AI Spend Visibility: How to Monitor and Control AI Usage Across Claude, ChatGPT, Copilot and Gemini

On 2 July 2026, Anthropic gave Claude Enterprise admins something every finance team has been asking of every AI vendor: a dashboard that shows what AI actually costs, per team and per user — with spend limits, alerts and an API to prove it. The timing is no accident. In the same quarter, GitHub Copilot moved to usage-based billing, Microsoft launched Copilot Cowork on metered credits, and Anthropic itself now prices heavy agentic work by consumption. The flat-rate era of workplace AI is ending — and AI spend visibility just became a core management discipline, not an IT afterthought. This guide shows you exactly where the usage and cost data lives in Claude, ChatGPT Enterprise, Microsoft Copilot and Google Gemini, how to build a cross-platform AI analytics dashboard on top of it, and how to train teams to use AI both more and more frugally.

By Toni Dos Santos, Co-Founder, Spicy Advisory

The flat-rate era of workplace AI is ending

For two years, budgeting for AI was easy: count seats, multiply by a monthly fee, done. That model is collapsing under the weight of agentic AI. A single autonomous coding session or a multi-step research agent can consume more tokens in an afternoon than a casual user burns in a month — and vendors have stopped absorbing the difference.

Look at what changed in just a few weeks of 2026:

• GitHub Copilot moved every plan to usage-based billing on 1 June 2026. Each plan now includes a monthly allotment of AI Credits (1 credit = $0.01), consumed by input, output and cached tokens at per-model rates. Copilot Business stays at $19/user/month — but that now buys $19 of metered credits, not unlimited use.
• Microsoft launched Copilot Cowork in June 2026 on usage-based Copilot Credits stacked on top of the $30/user/month Microsoft 365 Copilot license — agentic work is paid by consumption, not by seat.
• Anthropic meters Claude Enterprise usage beyond seat allotments as "extra usage", with spend reports that track exactly that overage — and its July 2026 admin update exists precisely because customers now need to manage a variable line item.
• The behavioural shift is documented: CNBC reported in late June 2026 that users of OpenAI and Anthropic tools are actively changing how they prompt and which models they pick to control costs — efficiency has become part of the user experience, not just the CFO's problem.

The consequence is simple and uncomfortable: your employees' daily habits now move a real invoice. Which model they pick, how they scope a prompt, whether they re-run a failed agent five times — all of it is metered somewhere. Companies that treat this as a procurement question will lose twice: once on waste, and once on adoption, because the natural reflex when costs are opaque is to ration AI rather than teach people to use it well. We wrote about that failure pattern in why AI adoption fails in companies — cost anxiety without visibility is one of its purest forms.

The answer is not less AI. It is visibility first, frugality second, adoption third — in that order, and this article covers all three.

What Claude's new admin analytics actually give you

Anthropic's 2 July 2026 release — "New analytics and cost controls are available for Claude Enterprise" — is the most complete spend-visibility package any of the four major workplace AI platforms ships today. It is worth understanding in detail, both because you may run Claude and because it sets the benchmark for what to demand from every other vendor.

Usage and cost, by team and by person

The admin analytics dashboard now shows usage and cost by group and by user, filtered by the SCIM groups your IT team already manages in your identity provider. Next to each cost line sits the output it bought: artifacts created, files edited, skills and connectors used. That pairing matters — a team whose costs doubled while output tripled is a success story, not a problem.

Ask the dashboard questions in plain language

Instead of exporting and pivoting, admins can ask the analytics interface things like "Which teams doubled their Claude usage this month?" or "Where are we getting the most value per seat?" and get charts back that can be exported and shared with stakeholders. If you have ever lost an afternoon reconciling license CSVs, you understand why this is the most quietly radical feature of the release.

Spend limits and alerts that prevent surprises

Admins can set spending limits at the organization level and per individual user. Spend-threshold alerts notify admins at 75% and 90% of an org-level limit — time to raise the cap before anyone is blocked mid-task. Users get their own in-app notifications at 75% and 95% and can request a limit increase from their admin without leaving Claude. This is the mechanism that turns a scary variable bill into a managed one.

