Quick Answer
AI cost visibility is the ability to see what AI spend is buying, by person, team, feature, customer and activity, before the bill arrives. The invoice can't provide it: it records tokens and hours, not decisions. It comes from capturing each AI event as it occurs, with business context attached, and reconciling to the bill.
Take 50 million input tokens and 10 million output tokens in a month. On Claude Haiku 5.5 that is $10. On Claude Fable 5.1, the same tokens cost $1,000. On OpenAI, GPT-6 Luna at standard rates runs the same month for $10, and GPT-6 Astra in Ultrafast mode runs it for $6,000. Four scenarios, one token count, a 600x spread. Every one of them reaches the invoice as “tokens.”
Most of the advice on this topic starts from that invoice and works backward: export it, tag it, split it by model and provider, build a dashboard.
We think that’s the wrong direction. A model is not something a CFO can fund or stop. The work the model did is. So the unit of AI cost visibility that matters is the business activity behind each call, seen while the call is still being made. The bill, with its model names, arrives 30 days later.
That’s the position CloudZero, the AI ROI company, built AI Signals around, and the one we run our own finance function on. Dan Carducci, our VP of Finance, said at the September 23, 2026 launch: “AI was the most evasive line in my forecast.”
He could report the total, but couldn’t say what drove it. The report itself didn’t change; what changed was the context behind the spend and which team drove it. This enables finance leaders like Dan to forecast the line rather than guess at it.
Five numbers to carry through what follows:
- 600x: the cost spread between the cheapest and most expensive way to run the same 60 million tokens in a month, on published OpenAI rates.
- $8,023: one week of an idle eight-H100 reservation on AWS, billed as compute.
- $2.67 trillion: Gartner’s 2026 worldwide AI spending forecast, up 49.5% in a year.
- 20%: organizations with a breach tied to shadow AI in IBM’s 2025 report; where shadow AI use was high, breaches cost $670,000 more.
- 35+: the business activities AI Signals classifies each AI call into, as it happens.
The survey-series companion, why the office that owns the AI budget sees least, covers what finance leaders told us in our 2026 survey about where visibility breaks.
What is AI cost visibility?
AI cost visibility is the ability to see who is spending on AI, on what work, through which models, and what it produced, before the invoice arrives. If you can answer those four questions in the week they happen, you have it.
OpenAI, Anthropic and AWS each give you a dashboard of provider totals for free, and none of them can tell you that the support team moved ticket summarization to a larger model on Tuesday. The AI cost management hub covers the budgeting and governance that sit on top; visibility is the layer underneath, and it has to cover every object AI spend arrives as.
| Cost object | Where it’s billed | What the bill records | What visibility requires |
|---|---|---|---|
| Model tokens (API) | OpenAI, Anthropic, Google, Bedrock, Vertex | Model, input and output token counts, tier, cached share | Which app, feature, customer session, person and activity generated the call |
| GPU hours | AWS, Azure, GCP, specialist GPU clouds, on-prem | Instance type, hours, reservation or on-demand | Which workload ran, at what utilization, for which team, and whether it ran at all |
| Per-seat AI tools | Cursor, Claude Team and Enterprise, ChatGPT Enterprise, Copilot, Notion AI | Seats, plan, overage credits | Who holds the seat, whether it is used, and what work it supports |
| Agent runs | Agent platforms, usage-based SaaS | Credits, tasks, runs | Which business process the agent serves and its AI agent cost per completed task |
| Embedded AI in SaaS | CRM, support, analytics vendors | A bundled subscription line | Which features are AI-metered and how much of the renewal is AI |
| Shadow AI | Expense reports, personal cards, unmanaged keys | Nothing, or a credit-card line | That it exists, who is using it, and what data it touches |
Every row has the same shape. The bill is a record of consumption. Visibility needs the record of intent, which lives somewhere else: application telemetry, gateway logs, identity systems, and the tags engineering applies (or doesn’t). The AI statistics cornerstone has the market-level numbers for each row.
2026 State of AI Spend Report
Why Finance Caps What It Can’t See
Learn how finance leaders are managing AI spend, and how they’re approaching FY2027
Why is AI spend so hard to see?
AI spend is hard to see because the decisions that move the number most are made at the point of use, in code and consoles, and the invoice flattens all of them into a quantity. Five of those decisions do most of the damage. The first four are verifiable on the vendor’s own rate card; the fifth never reaches a rate card at all.
1. Model choice. Anthropic’s rate card lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output. Claude Fable 5.1 is $10 and $50. A developer picking the bigger model for a routine summarization job multiplies that job’s cost by 100 and changes nothing on the invoice except the number.
