Read from the source
Agent calls, tokens and cost from your AI bills, status changes in Jira and events in GitHub. Timestamps give elapsed time, not hours worked.
How it works
A labeled estimate on day 0. Traced to the task by day 30.
Nothing to install on laptops, nothing in the path of your agents, nothing written back to your systems.
Rollout
Day 0
Connect, read-only
Connect Jira, GitHub and your AI bills with read-only scopes. AVYV joins them by time window and author and shows team-level cost per finished unit, labeled inferred.
Day 30
Add one ID
Engineers add one correlation ID to agent calls. Each agent attempt and its cost then join the work item it advanced, and the AI cost moves from inferred to observed. Review and rework keep the label for how they were measured. Nothing is added to Jira.
Day 60
Decide
Each task class passes or fails the automation gate, against criteria your CFO signed on day one. You keep the numbers either way.
How we know the human cost
Every figure says how it was measured.
The correlation ID ties each agent attempt and its cost to a work item. It can't see people's time, so that is sampled, and the label says so.
Measured on a sample
Review and rework hours come from a time study on a sample of work items, applied to the task class with its range.
Matched from evidence
AI cost tied to a work item by time window and author where no correlation ID links them. It stays labeled inferred until one does.
Set by Finance
Loaded rates per role. Your Finance team sets them, and AVYV doesn't measure them.
The instrument
Every number opens to the records that produced it.
The figure on the board slide and the rows behind it come from the same query. If a figure can't be opened to its records, AVYV doesn't show it.
HeadlinePROJ-1184 · Refactor auth module$200.28Finished and accepted
Cost of finished work, as a manager sees it.
Explain · one click
For Finance
- AI spend
- $0.28: 3 agent calls, 5,520 tokens observed
- Human review
- $200.00: 2.5 hours sampled at $80/hour declared by Finance
- Elapsed
- 5 hr 25 min from start to accepted observed
- How the AI cost was matched
- $0.22 matched by correlation ID observed and $0.06 by usage pattern inferred
For your architect
The one change at day 30. AVYV reads the ID from billing metadata. It never sees the prompt or the response.
Request header on each agent callX-Work-Item: PROJ-1184Work-Item: PROJ-1184Audit · the 5 source records behind this figure
| Source | Record | Value | Label |
|---|---|---|---|
| Jira | In Progress → In Review | 10:02 | observed |
| Jira | In Review → Done | 15:27 | observed |
| GitHub | Review hours on the linked pull request, from the time study | 2.5 hr | sampled |
| AI billing | 3 calls and 5,520 tokens | $0.28 | observed · inferred |
| Finance | Loaded rate for the engineering role | $80/hr | declared |
Illustrative work item. Every input carries its label and source: observed, sampled, inferred or declared.
In the product
A workflow's page in the AVYV portal.
What a team sees for one workflow: the agent cost of each finished item, and where the time goes between start and done.
Conceptual product preview. Sample data from a demo build, not customer data. Screens may still change.
Three answers a board will ask for.
What finished work costs
AI runs, human review, rework and failed attempts, per task type. The unit is work the business bought, not tokens or seats.
What next quarter costs
Volume, first-pass yield and reviewer capacity drive the forecast. You see the cost before you commit, and which team runs out of room first.
Automate or keep human
Automate a task type when the share that passes first time beats the agent's cost plus the cost of checking it: q > α + ρ. The answer comes from your own records, on numbers Finance accepts.
q = first-pass yield · α = agent attempt cost as a share of human production cost · ρ = verification cost as a share of production costHow the gate works →
30-minute walkthrough
Find out what finished work really costs.
Bring one workflow. In 30 minutes we'll show what your records can prove today, and label what is still an estimate.
Read-only. Metadata only, never prompts or code. 60-day proof of value, and you can walk away at no cost.