ComponentWhy it is unattributedcalls$ / run
Model · 15%Classification and summarization (Claude Haiku 4)Each news item or filing alert is sent to Haiku 4 for materiality classification (5-tier) and a 150-token structured summary.tokens in: 2,200 · tokens out: 180Visible on the invoice.1$0.003
Prompt · 24%Classification schema and entity metadataA 4,800-token prompt carries the materiality taxonomy, covered-name list, and 6 few-shot examples. Sent on every call with no caching enabled.tokens in: 4,800 · tokens out: 0Visible on the invoice.1$0.005
Orchestration · 11%Ingestion, dedup, and routing pipelineA dedup agent checks the item against a 24-hour seen-items cache; a routing agent sends Tier-1 alerts to a downstream escalation model. Both are billed calls.tokens in: 1,400 · tokens out: 200Dedup and routing agents are considered 'infrastructure' and billed to the platform cost center, not the signal workflow budget.2$0.002
Retrieval · 2%Entity resolution and prior-alert lookupEach item triggers an entity resolution query to map freetext company names to internal tickers, plus a prior-alert lookup to detect duplicates.tokens in: 900 · tokens out: 0Entity resolution queries are served by an in-house NER model whose inference cost is allocated to the infra team, not the trading workflow.2$0.001
History · 39%Coverage context re-injectionFor Tier-1 items, the last 3 alerts on the same ticker are prepended to provide trend context. This multiplies input tokens for the highest-value subset of items.tokens in: 7,500 · tokens out: 0Coverage context is injected by the alert-enrichment service; it shows up as increased prompt length in the model bill but is never attributed to this workflow.1$0.008
Connectors · 9%Slack, Teams, Zapier, and Salesforce routing connectorsEach material item is routed to Slack and, for Tier-1, cross-posted to Teams, logged to Salesforce, and handed to downstream Zaps. Routing carries a small payload-formatting call per destination; at 400 runs/day the connector volume dominates the marginal cost of the cheap classification model.tokens in: 700 · tokens out: 120Slack/Teams posts, Salesforce logging, and each Zapier task bill to their own platform plans; the routing volume scales with alert flow, not headcount, and never reaches the triage workflow budget.2$0.002
At $0.0205/run and 400 runs/day, news triage looks trivially cheap - until you realize 61% is uncached prompt overhead, routing infrastructure, coverage context, and Slack/Teams/Zapier connector traffic that scales with alert volume, not analyst headcount.