and price trends. Flags supply chain risks before they become stockouts. Negotiation intelligence from market pricing benchmarks.
Running ยท 84 suppliers
ReAct + Supply Chain
๐Ÿช
Store Operations AI
Optimises staffing schedules from demand forecasts, monitors planogram compliance via image AI, and tracks store KPIs in real time. Labour productivity +23%.
Running ยท 12 stores
Planning + Vision
๐Ÿ”„
Returns Intelligence
Predicts return likelihood at point of purchase, identifies serial returners, optimises returns routing, and extracts product quality signals from returns data. Returns rate โˆ’23% YoY.
Running ยท Returns โˆ’23%
Reflection + ML
Total SKUs
847
Stockouts
3
Overstock SKUs
47
Inventory Accuracy
99.2%
๐Ÿ“ฆ Top SKU Intelligence
Demand vs stock ยท AI forecast ยท recommended action
SKU-0847
Nike Air Max 90 UK8
STOCKOUT
SKU-1203
Patagonia Fleece M
2 days
SKU-2847
Winter Boots
84 units
SKU-3412
Levi's 501 32/30
Optimal
SKU-4891
Adidas Stan Smith W7
Optimal
SKU-5023
North Face Jacket XL
STOCKOUT
๐Ÿ“Š Inventory Health Summary
Across all 12 stores and distribution centres
Stockouts (3): ยฃ47K revenue at risk per day. POs auto-raised for all 3. Expected stock arrival: 48โ€“72h. Expedited shipping recommended for SKU-0847 (highest velocity).
Overstock (47 SKUs): ยฃ284K working capital tied up. Seasonal winter items. Markdown cascade recommended: 15% now โ†’ 30% in 14 days โ†’ 50% in 28 days. Predicted clearance: 94%.
Optimal (797 SKUs): 94.1% of SKUs within target service level. Safety stock calibrated to 97.5th percentile of demand variability.
SKUs Forecast
847
Accuracy (MAPE)
92%
Horizon
16 weeks
Forecast Signals
12
๐Ÿ“ˆ Demand Forecasting โ€” Signal Architecture
RetailOS demand forecasting combines 12 signal types per SKU per store: (1) Own sales history โ€” trend, seasonality, and weekly patterns. (2) Weather โ€” temperature drives clothing category demand, precipitation drives footwear. (3) Local events โ€” concerts, sports fixtures, school calendars. (4) Competitor promotions โ€” pricing intelligence triggers demand shifts. (5) Social signals โ€” trending products detected 1โ€“2 weeks before store demand. (6) Economic indicators โ€” consumer confidence and discretionary spend signals. Forecast horizon: 16 weeks at day-store level. Average MAPE: 8% (vs 22% manual planning). Safety stock reduction: 34% without service level degradation.
Price Changes Today
2,847
GM Uplift
+3.1pts
Revenue Lift
+8.4%
Competitor Prices Tracked
47K
๐Ÿ’ฐ Dynamic Pricing โ€” How It Works
Dynamic Pricing Engine adjusts prices in real time within human-approved guardrails. Inputs: demand elasticity by SKU and customer segment, competitor pricing (tracked every 4 hours), inventory levels, margin targets, and promotional calendar. When demand for SKU-0847 exceeds forecast by 40%, the engine raises price by 8% โ€” capturing margin before stockout. When SKU-2847 is tracking 62% below forecast, a 30% markdown is recommended to the category manager for approval. No price changes are made autonomously โ€” the engine recommends, category managers approve. All competitor price tracking is from public sources only.
Customers Profiled
47K
Basket Uplift
+ยฃ8.40
Email Conversion
11.4%
vs 1.8% generic
Revenue from AI Recs
ยฃ47K
Today
๐ŸŽฏ Personalisation Intelligence
Personalisation AI builds a real-time preference model for every customer from purchase history, browse behaviour, return patterns, and demographic signals. For each touchpoint (email, push notification, homepage, in-store screen), it selects the optimal product recommendation, promotion, and message. Customer A (trail runner, buys premium): shown new trail shoe launch with zero discount. Customer B (bargain hunter, price sensitive): shown clearance footwear with 30% off message. Same campaign budget, completely different execution. GDPR-compliant: all personalisation based on consented first-party data. Customers can view and delete their preference profile at any time.
At-Risk Customers
12
High LTV ยท 90-day flag
Win-back Success Rate
67%
LTV at Risk
ยฃ84K
Avg Customer LTV
ยฃ7,000
โค๏ธ Churn Prevention Intelligence
Churn Prevention Agent monitors 40+ engagement signals for every loyalty customer: purchase recency and frequency decline, email open rate drop, app session reduction, competitor mentions in support interactions, and NPS trend. Flags high-LTV customers showing early lapse signals 90 days before predicted churn โ€” while win-back investment still makes economic sense. Each at-risk customer receives a personalised intervention: exclusive early access offer, loyalty points boost, personal shopping appointment, or targeted markdown on their favourite category. Win-back success rate: 67% on high-LTV customers. ROI on retention spend: 12:1.
Active Suppliers
84
Lead Time Risks
3
On-Time Delivery
94%
Cost Savings (AI Negs)
ยฃ284K
๐Ÿค Supplier Intelligence
Supplier Intelligence monitors 84 suppliers across delivery performance, quality scores, price trends, and supply chain risk signals. Port congestion, manufacturing delays, and currency movements detected 4โ€“6 weeks ahead of impact. Negotiation intelligence: AI benchmarks each supplier's prices against market rates, identifies where margin is being left and surfaces the data for buyer negotiations. Supplier performance dashboards automatically generated for quarterly reviews. Alternative supplier recommendations provided when primary supplier risk exceeds threshold.
Stores Monitored
12
Labour Productivity
+23%
Planogram Compliance
94%
Footfall Conversion
34%
โ†‘8pts AI optimisation
๐Ÿช Store Operations Intelligence
Store Operations AI combines footfall prediction, labour scheduling, planogram compliance monitoring, and real-time KPI tracking for all 12 locations. Staffing: demand forecast drives optimal shift scheduling โ€” right staff, right time, right skills. 23% labour productivity improvement. Planogram compliance: vision AI checks shelf layout against plano during store hours and flags non-compliance to store managers. Conversion: footfall and basket data identifies where customers are dropping off in the purchase journey โ€” merchandising adjustments recommended with expected conversion impact.
Return Rate
8.4%
vs 12.1% pre-AI
Return Rate Reduction
โˆ’23%
Serial Returners ID'd
284
Saved P&L (Annual)
ยฃ840K
๐Ÿ”„ Returns Intelligence
Returns Intelligence reduces return rates by acting at three points: (1) Pre-purchase: size recommendation accuracy and detailed product content reduce size/fit returns by 34%. (2) Post-purchase: returns likelihood scored at order dispatch โ€” high-risk items get personalised product use guidance and care instructions. (3) Post-return: product quality signals extracted from return reasons feed directly to buying teams and suppliers. Serial returner identification enables policy enforcement without harming genuine customers. Return routing optimisation reduces processing cost by 18%.
Agents Active
14
Decisions/Hour
12,847
Revenue Impact
+18%
GM Uplift
+3.1pts
๐Ÿ“ก Live Agent Trace
๐Ÿ›ก Retail AI Governance
Pricing: category manager approval: All price changes require category manager sign-off. Dynamic pricing operates within pre-approved guardrails. No automated price changes without human approval.
Personalisation: GDPR consent: All personalisation based on consented first-party data only. No third-party data sharing. Customers can view, correct, and delete their profile.
Competitor pricing: public sources only: All competitor intelligence from publicly available pricing. No scraping of restricted sources.
AgentOps โ€” Live Agent Observability

