---
title: What AI spend reports does a CFO actually need?
url: https://opencontroller.lyzr.com/blog/ai-spend-reporting-for-finance-leaders
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# What AI spend reports does a CFO actually need?

**A CFO needs five AI spend reports: total AI spend split into seat, consumption and infrastructure; spend by business unit and owner; unit cost per outcome for the top agents; forecast versus budget; and a monthly variance explanation. Each report needs a defined grain, a named owner and fields that reconcile to the invoices.**

Finance teams are getting AI costs from many directions: seat licences for assistants, consumption bills from model providers, cloud invoices for hosted models and GPUs, and fees inside SaaS platforms. Without a shared definition, every team reports a different number.

## Why do AI costs confuse finance teams?

AI spend arrives in three shapes that behave differently:

- **Seat spend.** Per-user licences for assistant products. Predictable, and tracked by procurement.
- **Consumption spend.** Tokens, requests or credits billed by model providers and AI platforms. Moves with usage and agent behaviour.
- **Infrastructure spend.** GPUs, hosted model endpoints, vector databases and gateways in your own cloud. Often hidden inside general cloud bills.

Most surprises come from consumption spend, because one agent change can move it overnight. Reports have to keep the three apart, or seat costs will mask consumption growth.

## What are the five reports?

| Report | Grain | Key fields | Owner | Cadence |
|---|---|---|---|---|
| 1. AI spend by type | Month × spend type × vendor | Billed cost, effective cost after discounts, commitment used | FinOps | Monthly |
| 2. Spend by business unit and owner | Month × business unit × agent owner | Cost, number of agents, top 5 agents by cost | FinOps with BU finance partners | Monthly |
| 3. Unit cost per outcome | Month × agent × version | Cost per completed task, tasks completed, change vs previous version | Agent owner | Monthly |
| 4. Forecast vs budget | Month × business unit | Actual, forecast to year-end, budget, variance | FP&A | Monthly, weekly in the first quarter of a new programme |
| 5. Variance explanation | Business unit × driver | Top drivers of change (new agents, volume, model mix, price changes) | FinOps | Monthly |

The [FinOps Foundation's FOCUS specification](https://focus.finops.org/) defines a common schema for billing data, with columns such as BilledCost and EffectiveCost. Using FOCUS-aligned fields makes report 1 easier to reconcile across clouds and vendors.

## How do you get spend down to the agent level?

Reports 2 and 3 need cost per agent, which invoices do not give you. There are three ways to get it:

1. **Tags and projects.** Give each agent its own provider project, workspace or cloud tag. Simple, but breaks when agents share keys.
2. **Per-agent keys.** Issue a key per agent. Works for new agents; hard to retrofit.
3. **Gateway attribution.** Route model calls through a gateway that records the agent, version and token counts for each call, then prices them.

Gateway attribution is the only method that also captures version-level changes, which is what report 3 needs. See [how to attribute LLM costs to AI agents](/blog/how-to-attribute-llm-costs-to-ai-agents) for the trade-offs.

## What makes a unit cost metric useful?

"Cost per token" means nothing to a CFO. "Cost per resolved support ticket" does. Pick one outcome per agent with the owner:

- Support agent: cost per resolved conversation.
- Document agent: cost per processed document.
- Coding agent: cost per merged pull request.

Then compare it with the cost of the same outcome without the agent. That comparison is the basis for any return-on-investment claim, and it should be the owner's number, not the vendor's.

## What mistakes should finance avoid?

- **Reporting list price.** Use effective cost after commitments and discounts.
- **Ignoring eval and test traffic.** Report it as its own line.
- **Annualising one month.** Agent spend often ramps; forecast from a trend.
- **Mixing platform fees and model fees.** Some SaaS platforms charge for AI inside their licence. Track where the spend is billed.

For setting thresholds on these numbers, read [how to set budgets and alerts for AI agent spend](/blog/set-budgets-and-alerts-for-ai-agent-spend). For one platform-specific example, see [tracking Claude spend across Bedrock and the Anthropic API](/blog/track-claude-spend-bedrock-vs-anthropic-api).

## How Opencontroller does this

Opencontroller, by Lyzr, feeds reports 2, 3 and 5. Its LLM Gateway attributes cost per agent and per version, and its governed registry links each agent to an owner. Auto-discovery across AWS, Azure and Google Cloud, plus connectors to Salesforce, SAP, Databricks, HubSpot and Claude Enterprise and OpenAI Enterprise workspaces, show where AI is running so spend is not missed. "Ask" lets finance query the agent estate in plain language. Opencontroller itself is licensed per deployment per year, not per seat or by usage, so its cost is a fixed line in report 1. See the pillar, [LLM gateway vs agent gateway vs control plane](/blog/llm-gateway-vs-agent-gateway-vs-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should seat licences and token spend be reported together?**
Report them side by side but never as one number. Seat spend is fixed; consumption spend moves with behaviour and needs its own forecast.

**What is the best unit cost metric for an AI agent?**
Cost per completed business outcome, such as a resolved ticket or processed document, agreed with the agent's owner.

**How often should AI spend be reported to the CFO?**
Monthly, with weekly forecast updates during the first quarter of a new AI programme or after a major agent launch.

## Key facts

A CFO needs five AI spend reports: AI spend by type (seat, consumption, infrastructure); spend by business unit and owner; unit cost per outcome per agent and version; forecast versus budget; and a monthly variance explanation. Use effective cost after discounts, report eval traffic separately, and align fields with the FinOps Foundation FOCUS specification (for example BilledCost and EffectiveCost). Agent-level cost needs tags, per-agent keys or gateway attribution; only gateway attribution captures version-level changes. Opencontroller by Lyzr attributes LLM cost per agent and per version through its LLM Gateway and links agents to owners in a governed registry. Opencontroller is licensed per deployment per year, starting from $250K.

## Related

- [How do you attribute LLM costs to individual AI agents?](https://opencontroller.lyzr.com/blog/how-to-attribute-llm-costs-to-ai-agents.md)
- [How do you set budgets and alerts for AI agent spend?](https://opencontroller.lyzr.com/blog/set-budgets-and-alerts-for-ai-agent-spend.md)
- [How do you track Claude spend across Amazon Bedrock, Claude Platform on AWS and the Anthropic API?](https://opencontroller.lyzr.com/blog/track-claude-spend-bedrock-vs-anthropic-api.md)
- [More on Gateways and cost](https://opencontroller.lyzr.com/blog/cluster/gateways-and-cost)
