How to Track AI Costs Per Employee

The Attribution Problem

Ask most CFOs what their company spends on AI and they'll give you a seat count: "we have 80 ChatGPT licenses at $25 each." That number is wrong in both directions. Half those seats are dormant, and a handful of power users are generating API-level consumption that dwarfs the seat fees. Without per-employee metering, you can't tell a $25/month employee from a $400/month one.

The three places spend hides: vendor invoices (delayed, aggregated),expensed seats (flat fees, no usage data), and embedded AIinside other SaaS tools (invisible entirely). Solving attribution means intercepting usage, not reconciling bills.

Option 1: Reconcile Invoices (Baseline)

Pull vendor invoices and expense reports monthly. This is the floor — it catches license costs and flags zombie subscriptions, but gives you nothing on actual consumption or per-person usage. Fine for cleanup, useless for control.

Option 2: Provider Admin Consoles (Single-Vendor Only)

OpenAI, Anthropic, and Google all offer workspace dashboards with per-user views. If your org truly standardizes on one vendor, this works — but it leaves blind spots for every other tool, and vendors have no incentive to help you spend less on their own models.

Option 3: A Metering Gateway (The Real Answer)

Route every AI request through a gateway that issues each employee their own virtual API key. The gateway authenticates the request, meters tokens, applies policy, and forwards to the provider. Employees keep their preferred tools; you get complete attribution regardless of which model or provider they use.

1. Issue per-employee keys. Each employee gets a virtual key (e.g., sf-emp-*) mapped to their identity. Revocation is instant and per-person — employees never see real provider credentials.
2. Swap the base URL. Any OpenAI-compatible client works — point it at your gateway URL and use the virtual key. Same for a managed chat UI for non-technical staff.
3. Meter every request. Log model, tokens in/out, computed cost, and latency per request. Never store prompt bodies by default — counts are enough for spend.
4. Aggregate to rollups. Pre-compute hourly/daily totals per employee, team, and model so dashboards stay fast.
5. Attach budgets. Per-employee and per-team allowances turn visibility into control — see the fallback strategy for how to enforce without blocking.

What the Data Unlocks

• Chargeback/showback: real per-team AI costs for finance
• Anomaly detection: a sales rep burning $50/day on Opus is a conversation, not a mystery
• Right-sizing: most work doesn't need premium models — the data proves it
• Benchmarking: cost per employee against real signed contracts from comparable companies
• Vendor leverage: negotiate enterprise agreements with actual usage data instead of guesses

Common Mistakes

• Tracking only API usage: consumer chat subscriptions are the bigger blind spot — use a managed UI or browser extension to capture them
• Monthly aggregation only: a one-day spike looks identical to steady usage; keep request-level events for 90 days
• Blocking on exhaustion: hard stops train employees to route around you with personal accounts. Fall back to cheaper models instead
• Ignoring tier mix: per-model breakdown matters as much as totals — one team on Opus can outspend twenty on Haiku

SpendFriend implements all of this out of the box — virtual keys, metering, rollups, and budgets with fallback. Compare it against alternatives in our AI spend management software comparison.