๐Ÿ“ก Live Trace Feed

๐Ÿ“Š Session Metrics (24h)

Total Sessions2,847
Avg Latency1.4s
P95 Latency3.1s
Error Rate0.3%
Tool Calls12,284
HITL Escalations47
RAGAS GatePASS โœ“

๐Ÿ’ฐ Cost & Tokens

Cost (24h)ยฃ847
Input Tokens48.2M
Output Tokens12.4M
Cache Hit Rate67%
Cost/Sessionยฃ0.30

๐ŸŽฏ RAGAS Quality Scores

Faithfulness0.94 โœ“
Answer Relevance0.91 โœ“
Context Precision0.89 โœ“
Context Recall0.93 โœ“
Hallucination Rate0.8%

๐Ÿค– Agent Health

All agentsHealthy
OrchestratorActive
Tool registryOnline
MCP serversConnected
Memory storeHealthy
MLOps / LLMOps โ€” Model Lifecycle

๐Ÿง  Model Registry

claude-sonnet-4-5 PRODUCTIONPrimary
claude-haiku-4-5 ROUTINGFast path
claude-opus-4-5 SHADOWComplex
text-embedding-3-large RAGVectors

Automatic fallback routing. Versioned in MLflow. Prompt changes require RAGAS eval gate pass.