2. Speed tier. OpenAI’s pricing lists GPT-6 Astra at $10 input and $50 output at standard speed, $20 and $100 in Fast mode (renamed from Priority on July 30, 2026), and $60 and $300 in Ultrafast. Same model. The tier is a flag on the API call, and finance never sees the flag.
3. Prompt design. Cached input (context the application reuses instead of resending) is $1 per million on Astra against $10 uncached, a 10x difference. On Claude Haiku 5.5, prompts over 100,000 tokens are billed at five times the shorter-prompt rate. A prompt template change can quintuple a cost line overnight. The token-based pricing explainer covers how each lever is metered, and the LLM API pricing comparison lays the vendors side by side.
4. Idle capacity. AWS Capacity Blocks price a p5.48xlarge (eight H100s) at $47.757 an instance-hour. A block reserved for a training run that slipped a week costs $8,023 for doing nothing, and the bill shows a compute line that reconciles perfectly. On the newer p6-b200 at $113.666 an instance-hour ($14.208 per GPU-hour), the same idle week costs over $19,000. The H100 GPU cost guide shows where those rates sit against rental and on-prem.
5. Seats and shadow use. Per-seat tools bill whether the seat is used or not, and the AI that never reached procurement bills on an expense report or not at all. IBM’s 2025 Cost of a Data Breach report found 20% of organizations had a breach tied to shadow AI, with $670,000 in added breach cost where shadow AI was high.
The cost side of shadow AI is AI sprawl, and it never shows up as a line called “AI.”
All five are choices, and a total cannot tell you which of them were worth making. If you want those five run against your own last 90 days before you commit engineering time to any of this, CloudZero runs that pass on your bills and hands back the ranked layers.
What should you be able to see?
You should be able to see AI spend by the work it paid for, with the person and team attached, tied back to the invoice, in the week it happened.
Provider dashboards skip this part, because the model-level view is what every provider exports and the activity-level view has to be built. Here is the gap, in the questions a CFO asks.
| The question finance asks | What a model-level view says | What an activity-level view says |
|---|---|---|
| Why is AI spend up 18% this month? | Claude Sonnet usage rose | Back-end engineering’s Claude Sonnet spend up 3.4x week over week: four engineers, mostly test generation |
| Which teams are driving it? | Three API keys and two projects | Engineering 61%, sales 24%, marketing 15%, with the people behind each named |
| Is marketing’s AI spend working? | Marketing used $900 of GPT-5 | Marketing spent $900 on GPT-5 this week on content drafting, six times its trailing average |
| What should I fund more of? | The cheapest model | The activities with measured output, and the people doing them |
| What should I stop? | Nothing identifiable | The activity whose cost rose without any output rising |
CloudZero’s AI Signals classifies every AI event into one of 35+ business activities as it lands, so sales spend reads as prospecting or account research, engineering spend as code review or debugging, and marketing spend as drafting or campaign analysis.
Scott Castle, CloudZero’s Chief Product Officer, drew the line at launch: “Collecting AI cost data is table stakes.” The hard part is attaching the team and the work while the spend is still live, and that is what makes the number defensible. That is why AI ROI is an allocation problem before it is a measurement problem.
The activity lens is also why we treat AI spend visibility as an AI ROI question. A model tells you nothing about return. An activity with a cost and a measured output is a cost per outcome, and that’s a number a board can act on.
Is AI cost observability the same as AI cost visibility?
AI cost observability and AI cost visibility are the same goal described by different teams, and the differences are about who is asking and how late.
Engineering says AI cost observability, or LLM cost visibility when the scope is the model layer alone. Operations says monitoring. Procurement says tracking. Finance says visibility.
| Term | Who uses it | What they expect | What it misses alone |
|---|---|---|---|
| AI cost observability | Engineering, platform | Per-request cost in the telemetry stream, next to latency and errors | Business dimensions. Cost per call without the customer or margin impact |
| AI cost monitoring | Operations, finance ops | Thresholds, alerts, anomaly detection on totals | Cause. “Spend is up 40%” starts an investigation instead of ending one |
| AI cost tracking | Procurement, controllers | A ledger of seats, credits, commitments and renewals | Usage. A tracked seat says nothing about whether it is used |
| AI spend visibility | CFO, FP&A, business leaders | One view by team, feature, customer and activity, tied to the bill, in time to act | Nothing, provided the dimensions are attached at capture |
Build for the finance definition and the other three fall out of it: engineering filters the same events by model, operations sets thresholds on the dimensions finance budgets against, procurement rolls the events up to the vendor. Build for any one of the others and you will be adding the business dimension later, by hand, every month.
