How Anthropic Claude Powers IMPACT and CostKatana
At Hypothesize, we build products that solve real problems with AI. Two of our flagship offerings, IMPACT, built with P3M.AI, and CostKatana, rely on Anthropic's Claude models as a core intelligence layer. This case study details exactly how we use Claude in each product, from model selection and orchestration to reporting and cost tracking.
CostKatana: How Claude Is Used
1. Governed Agent (AI Task Orchestration)
The Governed Agent selects Claude models by task complexity:
- Claude Sonnet 4.5 – Coding, high complexity, high risk
- Claude Sonnet 4 – Medium complexity
- Claude Haiku 4.5 – Simple or low-complexity tasks (faster, cheaper)
This tiered approach ensures the right model for the job, using Sonnet 4.5 for production code and deep reasoning, and Haiku 4.5 for lightweight tasks to optimize cost and latency.
2. Auto Recommendation Agent
Uses Claude Haiku 4.5 (via AWS Bedrock) for cost optimization advice based on usage patterns.
3. Visual Compliance
Claude 3.5 Sonnet is used for visual and design compliance checks.
4. Chat and AI Routing (Cortex)
Claude models are exposed through the AI gateway as one of the supported providers. Requests are routed to the correct model via Bedrock and the AI router.
5. Pricing and Cost Tracking
Anthropic is configured as a provider in the pricing registry. Cost calculation and token counting use Anthropic's tokenizer (@anthropic-ai/tokenizer). Model pricing for Haiku, Sonnet, and Opus (3, 3.5, 4, 4.5, 4.6) is stored and used for cost estimates.
6. AWS Bedrock Integration
Claude is invoked through AWS Bedrock, which provides multiple Claude variants (Haiku, Sonnet, Opus), regional deployment control, and fallbacks when inference profiles are missing.
7. Marketing and Landing Pages
The marketing site highlights Claude as a supported provider in pricing tables and provider logos.
CostKatana Summary
| Use Case | Model(s) Used | Purpose |
|---|---|---|
| Governed Agent | Sonnet 4.5, 4, Haiku 4.5 | Task planning and execution by complexity |
| Auto Recommendations | Haiku 4.5 | Cost optimization suggestions |
| Visual Compliance | Sonnet 3.5 | Visual/design compliance checks |
| Chat / AI Gateway | All supported models | User-facing chat and AI routing |
| Cost Tracking | All models | Usage and cost analytics |
| Experimentation | Various | A/B testing and model comparison |
P3M (IMPACT): How Claude Is Used
Infrastructure
Claude is accessed via AWS Bedrock (not the Anthropic API directly). The default model is Claude 3.7 Sonnet (anthropic.claude-3-7-sonnet-20250219-v1:0). A fallback model, Claude 3.5 Haiku, is used when the main model fails. Configuration is via env vars: AI_MODEL=claude, CLAUDE_INFERENCE_PROFILE_ARN, FALLBACK_CLAUDE_INFERENCE_PROFILE_ARN. The system can be switched to Nova Pro by setting AI_MODEL=nova.
Cost & Usage
CostKatana is used for usage tracking. Per-organization AI credits are enforced, calls are blocked without sufficient credits. Token usage and costs are recorded in the AIUsage model.
Where Claude Is Used (Use Cases)
| Feature | Service | Purpose |
|---|---|---|
| KPI generation | aiKpiService | Generate KPIs from category name, description, industry context, and requirements |
| Criteria enhancement | aiKpiService | Improve/refine KPI criteria and task templates |
| Audit summaries | aiAuditService | Analyze audit data and create summaries |
| Audit analysis | aiAudit.controller | Deep analysis of audit results |
| Auditee reports | aiAuditeeReportService | AI reports for individual auditees using audit and assessment data |
| Anecdotal records | aiAnecdotalRecordService | Vertical, horizontal, and trend analysis of performance records |
| Overall reports | overAllReportService | Dashboard reports combining audits, tasks, and trends |
| Video audits | videoAuditService | AI analysis of video audit content |
| One-time audits | oneTimeAuditAIService | AI analysis for ad-hoc audits |
| Assessment reports | assessmentAIService | AI-generated assessment reports |
| Group analytics | groupAnalyticsService | AI insights and comparisons across groups |
| Dashboard insights | aiDashboardInsightService | AI insights for dashboards |
| AI chatbot | aiChatbotService | Intent detection, extraction, and conversational responses |
Claude powers content creation and analysis in P3M across reports (audit, auditee, assessment, video, one-time, and dashboard), KPI setup (generation, criteria refinement, task templates), anecdotal records (trend and horizontal analysis), group analytics, the AI chatbot, and dashboard insights. Note: AI relevance filtering for search results (e.g., image-based filtering) uses Amazon Nova Pro in AIRelevanceFilterService, not Claude.
Summary
Both CostKatana and P3M rely on Claude as a core AI layer: CostKatana for governed orchestration, cost optimization, and tracking; P3M for audit analysis, report generation, KPI refinement, and conversational AI. AWS Bedrock provides scalable, regional access to multiple Claude variants, while CostKatana tracks usage and costs across the stack.
Interested in how we can help you leverage Claude for your products? Get in touch.