๐Ÿ“ˆ Drift Detection

Faithfulness drift (7d)+0.02 stable
Latency drift (7d)+120ms watch
Output length driftWithin ยฑ5%
Sentiment driftNo anomaly
Alert thresholdฮ”>0.05 โ†’ PagerDuty

๐Ÿ”€ A/B Experiment Controller

Prompt v2.3 vs v2.4Running
CoT vs DirectStaging

Statistical significance (p<0.05) required before promotion.

๐Ÿช Feature Store

Vector IndexPinecone
Dimensions3,072
Indexed Docs284K
Retrieval P9542ms

๐Ÿ“ฆ Prompt Version Control

System promptsGit-tracked
Few-shot examplesVersioned
Eval datasetsDVC tracked
DevSecOps โ€” Security-First CI/CD Pipeline

๐Ÿš€ CI/CD Pipeline

๐Ÿ”SAST โ€” Semgrep + BanditPASS
๐Ÿ“ฆSCA โ€” SBOM + TrivyPASS
๐ŸงชUnit + Integration tests847/847
๐ŸŽฏRAGAS eval gate (โ‰ฅ0.92)0.94 โœ“
๐Ÿ”Secrets scan โ€” GitleaksCLEAN
๐ŸณContainer scan โ€” Grype0 CRITICAL
๐ŸšขDeploy โ†’ KubernetesDEPLOYED

๐Ÿ” Security Posture

RBAC โ€” Role-based accessEnforced
API keys โ€” HashiCorp VaultRotated 30d
mTLS โ€” Istio service meshActive
PII scrubbing โ€” NeMoActive
Audit log โ€” ImmutableCloudWatch
Pen testQuarterly
SOC 2 Type IIIn progress
ISO 27001Compliant

๐Ÿ— Infrastructure as Code

TerraformCloud infra
HelmK8s workloads
ArgoCD GitOpsSynced
Kustomize overlaysdev/stg/prd

โ™ป๏ธ Rollback & DR

RTO Target<15 min
RPO Target<5 min
Blue/Green DeployActive
Auto-rollbackError rate >1%

๐Ÿ“‹ Regulatory Compliance

GDPR Art. 22 HITLEnforced
EU AI Act Art. 9Documented
NIST AI RMFMapped
ISO/IEC 42001Compliant
AI Observability โ€” OpenTelemetry + Langfuse

๐Ÿ”ญ Observability Stack

L1TracesOpenTelemetry โ†’ Jaeger
L2MetricsPrometheus โ†’ Grafana
L3LLM TracesLangfuse (self-hosted)
L4LogsFluentd โ†’ OpenSearch
L5AlertsAlertManager โ†’ PagerDuty

๐Ÿ“Š SLO Dashboard

Availability SLO99.9% target
Current (30d)99.96%
Error Budget73% remain
P50 Response0.8s
P95 Response3.1s
P99 Response7.4s

๐Ÿšจ Active Alerts

Latency P95Normal
Error rate0.3% โœ“
Token budget84% remain
RAG recall0.93 โœ“
Latency drift+120ms watch

๐Ÿ”ฌ Langfuse Trace Explorer

๐Ÿ“ˆ Avg Span Breakdown

API Gateway12ms
Auth + RBAC8ms
RAG retrieval42ms
Guardrail check18ms
LLM inference1,240ms
Tool execution84ms
Total E2E1,452ms
Guardrails โ€” Responsible AI Framework

๐Ÿ›ก NeMo Guardrails โ€” Active Rails

โœ… Human-in-the-Loop (HITL) Gate
All consequential actions require human approval before execution. Confidence <0.85 always escalates. GDPR Article 22 compliant โ€” no fully automated consequential decisions.
๐Ÿ” PII Detection & Scrubbing
Microsoft Presidio + custom patterns. Names, emails, NI/SSN, card numbers scrubbed from all LLM I/O before logging. 47 entity types across 12 jurisdictions.
๐Ÿšซ Toxicity & Hallucination Filter
NeMo topic rails block off-topic responses. Factual grounding check cross-references every claim against retrieved context. Hallucination >5% triggers human review queue.
โฑ Rate Limiting & Abuse Prevention
Per-user token budgets at API gateway. 10ร— anomalous usage triggers suspension + security alert. Cloudflare WAF DDoS protection.

