Put every AI dollar in context.
Know which work earns the spend, where effort repeats and what is worth changing. agentacct attributes each usage record to the goal it served, finds the cohorts that repeat, and reviews every opportunity beside acceptance and time.
Where August’s spend went.
| Goal | Accepted | Spend |
|---|---|---|
|
Ready for enterprise customers Billing parity · 2 workstreams · Alex |
14 / 19 | $11,981.64 |
|
Ship changes faster and safely Mike · 3 workstreams |
21 / 23 | $7,568.94 |
|
Make the product fast and reliable Search quality · 1 workstream · Sam |
9 / 16 | $7,197.90 |
Usage estimate from client records · Provider fees, not airis pricing · Not an invoice · Sample data
Same usage records, grouped by workstream · 21 of 23 tasks accepted on this goal · Sample
| Agent | Sessions | Spend |
|---|---|---|
|
context-scout Repository discovery · in the 62-session cohort |
47 | $2,398.40 |
|
ci-fixer Failing-check repair |
34 | $1,904.30 |
|
test-runner Suite execution |
41 | $1,286.12 |
|
6 other agents Across the same three workstreams |
63 | $1,980.12 |
Attributed to the session the agent ran, not to the person who created the agent · Sample
01 / 04 The premise
A lower bill is not the goal. Cheaper accepted work is.
- 01 Every usage record is attributed to the session that did the work, then rolled up to the task, the workstream, the goal and the team.
- 02 Cost per accepted task counts every attempt, including the ones that failed or stopped, so cheaper failed work never reads as progress.
- 03 An opportunity is modeled potential. It stays at $0 realized until the change is reviewed, applied and compared on the same acceptance criteria.
02 / 04 Attribute, find, review
From a monthly total to a single proposal.
The same usage records carry the whole story. Nothing is re-entered, and every figure keeps the basis it came from.
01 Attribute the spend
Follow the money to the work.
Usage arrives from the clients your agents already run in. agentacct attributes each session to the task it worked on, then rolls the figure up to the workstream, the goal and the team. A session that no task claims stays visible as unassigned instead of being spread across the rest.
- Cost by goal, workstream, person and agent from one set of records
- A usage estimate, never an invoice; provider fees stay separate from airis pricing
- Missing attribution stays missing, and is shown that way
02 Find repeated waste
Find the repetition worth removing.
Across the month, 62 sessions each re-read the same repository before doing any work. agentacct groups them into a cohort, models what a shared skill would have removed, and keeps that cohort separate from the other two opportunities so no dollar is counted twice.
- Non-overlapping cohorts with modeled, not realized, potential
- A draft skill, workflow or hook with an owner and acceptance criteria
- Nothing changes in a runner until a person reviews the draft
Three non-overlapping cohorts across 3 teams. $0 realized until a change is reviewed, applied and compared.
Stop rediscovering the same repository.
Potential is the modeled reduction from the cohort’s current cost with the skill in place · Not realized savings · Sample data
03 Review the opportunity
Review it beside acceptance and time.
Before anyone applies the skill, it runs against the same 12 CI tasks as current practice: same input, same revision, same environment, same acceptance criteria. Unfinished tasks stay in the comparison, and four held-out cases check that the gain is not tuned to the sample.
- Cost per accepted task, not cost per attempt
- Failed and unfinished tasks retained in the cost basis
- Held-out cases and an explicit application state
- 1 Understand
- 2 Prepare
- 3 Compare
- 4 Apply & learn
03 / 04 What each number claims
Four figures. Four different claims.
Each figure on this page makes one bounded claim. The product keeps the vocabulary fixed, so an engineer and a finance partner read the same thing.
What the clients recorded.
Usage captured by the clients your agents run in, expressed in dollars. It estimates provider fees. It is not an invoice, and airis pricing is never part of it.
$26,748.48 · August 2026 · Sample
Every attempt, over what was accepted.
The usage of every attempt at a task set, failed and unfinished attempts included, divided by the number of tasks accepted. Failed and unfinished attempts stay in the total cost.
$48.12 → $26.96 · 12 paired tasks · Concept data
How much a cohort’s cost could fall.
The difference between current cost and modeled cost with the prepared change in place. Cohorts do not overlap, so potential is never counted twice.
$8,413.65 · 3 cohorts · Modeled
Only after compare and apply.
Nothing is realized until a change is reviewed, applied to future runs and compared on the same acceptance criteria. Until then the figure reads $0.
$0 realized · Repository context · Draft v1 · Sample
04 / 04 Related
Cost is one of three outcomes to improve.
A cost figure only makes sense beside what was accepted and how long it took. The other two stories use the same synthetic workspace.
Define acceptance, inspect the evidence, compare on equal terms.
03 · Outcome Delivery speedFind the delay, prepare a better sequence, compare completion.
Platform agentacctOne place to understand and improve agent work: cost, quality and speed in one connected workspace.
See it with your own spend in mind
Put your AI spend in context
with airis.
A guided look at attribution, cohorts and a fair comparison for a sample workspace, and a conversation about your teams and tools.