AI cost management enterprise playbook.

AI cost is doubling every nine to twelve months across most enterprises. Bring it inside FinOps before it becomes the largest shadow line on the cloud bill.

AI spend is the fastest growing line on the cloud bill at almost every enterprise we work with. The pattern is familiar. New technology, scattered ownership, no tagging, no chargeback, no caps. The result is a line item growing three to four times per year with no clear owner.

This playbook lays out the four cost levers, the controls that actually hold, and the org model that gets AI spend back under finance ownership without slowing the rollout.

The shape of enterprise AI cost

Where the spend lives

AI cost shows up in three places. Native foundation model APIs. Cloud provider AI services. SaaS tools with embedded AI features. The fastest growing piece is usually the third, because no one is tagging it.

• Foundation model APIs. OpenAI, Anthropic, Google, AWS Bedrock, Azure OpenAI. Tracked, but rarely tagged at the team level.
• Embedded AI in SaaS. Copilot, Einstein, Now Assist, and similar. Usually billed as add ons inside the larger SaaS contract.
• Internal model hosting. Self hosted open source models running on GPUs. Easiest to lose track of when the GPU cost moves to another team.

Unit economics by use case

Cost per use case varies by two orders of magnitude. Customer support deflection looks cheap per ticket. Long context document analysis can cost dollars per call. Without unit metrics, the executive view is a single growing number with no story.

The four cost levers

Four levers explain almost all of the variance in enterprise AI spend.

1. Model choice

The cheapest model that meets the quality bar wins. Most teams default to the largest model. For most use cases, a smaller or distilled model gets 90 percent of the quality at 10 to 20 percent of the cost.

2. Prompt and context design

Token count is cost. Cutting context length by half cuts cost by roughly half. Caching repeated context drops the bill further. Most teams over send context out of habit.

3. Negotiated rate cards

Retail pricing is the wrong starting point at enterprise scale. Negotiated rate cards routinely save 25 to 60 percent. The negotiated rate is also the floor you benchmark every subsequent renewal against.

4. Caps and quotas

Soft caps and quota alerts at the team level stop runaway use cases before they bend the budget. The cap rarely fires. It just shifts behavior.

The controls that actually hold

Pricing tactics without controls are a one quarter win. The controls keep the discipline through the next launch.

Tagging and attribution

Every model call should carry a team or product tag. Provider native attribution is usually weak. Build a thin proxy layer or use one of the AI gateway tools that already do this.

Chargeback and showback

Showback is the floor. Chargeback is what changes behavior. The first month of chargeback is usually the moment teams discover that the cheap experiment was actually a six figure habit.

Unit economics targets

Each use case gets a target cost per call or per outcome. The product team defends it. Without a unit target, every use case grows toward the budget ceiling.

The cheapest model that meets the quality bar wins. Most teams default to the largest model and find out later that the bar was lower than they thought.

Who owns AI cost

AI cost has the same ownership trap as cloud did. Finance owns the number. No one owns the behavior.

FinOps inherits AI

The cloud FinOps team is the natural home. They already run tagging, chargeback, and unit economics. Adding AI is incremental, not a new function.

Product owns the use case

Product or engineering leadership owns the unit target. Finance reports on it. FinOps automates the alerts. Procurement owns the rate card.

AI council for cross cutting calls

A standing council across finance, security, legal, and engineering decides model selection, rate card refresh, and the use cases that get unbounded spend versus capped spend.

Your 90 day AI FinOps plan

Three months of focused work gets most enterprises from no control to credible control.

Days 0 to 30. Visibility

Tag every model call. Build the dashboard. Identify the top ten teams and use cases. Confirm the rate card on every active contract.

Days 30 to 60. Levers

Switch the obvious model choices. Cut prompt and context bloat. Renegotiate the rate card where the data supports it.

Days 60 to 90. Controls

Stand up chargeback. Set unit targets per use case. Calendar the renewal. Hand the operating model to FinOps.

Suggested reading

• AI consumption billing and token cost control. Cross vendor reading from the wider library.
• Enterprise AI platform TCO comparison. Cross vendor reading from the wider library.
• Enterprise AI procurement framework. Cross vendor reading from the wider library.

What to do next

1. Pull a single view of every AI contract and add monthly run rate.
2. Tag every model call at the team or product level.
3. Identify the top five use cases and their unit economics.
4. Pick the obvious model switches that drop cost without losing quality.
5. Renegotiate the rate card where data shows the gap to benchmark.
6. Stand up showback this quarter, chargeback next quarter.
7. Set unit targets per use case and review monthly with product leads.
8. Talk to a buyer side advisor when annual AI spend crosses one million dollars.

Frequently asked questions

How fast is enterprise AI spend really growing?

Three to four times per year is the typical pattern across our portfolio. Some are growing faster. None are flat.

Does FinOps cover AI today?

Rarely. Most FinOps teams have AI on the roadmap but not yet in tagging and chargeback. Bringing it in is the first move.

Are negotiated rate cards really worth it under a million dollars?

Often yes. Even at the lower seven figure level we routinely see 25 percent or more come off retail. Below that it depends on the vendor.

Should we standardize on one foundation model vendor?

Standardize on a primary, qualify a secondary. Single vendor is fragile. Three or four vendors are unmanageable.

How do we charge back AI cost fairly?

Tag at the team and use case level. Charge on actual usage. Publish unit economics monthly. Disputes shrink fast once teams see their own numbers.

Is AI cost a procurement problem or an engineering problem?

Both. Procurement owns the rate card. Engineering owns the unit metric. Finance owns the report. The cleanest org model has all three pulling on AI together.