๐Ÿ“‹ Audit Trail & Explainability

๐Ÿ“ Immutable Decision Log
Every AI recommendation logged: input context, retrieved docs, reasoning chain, confidence, model version, user ID, timestamp. 7-year retention for regulated decisions.
๐Ÿ”Ž Explainability (XAI)
Every recommendation includes source citations, confidence intervals, alternatives considered, and limitation disclosures. SHAP attribution for structured ML models.
โš–๏ธ Bias Monitoring
Fairness metrics tracked across protected characteristics. Disparate impact analysis monthly. EU AI Act Article 10 data governance requirements met.
๐Ÿ› Regulatory Mapping
GDPR Art. 5/22 ยท EU AI Act Art. 9/10/13/14 ยท NIST AI RMF ยท ISO/IEC 42001 ยท IEEE 7001 Transparency. Compliance evidence pack generated quarterly.
0.3%
Hallucination Rate
Target <2%
100%
HITL Coverage
Consequential acts
0
PII Leaks (30d)
Target: 0
A+
Security Grade
Mozilla Observatory
Multi-Agent Architecture โ€” Mesh & Orchestration

๐Ÿ•ธ Agent Mesh Topology

Orchestrator
Agent 1
Agent 2
Agent 3
Agent 4
Agent 5
Agent 6

Orchestrator decomposes tasks, routes to specialists, aggregates results, handles conflicts. All inter-agent communication via typed schemas. No agent takes external action without Orchestrator validation.

โš™๏ธ Agent Patterns

ReAct โ€” Reason + Act loopsAnalytical
Reflection โ€” Self-critique cyclesHigh-stakes
Planning โ€” Hierarchical decompositionMulti-step
RAG โ€” Retrieval-augmented genKnowledge
HITL โ€” Human-in-the-loopAll consequential
Tool Use โ€” Function callingAll agents

๐Ÿ”„ Temporal.io Orchestration

Active Workflows2,847
HITL Signals Pending47
Retry PolicyExp backoff ร—3
Saga PatternCompensating txns
Durable ExecutionCrash-safe โœ“

๐Ÿ“จ Kafka Message Bus

Topics47 agent topics
Throughput12K msgs/s
Consumer Lag<100ms
Schema RegistryConfluent
Dead Letter QueueMonitored

๐Ÿ”Œ MCP Integration Layer

MCP โ€” Data sourcesActive
MCP โ€” CRM/ERPActive
MCP โ€” Document storeActive
OAuth 2.0 authAll connectors
JSON Schema validationAll tools
Evaluation Framework โ€” Continuous Quality Gates
0.94
Faithfulness
Gate โ‰ฅ0.92 โœ“
0.91
Answer Relevance
Gate โ‰ฅ0.88 โœ“
0.89
Context Precision
Gate โ‰ฅ0.85 โœ“
0.93
Context Recall
Gate โ‰ฅ0.90 โœ“

๐Ÿงช Eval Suite Composition

Golden dataset2,847 Q&A pairs
Unit evals (per agent)120โ€“400 cases
Integration evals84 end-to-end flows
Adversarial probes47 jailbreak tests
LLM-as-judgeclaude-opus-4-5
Human eval cadenceWeekly 5% sample

๐Ÿ” Eval-Driven Dev Flow

1
Change proposed โ†’ PR opened
Automated eval suite runs against golden dataset in CI. Results posted to PR.
2
RAGAS gate enforced
All metrics must meet thresholds. Failure blocks merge.
3
Canary deploy (5%)
Langfuse online evals on live traffic. Drift alerts trigger auto-rollback.
4
Full rollout + monitor
Weekly human eval sample. Monthly RAGAS full re-run.
Infrastructure โ€” Kubernetes ยท Scale ยท Resilience

โ˜ธ๏ธ Kubernetes Cluster

ClusterEKS / GKE / AKS
Node pools3 (system ยท app ยท GPU)
HPA targetCPU 70% โ†’ scale
KEDA triggersKafka consumer lag
Spot instances80% non-critical
Multi-AZ3 zones