Model defaults, so routine work doesn't run on the most expensive engine

Model defaults and entitlements let admins choose which Claude model new conversations start with — across chat, Cowork and Claude Code. Pointing routine drafting at a faster, cheaper model while reserving frontier models for complex work is the single highest-leverage frugality control on the platform, and it now takes one admin setting instead of a training memo.

An Analytics API for your existing FinOps stack

All usage and cost data is available programmatically through the Analytics API, so finance and IT can pull Claude spend into the tools they already run — Anthropic names Datadog Cloud Cost Management and CloudZero as integrations — and see AI spend alongside cloud spend. For engineering leaders, Claude Code analytics adds developer-specific metrics such as lines of code accepted and suggestion accept rate, under Admin settings > Claude Code.

A Compliance API for regulated teams

Enterprise organizations also get a Compliance API with real-time programmatic access to usage data and customer content, so compliance teams can build continuous monitoring and automated policy enforcement instead of quarterly sampling. If you operate under the EU AI Act, GDPR or sector rules, this is the audit trail your legal team will ask about.

Where to find it all: Analytics > Claude Chat for organization-wide usage, Admin settings > Claude Code for engineering metrics, and the Claude Help Center for the current field-by-field reference. Note one nuance: on seat-based Enterprise plans, spend reports cover extra usage (overage) — consumption inside seat allotments is covered by the seat fee. For the bigger picture of what Claude offers companies, see our complete guide to Claude for companies, and if your teams are hitting usage ceilings, our 18 tactics to stop burning Claude credits at work is the practical companion.

Platform by platform: how to monitor AI usage and spend on each tool

Claude set the pace, but your company almost certainly runs two, three or four platforms in parallel — and each one hides its analytics in a different place, with a different definition of "active user". Here is the working map, tool by tool — the table gives you the one-glance comparison, the sections below give you the click-paths.

ChatGPT Enterprise and Edu: Workspace analytics

Where it lives: Workspace settings > Workspace analytics, available to workspace admins.

What you get: seat allocation versus seats actually activated, activation rates, weekly active users and usage trends, plus aggregate metrics for messages, GPTs, tool usage and projects. It deliberately shows organization-level patterns, not content: admins cannot read individual prompts or conversations from analytics. OpenAI's own Enterprise user analytics guide walks admins through reading adoption and engagement from these views.

How to get the data out: on-demand CSV exports for Users, GPTs and Projects over a selected period. For raw, item-level records — legal hold, DLP, eDiscovery — the separate Compliance API is the right tool; analytics stays aggregated by design.

Spend levers: the activation-rate view is your money metric. Seats paid but never activated, or activated and dormant for 60 days, are the first line of recoverable spend. Review it monthly and recycle seats before renewing. If you are still weighing platforms against each other, our comparison of ChatGPT Enterprise vs Copilot vs Gemini covers the trade-offs, and our ChatGPT Enterprise training gets teams past the activation plateau.

Microsoft 365 Copilot: Copilot Analytics and the Copilot Dashboard

Where it lives: two levels. The Microsoft 365 admin center ships a Copilot usage report (enabled vs active users, last activity per app). The richer surface is the Copilot Dashboard in Viva Insights, part of what Microsoft brands Copilot Analytics.

What you get: metrics grouped into readiness, adoption, impact and sentiment — licensed employees, active Copilot users over a trailing 28-day window, adoption by app (Teams, Word, Excel, PowerPoint, Outlook, Chat), adoption by organizational attribute such as function or department, and usage intensity and retention (how often users return, average weekly actions). For deeper analysis, the Microsoft 365 Copilot adoption report template in Viva Insights' analyst workbench segments users by frequency and consistency of use — the closest thing Microsoft offers to a power-user/dormant-user census.

