The situation
The question nobody in the room can answer.
It's the last week of the quarter. Finance has the AI line in front of them: 140 seats, a usage bill that grew faster than the team did, and eleven expense claims for tools nobody bought centrally. Nearly $50,000 in all. The question is short. What did we get for it?
Engineering answers with the numbers it has: how many engineers use the tools each week, and a survey where most of them say they feel faster. All of it is true, and none of it answers the question. Adoption says the tools are being used. It doesn't say what they were used on, or whether that work ever shipped.
That isn't carelessness. It's how AI spend arrives: in three shapes, and none of them carries the answer on its own.
Why it's hard
Three shapes of spend, three different blind spots.
Seats
- Tells you
- Who has access, and roughly how often they use it.
- Doesn't tell you
- What any of it was used on. A seat costs the same whether it built a feature or sat idle.
Usage
- Tells you
- How much was consumed, and through which key or account.
- Doesn't tell you
- Which piece of work it served. A token carries no ticket number.
Outside procurement
- Tells you
- Very little, and usually late, when the expense claims arrive.
- Doesn't tell you
- That it exists at all, until somebody adds it up.
The ladder
Four answers, and each one needs the one before it.
Every question anyone asks about AI spend is one of four. They come in order: you can't say whether spend was worth it until you know what it bought, and you can't know that until you know which work it paid for.
- 0
What did we pay?
NeedsThe invoice. Everyone has this one.
- 1
Who spent it?
NeedsEvery seat and every key tied to a team.
- 2
On what?
NeedsEvery run tied to the piece of work it served, at the time it runs.
- 3
Did it ship?
NeedsThat work's outcome on the same record as its spend.
- 4
Was it worth it?
NeedsAn estimate made before the work was approved, to measure against.
The questions
Five questions to ask before your next review.
Put these to whoever owns your AI tools. If the answers take a week of spreadsheets, that tells you where you are on the ladder.
Which team spent it?
Answer 1Seats are billed to people and usage to keys. If two teams share a key, or a platform team holds keys on everyone's behalf, the answer is a guess.
Which piece of work did it pay for?
Answer 2Unless each run is tied to the work item it served while it runs, no report afterwards can put the two back together.
How much went on attempts you threw away?
Answer 3An agent that needs three attempts costs three times as much, and the invoice still shows one line. Retries, abandoned branches and reverted changes are real spend with nothing to show for it.
Did the work it bought actually ship?
Answer 3Spend on a change that was never merged, or was reverted a week later, bought nothing. You only see that when spend and outcome sit on the same record.
Could you have known the cost before you said yes?
Answer 4Without an estimate there's no such thing as an overrun, only a bill. With one, a figure can be high or low, and someone can say why.
Try this on your organisation
How far up the ladder does your spend go?
Enter a typical month and answer four questions about how your spend is recorded today. It takes two minutes, and nothing you type leaves this page.
Add your monthly figures to see where your answers stop.
How this is worked out
- Seat spend reaches answer 1 if you know which team holds each seat. It never goes further: a seat is paid for by the month, not by the piece of work.
- Usage spend reaches answer 1 if each team has its own key, or if runs are tagged with their work item.
- Usage spend reaches answer 2 only when runs are tagged with their work item, answer 3 when you can also see which items shipped, and answer 4 when those items were estimated first.
- Spend outside procurement stays at 0. Nobody can explain spend they can't see.
This counts what you can explain, not what you wasted. Spend you can't explain may well be buying good work. You just can't show it.
With Wakalix
Spend that's attached to the work while the work runs.
The four answers can't be rebuilt from an invoice afterwards. They have to be recorded as the work happens. That is what Wakalix does, on the AI tools your teams already use.
- 1Who spent itEvery work item belongs to a team's cycle, so its spend does too, whichever AI tool that team works in.
- 2On whatEach run is measured from the AI tool's own usage records and attributed to the work item and the agent that did it. It's never an agent's account of itself, and it works the same on a flat-rate subscription.
- 3Did it shipEvery attempt is counted, and each item carries its outcome (accepted, in review or sent back) right next to what it used.
- 4Was it worth itEvery item is estimated when it's planned, from the typical cost of each agent's skills, so you approve the cost with the plan and watch the measured figure against it.
Your AI subscriptions and their bills stay your own. Wakalix never runs the model, and never adds anything of its own to the figure.