# Charging LLM spend back to the team that spent it

> Charging AI spend back by headcount is a guess. In the modelled split it undercharged the heaviest team by $11,400 a month out of $24,000.

URL: https://getculpa.com/llm-cost-chargeback
Last reviewed: 2026-08-05
Rates effective: 2026-07-02

## Answer
LLM cost chargeback means allocating shared model spend to the internal teams that caused it, using recorded per-call attribution rather than a headcount split. One key gives you one number, and a number resists being divided fairly. Culpa, a local-first LLM cost, margin, and forecast ledger, tags every call with its team and prices it individually.

## Why this happens
Most companies start with one API key, because one key is easy, and the bill that comes back is one number. Finance still has to split it, so it gets split by the only thing anyone can measure, which is usually headcount and occasionally an argument. Both are guesses, and AI usage is the wrong shape for a guess: it concentrates. A six-person team running a document pipeline can outspend a twelve-person team using a chat assistant several times over, and a headcount split will move money in exactly the wrong direction while looking scrupulously fair. The damage isn't the accounting, it's the incentive. A team charged for someone else's usage has no reason to reduce its own, and the team actually driving the bill sees a discount on its own consumption. You can only charge back what you attributed, and the cleanest attribution is recorded at call time rather than reconstructed afterwards.

## What this usually looks like
- One API key serves every team and finance splits the bill by headcount.
- A team disputes its allocation and nobody can produce evidence either way.
- Nobody knows which internal team drove last month's increase.
- The teams told to cut usage aren't the ones generating it.
- Your allocation method hasn't changed since the first month you used it.

## Common mistakes
- Splitting shared AI spend by headcount. Why it hurts: Usage concentrates and headcount doesn't, so the split moves money the wrong way. Do instead: Attribute per call and charge the actual figure, even if it embarrasses somebody.
- Issuing one key per team and stopping there. Why it hurts: It splits the bill and tells you nothing about which feature or customer inside a team spent it. Do instead: Keep the key split and record the team on the call as well, so you can go a level deeper.
- Charging back a number nobody can audit. Why it hurts: An allocation you can't show the working for gets argued rather than acted on. Do instead: Publish the per-call detail behind each team's figure alongside the figure.
- Waiting for a clean allocation model before charging anything back. Why it hurts: Spend keeps concentrating while the model is designed, and the incentive stays broken. Do instead: Start showback now, visible and uncharged, and let the numbers make the case.

## Self-check
- Ask how last month's AI spend was split, and what evidence backed the split.
- Name the team you believe spent the most, then check whether you can prove it.
- Find out whether one key serves several teams, and how many.
- Work out what each team would owe on measured usage and compare it to what it was charged.

## Headcount against measured usage, on the same $24,000 (illustrative)
A modelled month at $24,000 of shared spend across four teams of 12, 8, 6 and 4 people, priced on Claude Haiku 4.5 at real rates of $1.00 and $5.00 per million from the price book effective 2026-07-02. Team C runs a document pipeline and the others run assistants. Token volumes, team sizes and the usage split are modelled.
  the month: 19,000M input x $1.00/M + 1,000M output x $5.00/M = $19,000 + $5,000 = $24,000
  headcount split over 30 people: $24,000 / 30 = $800 a head
  so A pays $9,600, B $6,400, C $4,800, D $3,200
  measured usage: A $2,400, B $1,800, C $16,200, D $3,600
  team C: charged $4,800, used $16,200, undercharged by $11,400
  team A: charged $9,600, used $2,400, overcharged by $7,200
Team C is a fifth of the company and two thirds of the bill, and the fair-looking split hands it a $11,400 monthly discount paid for mostly by team A. Nobody chose that. It fell out of the only method available when the only number available is the total.

## Cost figures
Every figure carries its confidence and its source. No figure on this page is provider-reported.
- $11,400 per month, modelled amount a headcount split undercharges the heaviest team on a $24,000 month [calculated] Source: 19,000M input and 1,000M output tokens on Claude Haiku 4.5 at real rates of $1.00 and $5.00 per million from the price book effective 2026-07-02 gives $19,000 + $5,000 = $24,000. Split across 30 people that's $800 a head, so a team of 6 is charged $4,800 against modelled measured usage of $16,200, a shortfall of $11,400. Token volumes, team sizes and the usage split are modelled.

## FAQ
Q: How should we split shared LLM costs between teams?
A: By measured per-call usage, not headcount. AI usage concentrates and headcount doesn't, so a per-head split systematically overcharges light users and discounts heavy ones. In the modelled example it undercharged one six-person team by $11,400 on a $24,000 month.

Q: Is one API key per team enough?
A: It's a good start and it stops at the team boundary. A key tells you which team spent the money and nothing about which feature, customer or release inside that team did. Keep the key split and record the team on the call too, so you can go a level deeper when someone asks.

Q: What is the difference between showback and chargeback?
A: Showback publishes what each team used without moving money. Chargeback moves it. Showback is the sensible first step because it needs no budget agreement and the numbers usually make their own argument once people can see them.

Q: Can we work out last quarter's allocation retroactively?
A: Only as far as your keys already separate. Where several teams shared one key, nothing on the provider's side tells them apart, because the team was never recorded there. Your own request logs may still carry enough to reconstruct part of it. That's the part worth fixing before the next quarter rather than after it.

## Sources
- Anthropic pricing: https://platform.claude.com/docs/en/docs/about-claude/pricing

Run the free Cost Leak Scan: https://app.getculpa.com/scan?source=pseo&slug=llm-cost-chargeback&cluster=problem

Machine-readable index of every guide: https://getculpa.com/api/pages
Human-readable index of every guide: https://getculpa.com/guides
Site overview: https://app.getculpa.com/llms.txt

Privacy: Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.