Cost is also the pillar most AI observability programs skip, which is why engineering-led telemetry so often ends at latency and never reaches the ledger.
How does real-time AI cost visibility work?
Real-time AI cost visibility works by capturing each AI event at the moment it happens, rather than the bill at the end of the month, and attaching to the event the context billing cannot carry. The event becomes the unit of record, and the invoice becomes the thing you reconcile to at close.
A monthly AI invoice is a summary of millions of events that already happened. Each event, when it occurred, knew things the invoice forgets: the application, the feature, the customer session, the user, the agent, the prompt template, the cache state. That context lives in telemetry and gateway logs. Capture it when the event lands and you keep it. Wait for the bill and it’s gone, and you’re back to tagging after the fact.
CloudZero captures those events from three places: macOS machines (for Claude Code and other local AI tools), LLM gateways (the proxy every API call passes through, so it sees every request, with no application changes), and OpenTelemetry streams (the standard way applications emit their own event data, including feature and customer context). Supported inference providers at launch included Anthropic, OpenAI, Google Vertex and Gemini, AWS Bedrock and Fireworks, with LiteLLM and Bifrost as gateways.
Per the launch release, no prompt or session data leaves the customer’s own cloud environment. The AI Signals documentation has the integration detail.
Cadence matters as much as capture, because different cost questions need different speeds. A spike needs an answer today; a margin review needs the complete picture. The AI Signals platform describes three speeds for this reason: a livestream of individual events as they occur, working answers within minutes for ad-hoc questions, and full allocation across every dimension within hours, in time for close.
That one allocated number is then delivered where decisions get made: the documented surfaces include Slack, Jira, executive dashboards, Claude Code and any MCP-compatible tool.
This is where the two approaches to AI cost monitoring diverge. The detect-then-investigate model, which most cost tools and their new agent features follow, starts from a finding after a total has broken a rule, then investigates its cause and owner.
The see-it-as-it-happens model attaches the cause at capture time, so the alert arrives as a finding with the team and the work already named. Both approaches beat a monthly surprise. Only the second lets the person accountable see the pattern break while it is still small.
A free CloudZero demo walks through the platform against sample data, if you’d rather see the livestream than read about it.
Which metrics does AI cost visibility produce?
AI cost visibility produces nine numbers a finance team can put on a page, and the first one tells you whether the other eight can be trusted. None of them exist at bill level. All of them fall out of event-level capture with business context attached, and together they are what AI cost reporting to a board should look like.
| Metric | What it answers | Where it comes from | Healthy direction |
|---|---|---|---|
| Unallocated share of AI spend | How much of the total has no owner | Spend with no team, person or feature attached ÷ total | Falling toward single digits |
| Spend by business activity | What the money paid for | Classification of each event (prospecting, code review, drafting) | Shifting toward activities with measured output |
| Cost per request or query | What one unit of AI work costs | Event cost ÷ requests, by feature and model | Flat or falling as volume grows (see AI cost per query) |
| Cost per customer or tenant | Whether any customer is unprofitable on AI alone | Event cost by session or customer account | No customer whose AI spend exceeds its contribution margin (see AI cost per customer) |
| Cost per outcome | What a completed task, ticket or deal costs in AI | Event cost ÷ completed outcomes, with failures in the numerator | Falling as prompts and models are tuned |
| Cached-token share | How much input cost is being avoided | Cached input tokens ÷ total input tokens | Rising; a drop flags a prompt or template change |
| Model and tier mix | Whether spend growth is volume or upgrades | Spend by model and speed tier over time | Stable, or shifting down-tier where quality allows |
| GPU utilization against reservation | How much reserved capacity is idle | Utilization metrics next to reserved hours | Idle hours trending to zero |
| Seat utilization | Whether per-seat AI is being used | Active users ÷ paid seats, from the vendor admin API and SSO | Rising, or seats reclaimed |
Two of these measure the visibility itself: the unallocated share (how much spend has no owner) and the lag between an AI event and its appearance in reporting, which should be measured in hours.
The unallocated share is the one to publish first. A 40% unallocated line tells the board the other metrics describe 60% of the spend. Getting it under 10% is the first milestone, and the 30-day plan that follows is built to move it. Once it is there, the AI ROI measurement guide covers connecting these numbers to outcomes.
How do you get AI cost visibility in 30 days?
You can get from bill-level to resource-level AI cost visibility in 30 days with the exports you already have, and to activity-level with event capture, if you sequence the work by dollar weight rather than by what is easiest to tag. The plan assumes a finance owner and one platform engineer, part time, and every week is measured by one number: the unallocated share.