How to get the data out: dashboard exports and Viva Insights queries feed Power BI; the admin-center report exports to CSV. Mind the latency: the dashboard refreshes daily but reflects the previous 28-day period with up to six days' delay.

Spend levers: at $30/user/month, a Copilot license that shows no activity for 28 days is $360/year of pure waste — the adoption-by-app view tells you exactly where to intervene with training or reclaim licenses. And with Copilot Cowork and GitHub Copilot both metered in credits, set the budgets and alerts in your Microsoft and GitHub billing consoles before agents go live, not after the first surprising invoice. Our guides to Microsoft Copilot Cowork and Copilot workflows in Excel and PowerPoint pair well with Copilot training here.

Google Gemini and Workspace: Gemini reports in the Admin console

Where it lives: Admin console > Menu > Generative AI > Gemini reports, documented in Google's "Review Gemini usage in your organization".

What you get: organization-level and user-level views — how many people actively use Gemini, what percentage of eligible licensed users that represents, adoption of Gemini features app by app (Gmail, Docs, Sheets, Meet), and identification of power users. Since a February 2026 update, admins also see threshold reports: how many users have hit their AI feature limits — an early-warning signal for both frustrated users and future upsell pressure. Reports can be filtered by organizational unit or group, with a 2–3 day data lag.

How to get the data out: three routes of increasing power. Export report views to Sheets or CSV; query Gemini for Workspace log events via the Audit and investigation tool or the Reporting API (Admin SDK); or enable the full pipeline under Reporting > Data integrations to export logs to BigQuery (requires a Google Cloud project with billing) and build whatever you want in Looker Studio on top.

Spend levers: since Gemini is bundled into Workspace Business and Enterprise editions, the waste pattern is inverted — you have already paid for AI most teams don't know they have. Here the dashboard's job is to find under-use and fix it with enablement; our guide to Gemini workflows across Gmail, Docs and Sheets and our Gemini for Workspace training exist for exactly that. The threshold reports meanwhile tell you when power users genuinely need a higher tier.

Claude (Anthropic): the new benchmark

Where it lives: Analytics > Claude Chat and Admin settings > Claude Code, plus the Analytics API and Compliance API — all covered in detail above. In one line: per-group and per-user cost next to output, plain-language analytics chat, org- and user-level spend caps with staged alerts, model defaults, and programmatic export to FinOps tools. Use it as your reference when evaluating what the other three give you.

Building a cross-platform AI analytics dashboard

Four consoles, four definitions of "active user", four billing models — and a CFO who wants one number. Closing that gap is less a tooling problem than a discipline problem. In our client work we see three maturity levels, and the mistake is trying to jump straight to level three.

Level 1 — the monthly console ritual (start this week). One owner (IT ops or the AI lead) opens all native dashboards on the first business day of each month and fills a single spreadsheet with identically defined KPIs: seats purchased, seats active in the last 28 days, activation rate, cost per active user, percentage of credit/usage allotment consumed, and overage spend. Twenty minutes per platform. The spreadsheet is ugly and it works — most companies discover 15–30% of paid seats are dormant on the first pass.

Level 2 — automated exports into one view. Wire the machine-readable routes: Claude's Analytics API, ChatGPT Enterprise CSV exports, Copilot Dashboard/Graph exports into Power BI, and Gemini logs into BigQuery feeding Looker Studio. Two prerequisites make or break this level: a single department taxonomy (the same team names everywhere), and SCIM groups kept clean in your identity provider, because Claude's per-group costing and Viva's org attributes are only as good as the groups you feed them. You can even prototype the visual layer with AI itself — we showed how in building live dashboards with Claude artifacts.

Level 3 — FinOps integration. Pipe everything into the cost tooling your infrastructure teams already trust — Datadog Cloud Cost Management and CloudZero both ingest Claude's Analytics API today — and manage AI like any other cloud line: unit economics (cost per active user per week, cost per completed workflow), budgets with alerts, and a quarterly license-rebalancing review where dormant seats become credits for the teams generating measurable value. This is also where AI spend meets ROI measurement; our CFO's guide to measuring AI ROI covers the value side of the same equation.

