Workforce analytics — per-user AI usage

The boss’s question is simple: which employee is using the AI, for what, and how much is it costing? DuDuClaw answers it from the same token-usage telemetry it already collects, now attributed per end-user and per channel.

Privacy boundary (read this first)

What this feature exposes is company AI-resource usage — the same thing a company already audits when it logs corporate email or VPN access. Two rules keep it from becoming surveillance:

• Aggregate first. The default views are totals per user (request count, token count, cost). Nobody’s individual messages show up in the usage report.
• Detail needs operator rights. Drilling into what a specific person asked is an operator-scoped action and is itself audited.

Any downstream feature that scores or flags individuals (question-quality coaching, private-use detection) is opt-in, off by default, and documented as “for a heads-up, never as grounds for discipline”.

Per-user cost attribution

Every channel reply now records the end-user id and channel alongside the agent and token counts. The token_usage table gained two additive columns (user_id, channel) via an idempotent migration — existing rows keep working, they just read as unattributed.

Query it:

> cost_users            # MCP tool, admin scope
{ "hours": 24 }

Returns each user ranked by cost:

[
  { "user_id": "u-alice", "total_requests": 42, "total_input_tokens": 130000,
    "total_output_tokens": 8000, "total_cost_millicents": 5400 },
  { "user_id": "(system)", "total_requests": 12, ... }
]

(system) buckets non-human traffic (sub-agent dispatch, evolution, utility calls) that has no end-user.

What’s attributed today

The channel-reply path (the surface most channels share) attributes user + channel. Paths without a human user — sub-agent dispatch, cron, evolution — record as (system) by design. Wiring the remaining per-channel media/tool sub-paths to the same attribution is an ongoing sweep; where a path is not yet wired, spend still lands on the agent, just not on a specific user.

Roadmap (opt-in, off by default)

• Question-quality coaching — a periodic batch scores a sample of each user’s messages (clarity, has-an-actionable-goal) with a cheap model and surfaces a “suggest training” list. Aggregate score only; no message text in the list.
• Usage-anomaly flag — a user whose spend deviates from their own baseline by N sigma is flagged, using the existing burn-rate anomaly math. Zero semantic judgement, so no false accusation risk.
• Private-use detection (semantic) — the false-positive guards are implemented (workforce_private.rs): the feature refuses to run without an operator-defined business-scope baseline (fail-closed), only flags high-confidence “suspected private” (never “undetermined”), honours an exempt list, and auto-expires unconfirmed flags after 30 days. Flags are advisory (“建議關注”), never employee-visible, and explicitly not grounds for discipline. The Haiku classification batch and the operator-only review UI sit on top of these guards.