Week 1: inventory where AI spend arrives. Pull every AI charge from cloud bills, provider invoices, SaaS renewals and expense reports for the last 90 days. Sort it into the six cost objects: tokens, GPU hours, seats, agent runs, embedded AI, shadow use. Rank by dollars. The ranking decides the order of everything that follows.
Week 2: instrument the heaviest layer. For most teams that is the API layer: put a gateway in front of the heaviest call paths (the LLM gateway comparison covers the main options and what each costs) and turn on provider usage reporting, which OpenAI, Anthropic and Bedrock each expose. For a GPU-heavy shop it is utilization metrics next to every reservation.
Week 3: map keys, clusters and seats to people and teams. Build the lookup tables that connect each API key and each cluster to an owner and a team, from your identity provider and repo ownership. Reconcile seat lists against single sign-on and reclaim the unused ones. This is the week the unallocated share first drops, and the week AI spend tracking stops being a spreadsheet.
Week 4: attach the business context and reconcile to the bill. Add feature and customer attributes on the heaviest application paths, classify activities (by hand from key and app ownership if you’re not on a tool that does it), and run the first reconciliation of event-level cost against the invoice. Then set the first monitors on the dimensions finance budgets against, not on the totals.
Two places teams stall. The first is waiting for perfect tagging before publishing anything; publish whatever share is allocated and label the rest, so its size is visible and shrinking. The second is building the dashboard before the capture; an AI cost dashboard on top of monthly exports will tell you the same thing the bill did, a month late.
What should an AI cost visibility tool do?
An AI cost visibility tool earns the label if it answers “who, on what, for what, and what did we get” from live events, reconciles that answer to the invoice, and puts it in front of the person who can act. Many products sold as AI cost visibility tools do one of those three. Here is the checklist we would hand a CFO evaluating the category, in the order it matters.
- Captures events at the source. Ingests from gateways, desktop agents and telemetry, so model, tier, cache state, person and feature are known at the moment of spend. A tool that starts from billing exports is a day or more behind and carries none of the business context.
- Classifies by business activity. Turns a “gpt-6-astra” call into “code review” or “prospecting.” Without this, every report is a model report.
- Attributes to people and teams without tags. API tokens, seats and agent runs do not carry resource tags. Attribution has to come from identity, repo ownership and gateway metadata.
- Covers all six cost objects. Tokens, GPU hours, seats, agent runs, embedded AI and shadow use. A tool that stops at API tokens is reporting the easiest fraction.
- Reconciles to the invoice at close. Real-time figures are estimates, so the allocated total has to tie to the bill at month end. If the tool’s total and the bill’s total disagree and stay that way, nobody will trust either.
- Alerts with the cause attached. “Marketing’s GPT-5 spend is six times its trailing average, on content drafting” is a finding. “Spend is up” is a ticket.
- Keeps prompt and session data inside your environment. Visibility should not require exporting what your people typed.
- Reaches the decision wherever it is made. The coding tool, Slack, the board deck, the production system.
AI Signals was built around the first three of those: events first, activity second, attribution third. The AI cost management tools listicle compares the category, and if you are weighing assembling this yourself from gateway logs and a warehouse, the build vs buy framework covers that math.
How does AI cost visibility fit with allocation and AI cost management?
AI cost visibility supplies the inputs, allocation assigns them, and management is the program that runs on both; keeping the three straight is what stops a visibility project from turning into a tagging project.
AI cost allocation is the methodology for assigning the spend you can see to teams, products, features and customers, usually with shared-cost rules. The allocation guide covers it in depth; visibility’s job is to make sure allocation has events to work with rather than invoice lines.
Mature cloud cost visibility is a head start on the process but not on the data. Tokens, seats and agent runs carry none of the tags cloud allocation relies on, which is why teams with solid cloud unit economics find themselves starting over on AI.
AI cost management is the budgeting, forecasting and governance layer, and it is where Gartner’s numbers land: worldwide AI spending of $2.67 trillion in 2026, up 49.5%, with infrastructure at $1.48 trillion and generative AI model spend at $28.3 billion, up 117%. John-David Lovelock, Distinguished VP Analyst at Gartner, called the buildout the “largest infrastructure project humanity has ever undertaken.”
Managing a line growing 49.5% a year without seeing inside it is the situation most finance teams are in, and the one the AI cost optimization playbook assumes you have already fixed.
What should you do first?
Count how long it takes your team to answer “why did AI spend move last week?” If the answer is “after the bill arrives,” the five rate-card decisions (model, tier, prompt design, idle capacity, seats) are being made without you, and the 30-day plan is where to start. If the answer is “in minutes, by activity, with names,” you are already running on the signal, and the question becomes what to fund more of.
Frequently asked questions about AI cost visibility
