HR and benefits helpdesk chatbot

Employee-facing chat over Workday data and policy docs, answering PTO, benefits, payroll, and policy questions inside Slack or Teams with multi-turn memory.

HR helpdesk chatbot: $13.1k/yr, 67% invisible in session-history compounding, nightly policy re-embedding, and Workday fetch hops. Firedog shows the per-session cost curve People Ops never sees.

Model 17%Prompt 15%Orchestration 4%Retrieval 17%History 47%

ComponentWhy it is unattributedcalls$ / run

Model · 17%Conversational inference (Claude Sonnet 4)Each employee turn sends the retrieved policy context plus the running conversation to Sonnet 4 and returns a grounded, sourced answer.tokens in: 26,000 · tokens out: 1,400Visible on the invoice.6$0.099

Prompt · 15%HR persona, citation rules, and Workday field mapA 5,000-token prefix defining tone, policy-citation rules, and the Workday field mapping (PTO, comp, benefits) is resent on every turn without caching.tokens in: 30,000 · tokens out: 0Visible on the invoice.6$0.090

Orchestration · 4%Intent classification and Workday data fetchEach turn hits a lightweight classifier (leave vs. payroll vs. policy) that decides whether to call Workday for employee-specific data such as a live PTO balance. The classifier and Workday-fetch hops are billed separately.tokens in: 5,000 · tokens out: 400Intent-classification and Workday-fetch calls are made by the chatbot middleware and land on a platform cost center separate from the People Ops budget.6$0.021

Retrieval · 17%Policy-document vector lookups per turnEvery turn retrieves the top policy and handbook chunks from a Box/Workspace-backed index. The index is rebuilt nightly including unchanged documents.tokens in: 24,000 · tokens out: 0The nightly full re-embed of the policy library regardless of what changed sits on the embedding-provider invoice, not the chatbot budget.6$0.10

History · 47%Full session transcript re-sent each turnThe complete conversation from turn 1 is resent on every subsequent turn. By turn 6 this is 40k+ tokens of prior dialogue re-billed as context.tokens in: 90,000 · tokens out: 0Session history is a first-class feature of the chat framework; its token volume is never disaggregated from per-turn inference cost in any dashboard People Ops runs.6$0.27

The HR helpdesk costs ~$13.1k/yr to run; 47% is full session history re-billed on every turn and another 20% is nightly policy re-embedding and Workday fetch hops nobody set a policy on.

$0.58

cost per run

$13k

per year

67%

unattributed today

$1.16

per user, per day

Scenario: Company-wide HR assistant, ~45 weekly active employees. 45 active users, 2 runs/day, 250 working days. 6 documents/run. Avg doc: 5,000 tokens.

Default model: claude-sonnet-4

Trigger: Ad-hoc employee questions throughout the day, spiking during enrollment, review, and payroll cycles. Average session runs 5-7 turns.

Who runs it: The whole company as end users; owned by People Ops. Employees ask HR questions in Slack/Teams; the assistant grounds answers in Workday records and the policy library. Session depth compounds cost non-linearly.

What Firedog shows

Per-session cost curve showing how cost-per-turn climbs as employees ask follow-up questions.
How much of each session is session-history re-injection versus new answer generation.
Projected saving from sliding-window history compression plus incremental policy-index updates.

Decisions this informs

Compress session history into a rolling 2,000-token summary after turn 4; test answer quality on representative sessions before rollout.
Cache the 5,000-token HR persona and Workday field-map prefix; at 6 calls/run this is a clean recurring saving.
Switch the policy index to incremental re-embedding so only changed handbook sections are reprocessed nightly.

HR and benefits helpdesk chatbot

What Firedog shows

Decisions this informs

Understand every AI dollar.