Operations

Meeting summarization and action-item routing

Turn Zoom, Google Meet, and Teams transcripts into structured summaries and route action items into Slack and Salesforce automatically.

Meeting summarization: half of an $8.4k/yr company-wide bill is invisible integration hops and recurring-meeting context. Firedog tracks cost-per-summary across Zoom, Meet, and Teams.

Model 38%Prompt 13%Orchestration 14%Retrieval 13%History 22%

ComponentWhy it is unattributedcalls$ / run

Model · 38%Summarization and action extraction (Claude Haiku 4)Each meeting transcript is sent to Haiku 4 to produce a structured summary, decisions list, and typed action items with owners.tokens in: 18,000 · tokens out: 1,100Visible on the invoice.2$0.024

Prompt · 13%Summary template and action-item schemaA 4,000-token prefix defining the summary sections, action-item schema, and Slack/Salesforce routing rules is resent on every call with no caching.tokens in: 8,000 · tokens out: 0Visible on the invoice.2$0.008

Orchestration · 14%Transcript fetch, cleanup, and Slack/Salesforce routingA connector agent pulls the transcript from Zoom/Meet/Teams and cleans speaker labels; a routing agent posts the summary to Slack and creates follow-up tasks on the linked Salesforce record. Each is a billed call.tokens in: 5,000 · tokens out: 800Transcript-fetch and routing hops are billed to the 'integrations' cost center alongside the Zoom and Slack API usage, never to the meeting-assistant budget.3$0.009

Retrieval · 13%Prior-meeting and Salesforce context lookupFor recurring or account-linked calls, the agent retrieves prior meeting notes and the related Salesforce opportunity context to keep summaries consistent.tokens in: 4,000 · tokens out: 0Context lookups run against a separate vector store; their cost sits on the vector-DB invoice and is never joined to the per-summary spend.2$0.008

History · 22%Recurring-meeting continuity contextFor standing meetings, the last 2-3 summaries are prepended so the model can track open items across sessions, multiplying input tokens on the highest-value recurring calls.tokens in: 14,000 · tokens out: 0Continuity context is injected by the meeting app as a 'thread memory' feature; the extra prompt length shows up in the model bill but is never attributed to this workflow.1$0.014

Meeting summaries look free at ~$0.06 each - until 90 people generate them 6x a day and half the ~$8.4k/yr bill is Zoom/Slack routing hops and recurring-meeting context that scales with headcount, not value.

$0.063

cost per run

$8.4k

per year

50%

unattributed today

$0.38

per user, per day

Scenario: Company-wide meeting assistant, ~90 active users. 90 active users, 6 runs/day, 250 working days. 1 documents/run. Avg doc: 16,000 tokens.

Default model: claude-haiku-4

Trigger: Every completed Zoom/Meet/Teams call with recording enabled. A typical employee generates 4-8 summarizable meetings per day.

Who runs it: Sales, customer success, and product teams across the whole company. Every recorded call is summarized, decisions logged, and follow-ups pushed to the right owner - high volume, low cost per item, but it scales with headcount and meeting load.

What Firedog shows

Total daily and monthly summarization spend with per-run breakdown across model, integration hops, retrieval, and continuity context.
Which recurring meetings drive disproportionate cost through accumulated continuity context.
Projected annual saving from caching the summary schema prefix across every call.

Decisions this informs

Cache the 4,000-token summary and routing schema prefix; at company-wide volume this is the single cheapest saving to capture.
Limit continuity context to standing meetings only; one-off calls do not benefit from prior-summary injection.
Batch the transcript-cleanup and routing hops into a single call to cut orchestration overhead per meeting.

Meeting summarization and action-item routing

What Firedog shows

Decisions this informs

Understand every AI dollar.