Blog · FinOps and IT procurement · 20 Aug 2026

Which Department Is Burning Your AI Budget?

Engineering agent loops can burn budget while sales barely uses seats. A blended invoice hides both problems: paid licenses that sit idle, and power users whose prompt cost dwarfs the seat fee. ConvoMargin measures cost-per-prompt vs output so procurement and FinOps can see ROI before renewal.

Why do bulk AI seats hide shelfware and overage at the same time?

CFOs often buy Microsoft 365 Copilot, Claude, or similar seats in bulk (commonly about $30 per user per month) to standardize access. That purchase can produce two failures on one bill. Some seats are never used. Other seats run heavy multi-step agents that consume API overage or hidden token cost far above the license line.

Seat count is not usage. License coverage is not output. If you only ask “did we buy enough seats?”, you will renew shelfware and miss the teams that actually drive spend. The useful questions are who logged in, who prompted, and what those prompts cost relative to the work produced.

TLDR: One invoice can mix unused seats and expensive loops. Split them. Preview the math.

What does a $30 seat miss about true cost-per-prompt?

A seat fee is a capacity charge. Prompt cost is a usage charge. In vendor-hosted copilots the two can be bundled. In product-built agents they are not: you pay the model provider per token on top of whatever tools you licensed. Either way, average cost per employee is the wrong unit.

Measure cost per completed task, per department, and per prompt family. A legal team that drafts once a week is a different ROI than an engineering org that runs overnight agent loops. If you cannot see that split, you will either over-buy seats or under-price internal usage. ConvoMargin is built for the prompt-and-conversation grain, not for cancelling Microsoft licenses automatically.

TLDR: Seat price is not cost-per-prompt. Join tokens to team and task. SaaS P&L post.

How do you tell which department is burning the budget?

Start with allocation, not blame. Tag prompts or API keys by department, product, or cost center. Compare token dollars to a simple output proxy: tickets closed, specs shipped, deals assisted, or hours saved if you already measure that. Blended bills make engineering look “expensive” even when sales holds unused seats that should have been returned.

CloudZero and Benchmarkit have reported that only 43% of organizations track AI spend by customer and 22% to the transaction (CloudZero). Department-level prompt cost is the same gap inside the company. Until you have it, renewal meetings argue from totals.

TLDR: Tag spend by team, then compare to output. Totals are not ROI. Modeled bands.

What should you measure before a renewal?

Before you sign another year of seats, you want a short list: unused paid seats, power-user loops whose cost exceeds the seat, departments with no mapped output, and prompt templates that retry or retrieve more than they need. Flag unused seats so humans can reclaim them. Do not treat a dashboard as an auto-cancel.

An Audit Sprint is the current ConvoMargin path for a 48-hour read of cohort and prompt drivers. Telemetry middleware is on the waitlist for ongoing margin vs plan. The public calculator will not ingest your M365 admin center. It will show you how fast conversation-level cost can swamp a fixed price, which is the same shape as a seat that looks cheap until the loops start.

TLDR: Unused seats and hot loops are different fixes. Measure both. Book an audit.