For finance
The AI spend number your board can't poke a hole in
An AI cost report is one page that shows what your AI tools cost and what you get back. Every dollar on one side. What it bought on the other. Every figure says how we know it. We build it for you, and we lead with the half you can bank this quarter.
Free Snapshot, no card. We can only read. Nothing writes back to your systems.
Start with the recovery, because an invoice confirms it. Seats billed and never opened — a seat is a paid license for one person. Token spend nobody planned for. A token is the unit AI vendors bill for, roughly a few characters of text. The same thing bought twice under two vendor names. Subscriptions on personal cards.
That runs 10 to 15% of seat and subscription spendacross the public benchmarks and audits we've run. On token-heavy stacks the recoverable share sits lower and the routing findings carry more of it. We read it straight out of billing and vendor consoles, and it lands in days. That number is counted. You can put it in a forecast.
The productivity half is harder, and it comes second on purpose. We will not give you a single number, and that is the feature.
Any tool that hands a CFO a single ROI figure is selling something.
Why AI cost visibility lags AI spend by 5×
Leaders who can see AI costs are 5× more likely to prove ROI (KPMG, Q2 2026).
KPMG surveyed 2,145 C-suite leaders across 20 countries in Q2 2026. Seven percent report proven ROI on AI. Among leaders who can see their AI costs clearly, that rises to 15%. Among those who cannot, it falls to 3%. Forty-two percent can see only part of what they spend.
Read that as a finance problem, not an AI problem. You cannot work out a return when you do not know the number you are dividing by.
Source: KPMG Global AI Pulse, Q2 2026.
Every figure carries a proof level
Every number we show carries a proof level, so you know how much guessing sits behind it. There are three, and they do not mix.
Recovery is counted. Productivity is estimated, shown as a band 20% either side. It never appears without the cost of going faster beside it: rework within 21 days, bugs per PR, unreviewed merges, incidents per PR. Rework is work that has to be redone.
One more rule, because it is where most AI ROI decks break. We count each dollar once. A saved hour is never booked as both new feature capacity and headcount avoided.
More: The full method, layer by layer · Honest AI ROI modeling: ranges, not fake precision
Free template
The one-page AI cost report, blank and yours
The structure we produce, as a one-pager you can fill in by hand today. Spend by vendor and category. Recovery by proof level. Estimated upside as a band. The cost of going faster. And the rule that we count each dollar once, printed on the page. Shaped for your board. No account needed.
No email wall. If it is useful without us, take it.
What it costs and when it clears
Observe is $16 per engineer per month, per engineer covered, never per employee. At 200 engineers: $3,200 a month, $38,400 a year.
Break-even is $3,200 a month of recovered spend. That is 107 unused ChatGPT Business seats at $30, or 80 idle Cursor Teams seats at $40, or one extra usage charge you stop repeating.
Whether it clears depends on what you spend, and the honest answer differs by company.
Where Observe pays for itself, and where it does not.
Observe pays for itself above roughly $200 per engineer per month of AI spend. Below that, use the free Snapshot — it's free for a reason.
Below that, the recovery half alone may not cover Observe. Run the free Snapshot, weigh both halves, and come back when the bill grows. We would rather tell you that now than at renewal.
See full pricing · Where AI cost actually builds up · Run your numbers on the savings calculator
Three questions you are going to ask
Is this counted or estimated?
Both, labelled separately, never blended into one figure. The recovery line is counted from invoices, consoles and billing exports. The productivity line is estimated and says so on its face, as a range. If a number cannot name its source, it does not ship.
Can I show this to auditors?
That is what it is built for. Every figure shows its formula and its inputs. Our connectors read. They never write. Nothing in your records changes because we counted it. SOC 2 — a security audit standard — as of 26 July 2026 sits at 20 of 34 controls met and enforced in code. The readiness report is yours on request, rather than implied by a badge.
How do I know your numbers are not flattering the vendors?
Because the model is built to come back negative when the truth is negative, and the sample report shows it doing exactly that. We sell no AI tools. We resell nothing. We take no vendor referral fees.
Questions finance asks
How long until I have a number I can defend?
Recovery findings land in days, because they come out of systems of record. The estimated productivity band needs about a quarter of history before it means anything, and we will say so rather than fill the gap.
Does this replace our SaaS management platform?
No. Those tools tell you what you bought. This tells you what got used, per team, and what that means for the next renewal.
Who should own this inside the company?
Usually finance owns the number and engineering owns the decisions it drives. See who owns the AI budget.
What is actually free?
Snapshot: a one-page AI cost report, no card, free account needed. It is a product, not a trial. Observe adds ongoing tracking at $16 per engineer per month. Control is $40.
Bring a number that holds up
The counted half is on your invoices right now. Everything after that is a conversation you get to have from a stronger position.
Finance teams sometimes call this an AI P&L.
Snapshot is free, no card. Observe $16 per engineer per month. Control $40 per engineer per month.