AI frugality: the discipline nobody trained your teams for

Here is the part most companies get backwards. When the AI bill becomes variable, the instinct is to restrict: lower limits, fewer seats, approval workflows. That reflex kills adoption — and adoption, not thrift, is where the return on AI lives. The companies that win under token-based billing do something different: they make costs visible and then teach frugality as a skill, the way manufacturing taught lean.

Frugal AI use is not using AI less. It is maximizing value per token:

• Right model, right task. Routine summarization does not need a frontier model. Claude's new model defaults enforce this org-wide; on other platforms it is a habit to train. This one change typically cuts cost per task by half or more without touching output quality.
• Scoped prompts and context hygiene. Vague prompts produce long, wrong answers that get re-run three times. Teaching people to state the task, the format and the constraints up front is a cost measure disguised as a quality measure.
• Reuse beats re-prompting. Skills, projects, custom GPTs and shared prompt libraries mean the organization pays once for good instructions instead of paying every employee to rediscover them. Our piece on 18 tactics to stop burning Claude credits is a full playbook of these habits.
• Know when not to use AI. A template, a saved search or a five-line script is free forever. The most expensive query is the one that didn't need a model at all.
• Kill duplicate and shadow spend. Personal ChatGPT Plus subscriptions expensed alongside an enterprise contract, or unsanctioned tools processing company data, are both a cost leak and a governance risk — our guide to shadow AI covers how to surface them.

Employees are now on the front line of this whether you prepare them or not: Claude users literally receive in-app alerts at 75% and 95% of their own spend limit. An employee who understands what drives those numbers adjusts their model choice and prompting; one who doesn't just stops using AI on the 25th of the month — the worst possible outcome for everyone. In our training rooms across 50+ companies, the pattern is consistent: teams trained on cost-aware usage increase their AI activity while their cost per task falls. Frugality and adoption are the same curriculum. That's also the argument for treating this as an enablement project, not a policing project — a point we develop in AI adoption best practices and operationalize in a right-sized governance framework.

Your 30-day AI spend visibility plan

1. Week 1 — inventory the meter. List every AI tool the company pays for, its billing model (seat, credits, overage), renewal date and owner. Include the credit allotments: what does $19 of GitHub AI Credits or a Claude seat allotment actually cover for your usage pattern?
2. Week 1 — switch on the native analytics and the guardrails. Claude: set org- and user-level spend limits and confirm the 75%/90% alerts reach a mailbox someone reads. GitHub/Microsoft: set credit budgets before agents scale. Google: check the threshold reports. OpenAI: pull your first Users export.
3. Week 2 — define shared KPIs and one taxonomy. Active user (28-day), activation rate, cost per active user, % allotment consumed, overage. Align department names and SCIM groups across platforms.
4. Week 3 — stand up the Level-1 dashboard and hold the first review. Thirty minutes with finance, IT and one leader per department. Decisions, not admiration: reclaim dormant seats, set two model defaults, pick one team for deeper enablement.
5. Week 4 — train the habits and re-measure. Run a frugal-usage session with the highest-spend team, publish a one-page cost-aware AI playbook, and compare week-4 numbers to week-1. Expect the counterintuitive result: activity up, cost per task down.

Where Spicy Advisory fits

Spend visibility, frugal usage and adoption are one project, and it sits exactly where we work. Spicy Advisory is a founder-led AI advisory and training boutique — 1,500+ professionals trained across 50+ companies including L'Oréal, EssilorLuxottica and IGN, rated 4.98/5, in English and French. For AI spend, we work in three steps: an AI audit that maps your real usage, spend and adoption gaps across Claude, ChatGPT, Copilot and Gemini; a dashboard sprint that stands up the cross-platform view with your IT and finance teams; and role-based training that turns cost-aware, high-output AI use into a daily habit rather than a memo.

Frequently asked questions