๐Ÿ’พ Data Architecture

PostgreSQL (RDS)Operational
Redis (ElastiCache)Session + cache
Pinecone / pgvectorVector search
S3 Intelligent TierDocuments
Kafka (MSK)Event streaming
Snowflake / BigQueryAnalytics DWH

๐Ÿ’ฐ Cost Architecture

LLM API (Anthropic)~45% of AI cost
Vector DB~12% of AI cost
Compute (K8s)~28% of AI cost
Prompt cache savingsโˆ’67% input tokens
Haiku fast-path savingโˆ’40% LLM spend
Est. monthly totalยฃ8โ€“28K

๐Ÿ” Disaster Recovery

1
Primary failure detected (<2 min)
Route53 health check fails โ†’ DNS failover. Temporal promotes standby. Kafka MirrorMaker live.
2
DR validates (<5 min)
Smoke tests auto-run. PagerDuty alert to on-call. RTO target: 15 minutes.
3
Data reconciled (<15 min)
PostgreSQL read replica promoted. S3 cross-region lag <5min. RPO: 5 minutes.

๐Ÿ“Š Capacity Planning

  • Baseline: 3 app nodes ยท 2 vCPU ยท 8GB RAM each
  • Scale trigger: Kafka consumer lag >10K msgs
  • Max scale: 20 nodes via KEDA + HPA
  • LLM concurrency: 50 parallel sessions managed
  • Vector search: Pinecone p1 โ†’ p2 at 500K docs
  • DB connections: PgBouncer pool (max 500)
Documentation โ€” Deployment Guide & Runbook

๐Ÿš€ 10-Week Deployment Guide

1
Week 1โ€“2: Data Foundation & Infrastructure
Deploy K8s cluster. Provision Temporal.io, Kafka, PostgreSQL, Pinecone. Connect source systems via MCP. Establish data governance and RBAC. Run baseline eval on golden dataset.
2
Week 3โ€“4: Core Agents Live
Deploy first 3 highest-value agents. Wire HITL approval workflows in Temporal. Configure NeMo guardrails and PII scrubbing. Set up Langfuse tracing and RAGAS eval gate.
3
Week 5โ€“7: Full Agent Mesh
Deploy all agents. Configure Orchestrator routing. A/B test prompt variants. Enable drift detection. Train end-users on HITL workflow.
4
Week 8โ€“10: Production Hardening
Pen test + SAST/DAST scan. Load test 10ร— baseline. Configure PagerDuty. Compliance review (GDPR, EU AI Act). Produce runbook. Go-live.

๐Ÿ— 7-Layer Platform Stack

L7PresentationReact ยท Next.js ยท SSO
L6API GatewayFastAPI ยท OAuth2 ยท WAF
L5OrchestrationTemporal.io ยท LangGraph
L4Agent RuntimeNeMo ยท RAGAS ยท Tools
L3Model + ToolsClaude API ยท MCP servers
L2Data + IntegrationKafka ยท PostgreSQL ยท Redis
L1ObservabilityOTel ยท Langfuse ยท Grafana

๐Ÿ”Œ Integration How-To

  • MCP server per data source (REST/GraphQL/gRPC)
  • OAuth 2.0 service account per enterprise system
  • Kafka topics per agent capability namespace
  • Schema registry for typed message contracts
  • Data lineage via OpenLineage โ†’ Marquez
  • Webhooks for real-time event ingestion
  • dbt + Airflow for batch data refresh

๐Ÿ‘ค RBAC User Roles

ViewerRead dashboards
AnalystRun queries + export
ApproverHITL decisions
ManagerConfig + agents
AdminFull platform
AI EngineerModels + prompts

IdP via Okta/Azure AD. MFA enforced for Approver+.

๐Ÿ“ž Incident Runbook

  • High latency (>5s): Check Langfuse trace โ†’ vector store โ†’ LLM API status
  • RAGAS gate fail: Roll back last prompt change โ†’ notify AI engineer
  • Error spike: Circuit breaker โ†’ fallback to previous version
  • PII leak: Suspend session โ†’ DPO notification within 24h
  • HITL queue backup: Escalate to senior approver
  • Cost overrun: Auto-throttle โ†’ route to Haiku