AI usage tracking: how companies monitor AI spending across every team in 2026

AI usage tracking is the practice of recording every AI request a company makes — which model, which provider, which team, which feature, how many input and output tokens, and what it cost — and turning that record into a live view of AI spending that finance and engineering can both act on. Monitoring AI spending at a company is not the same problem as monitoring cloud spend. Cloud resources are provisioned; AI usage is emitted, request by request, by dozens of features and increasingly by autonomous agents. In 2026 the median mid-market company runs AI workloads across three or more providers, and roughly 40% of that spend cannot be attributed to a named owner on the first day of an audit. This guide is the operating manual for closing that gap.

Why AI spending is invisible by default

Provider consoles are organized around the billing account, not around your organization. OpenAI shows you tokens by API key and project; Anthropic shows you workspace totals; Google, Azure OpenAI, and AWS Bedrock each report in their own units, on their own cadence, with their own definition of a day. Nothing in that stack knows that key sk-live-7 belongs to the support copilot owned by the customer experience team, or that a 30% jump last Tuesday came from one enterprise tenant retrying a failed agent loop. Monitoring AI spending therefore always requires a layer above the providers — the moment you have two providers, the invoice stops being a report and becomes a puzzle.

The five signals every AI usage tracking system must capture

First, tokens split by direction: input, output, and cached input, because output typically costs 4–5× input and cached input costs 10–50% of standard input. A total token count that does not separate these three cannot explain a bill. Second, model and provider identity per request, including the exact model version — silent version upgrades change price and verbosity. Third, an owner dimension: team, product, feature, environment, and where applicable customer or tenant. Fourth, request outcome: success, error, timeout, retry. Failed and retried calls are billed and are one of the most common sources of unexplained spend. Fifth, latency and request shape, which is how you tell a genuine usage increase from a prompt that quietly grew by 4,000 tokens of context.

The three ways companies collect this data, and when each is right

Provider usage APIs are the source of truth for money: pull daily aggregates from each provider and reconcile them against the invoice. They are authoritative but coarse and lag by hours. A proxy or gateway in front of your model calls gives per-request truth in real time with no application changes, and is the fastest way to get tenant and feature attribution — this is how most companies get from 60% to 100% attributed spend. In-application instrumentation, where each call site logs its own metadata, gives the richest labels but requires discipline across every codebase. Mature setups run all three: instrumentation for labels, a gateway for real-time enforcement, and provider APIs for financial reconciliation. If you only do one thing this quarter, do the gateway.

Metrics that actually change decisions

Raw monthly spend is a headline, not a metric. Track cost per unit of business value — cost per resolved support conversation, per document processed, per active user, per closed ticket — because it is the only number that tells you whether growing spend is healthy. Track cost per team and per feature, weekly, with a named owner on each line. Track the share of traffic served by non-frontier models (healthy: 55–75%), cache hit rate on stable-prefix workloads (healthy: above 60%), share of non-interactive work running through batch endpoints (healthy: above 80%), and forecast accuracy against the actual invoice (healthy: within ±5%). Finally track the tail: the p99 cost per session and your single most expensive tenant, because that is where budget emergencies come from.

Allocation and chargeback: making AI spend land on a real owner

Attribution turns monitoring into accountability. The practical model is a three-level tag hierarchy applied at request time: organization → team → feature, with an optional tenant dimension for usage-based products. Shared infrastructure — embeddings, evaluation runs, internal tooling — goes into a clearly labelled shared pool that is either split by consumption or absorbed centrally, but never left unassigned; an unassigned pool always grows. Then choose your governance mode. Showback publishes what each team spent and is the right starting point, because it changes behaviour within two reporting cycles without any budget renegotiation. Chargeback moves the cost into each team's own budget and is appropriate once your attribution is reconciled to within a few percent of the invoice. Moving to chargeback on unreliable data destroys trust in the numbers, so earn accuracy first.

Budgets, alerts, and anomaly detection that people do not mute

Static thresholds fail in both directions: they fire constantly during growth and stay silent during a genuine 3× spike on a small feature. Use three layers. A monthly budget per team with alerts at 50%, 80%, and 100% of forecast pace gives finance predictability. Anomaly detection on the per-feature daily rate — comparing against a trailing 7- or 28-day baseline, and flagging deviations rather than absolutes — catches the runaway loop the same day. A hard circuit breaker per tenant and per feature converts worst-case AI spend from unbounded into a number you chose. Route alerts to the owning team, not to a central channel: an alert with no owner is a mute waiting to happen.

Shadow AI: the spending you are not tracking at all

Individually expensed AI subscriptions, personal API keys used in production scripts, and unsanctioned copilots typically account for 10–25% of real company AI spend and 100% of the governance risk, since prompts leaving through untracked channels are also data leaving through untracked channels. Find it with three sweeps: search expense reports and corporate cards for AI vendors, ask each provider for a full list of API keys and match every one to a named owner and workload, and inspect egress or SSO logs for AI domains. Then give people a sanctioned path that is faster than the workaround — a self-serve internal gateway key with a default budget attached — because prohibition alone reliably fails.

A 30-day rollout plan

Days 1–5: connect every provider account read-only, pull 90 days of history, and reconcile the totals against invoices so nobody can dispute the baseline. Days 6–12: introduce a gateway or middleware for the top three workloads by spend and stamp every request with team, feature, and environment. Days 13–18: publish the first weekly showback report with a named owner per line, and hold one 30-minute review. Days 19–24: set per-team budgets from the reconciled baseline plus expected growth, wire three-tier alerts to owning teams, and enable anomaly detection on per-feature daily rates. Days 25–30: run the shadow-AI sweep, add circuit breakers for the top tenants, and agree the one cost-per-value metric your leadership team will see monthly. Companies that complete this sequence typically identify 20–35% of spend they can remove immediately, mostly through over-sized models, retried requests, and forgotten features.

Common mistakes worth avoiding

Tracking spend without tracking quality, which turns cost reduction into invisible product regression — always pair a cost dashboard with an eval score. Treating cached and uncached input as one number, which hides the highest-leverage optimization you have. Reporting monthly when AI spend moves weekly. Building a beautiful dashboard nobody owns; a metric without a name attached to it does not change behaviour. And instrumenting only the primary chat feature while agents, background jobs, evaluations, and embeddings — often the majority of requests — go unmeasured.

Frequently asked questions

How do you monitor AI spending across multiple providers? Normalize each provider's usage into one schema — provider, model, input tokens, output tokens, cached tokens, timestamp, owner tags — then price it with a maintained rate card and reconcile against invoices weekly; this is the core of what SpendTensor does out of the box. What is the difference between AI usage tracking and AI cost management? Tracking is the measurement layer (what was used, by whom); management is the decision layer built on top (budgets, routing, caching, chargeback). You cannot do the second without the first. How accurate can attribution be? With gateway-level instrumentation, 95–100% of spend attributable to a team and feature is a realistic target; below 85% means your alerts and chargeback will be contested. How often should you review? Weekly for engineering owners, monthly for finance, plus real-time anomaly alerts for spikes.

The bottom line

Monitoring AI spending at a company is a measurement problem before it is a savings problem. Capture tokens by direction, attribute every request to an owner, reconcile to the invoice, publish it weekly, and put a ceiling on the tail. Companies that do this consistently pay 30–60% less per quality-adjusted AI request than companies that read their provider console at month end — not because they use AI less, but because every dollar is visible while it is still being spent.