AI cost · Showback
DashboardFinance & ITShowing each team what its AI work costs
One AI bill for the whole workspace tells nobody what to change. This dashboard shows each team the cost of its own workflows and Relays, warns before a budget runs out and points at the jobs worth moving to a cheaper model.
Valnora & Elric team5 min read
Top workflows by cost, month to date
| Workflow | Team | Relay | Runs | Cost | Status |
|---|---|---|---|---|---|
| Lookbook video drafts | Marketing | None | 96 | US$1,140 | Review |
| Ticket reply drafts | Support | Support triage | 9,870 | US$520 | On track |
| Product copy rewrite | Marketing | None | 1,240 | US$410 | Review |
| Deal notes summary | Sales | Pipeline watch | 3,310 | US$290 | On track |
| Variance commentary | Finance | None | 42 | US$95 | On track |
5 lines
One cost line per team instead of one bill
Illustrative: five teams, each seeing its own workflows and Relays.
80%
Budget share that triggers the alert
Illustrative threshold. Set your own per team.
1 per week
Usage export and import
Illustrative cadence. A daily import gives a fresher view.
Illustrative example. Figures show how the workflow is measured; they are not customer results.
At the end of the month, the finance lead gets one AI usage figure for the whole workspace. It went up. Marketing says it is support's reply drafts. Support says it is marketing's video. Nobody can say, so nothing changes.
Showback fixes the first half of that problem: each team sees the cost of its own work. This post builds a showback dashboard in Agent Serve from data the workspace already records, adds a monitor on each team's budget, and shows how to move the heaviest jobs to cheaper models. All figures on the dashboard are illustrative.
Where the numbers come from
From run logs to a cost line per team
Data · Usage in
- 01Run logsEach workflow run records its trigger, duration and total cost. Logs can be filtered by folder and exported as CSV.
- 02AI usage tableOne row per run from the weekly export: date, workflow, trigger, cost.
- 03Owners tableMaps each workflow to a team, and to the Relay that runs it, if any.
Dashboard · Showback
- 04Showback dashboardCost by team, by workflow and by Relay, month to date against budget.
- 05Budget monitorAlerts a team lead in Slack and email when month-to-date cost passes 80% of budget.
- 06Cost review RelayStarts on the alert, reads the usage table and drafts which jobs to move to a cheaper model.
- 07Team lead decidesChanges the model on a workflow block, or accepts the cost.
Every workflow run records its trigger, duration and total cost in run logs. That is the raw material. Usage on the platform key is metered to the workspace; with your own provider keys, the run still records its cost and your provider bills you directly. This setup does not invent anything on top. It groups what is already there.
Two small pieces of structure make grouping possible:
- One folder per team. Logs can be filtered by folder, so the export for Marketing contains only Marketing's workflows.
- An owners table in Tables that maps each workflow to a team and, if a Relay runs it, to that Relay. Relay work shows up as the workflow runs it starts.
Build the usage table
Once a week, filter run logs to the last seven days and export them as CSV. The export includes the date, workflow, trigger and cost of each run. Import it into an AI usage table. A ten-minute job for one person; run it daily if you want a fresher view.
Then define two named metrics: cost month to date, and cost as a share of budget. Every panel, alert and report reads the same definitions, so the tile and the alert never disagree.
The panels
- KPI tiles. Workflow run cost month to date, teams over 80% of budget, the most expensive workflow and run count.
- Cost by team against budget. The bar chart answers who is close to their limit with how much of the month left.
- Top workflows by cost. Workflow, team, Relay, runs and cost. A workflow with few runs and high cost is usually a media job; one with many runs and moderate cost is usually a step using a bigger model than it needs.
Row-level access lets each team lead see their own rows while finance sees everything. A scheduled report sends the PDF to each lead on the first of the month.
A monitor on the budget
Add a monitor on the budget-share metric for each team, alerting at 80%. Monitors re-check as often as every five minutes, though with a weekly import the number only moves when new rows land. Alerts go to email, in-app and Slack.
Connect a Relay to the alert, running in Observe mode. It reads the usage table and drafts a short review for the team lead: which workflows drive the cost and what could change. It cannot change a workflow, so it suggests and the lead decides.
What the solution produces
Cost review drafted after the budget alert (illustrative)
MemoMarketing is at 91% of its monthly budget with 38% of the month left.
Lookbook video drafts is nearly two thirds of the team’s spend from 96 runs. Each draft uses a premium video model at full length.
Product copy rewrite runs 1,240 times on a large reasoning model for a short rewrite.
Suggestion 1: draft videos with a Fast video model variant and keep the premium model for the final cut.
Suggestion 2: test the copy rewrite on a smaller model against twenty approved examples before switching.
Nothing has been changed. The team lead decides.
Moving heavy jobs to cheaper models
The review usually points at one of two patterns, each with a fix in the product:
- Premium media models used for drafts. Set a Fast video variant, such as Veo 3.1 Fast, on the drafting step and keep the premium model for the final cut. Generations above the cost confirmation threshold already ask before running in Chat.
- Large reasoning models on routine steps. Choose a smaller model on that workflow block. Test it on twenty approved examples first, and switch only if the results hold.
After the change, the next export shows whether it worked. That loop, measure, change one block, measure again, is the point of showback.
How to measure it
5 lines
One cost line per team instead of one bill
Illustrative: five teams, each seeing its own workflows and Relays.
80%
Budget share that triggers the alert
Illustrative threshold. Set your own per team.
1 per week
Usage export and import
Illustrative cadence. A daily import gives a fresher view.
Illustrative example. Figures show how the workflow is measured; they are not customer results.
What showback is not
Controls and governance
Roles and permission groups
Admin, write and read roles per workspace, plus permission groups that limit which models, blocks and integrations a team can use.
Row-level access
Row-level policies narrow what each person sees in tables, dashboards, exports and scheduled reports.
Run logs
Every workflow run records each block’s inputs, outputs and errors, so a reviewer can trace how a result was reached.
Audit trail
Proposals, approvals, declines and executed actions are written to the workspace audit log, which admins can export.
Bring your own keys
Use your organisation’s own model provider keys at organisation or workspace level.
More on how Agent Serve handles access, data and audit on the security page.
It is not a billing system. It does not move money between budgets or bill teams. It also covers workflow runs only: Chat usage and a Relay's own reasoning are part of workspace usage but are not broken out here. For the levers themselves, see keeping AI model costs low, and for setting per-team limits before spend happens, access policies per team.
Questions teams ask
Is this chargeback?
No. It shows each team what its work costs; it does not move money or produce invoices. If finance wants to recharge teams, the usage table is the input to that process, which happens outside Agent Serve.
Does it include Chat usage?
Not in this setup. It covers workflow runs, which record cost in run logs. Chat and a Relay’s own reasoning are part of workspace usage, but they are not broken out here.
What if we use our own provider keys?
Run logs still record the cost of each run, so the showback works the same way. The money itself is billed by your provider under your agreement rather than metered to the workspace.
What you need
- Workflows organised into one folder per team, so logs can be filtered by team
- An owners table mapping each workflow to a team and, where relevant, to a Relay
- A monthly AI budget per team, agreed with finance
- A person who exports run logs and imports them into the AI usage table each week
- A Slack channel or email list per team lead for budget alerts
- A small test set for any workflow you plan to move to a cheaper model