AI cost review

A preview of how this template lays out — the boards, their custom workflow states, and where the sample issues land. WIP caps show a badge.

• Investigate the workspace approaching its monthly AI budget

  Quotas gate AI features when the budget is exhausted — decide whether to raise the cap or rein in usage before it blocks the team.

  • type:budget
  P12pt

• Review which agent surfaces drive the most inference spend

  Attribute spend by surface (triage, summarize, skills) and decide where to optimise prompts or routing.

  • type:attribution
  P22pt

• Decide the AI budget for the next quarter

  • type:planning
  P11pt

These rules aren't created when you apply the template — they're recipes you can wire up in Settings → Automations once your board exists.

• When

  A workspace crosses 80% of its monthly AI budget

  Then

  Open a review card so a human can decide before quotas start gating features

• When

  AI spend for a period is reviewed and approved

  Then

  Record the decision and the rationale for the audit trail

• When

  A new quarter begins

  Then

  Create the budget-setting card and pull in last quarter's actuals

The column order separates noticing a problem from deciding what to do about it, on purpose. 'To review' and 'Investigating' exist as two lanes instead of one because the instinct when spend spikes is to jump straight to an action — raise the cap, throttle a feature — before anyone has actually attributed where the dollars went. A board that only has 'flag it' and 'fix it' skips the step that makes the fix correct instead of just fast.

1. 1.Let the 80%-of-budget signal open a card, not close a decision

   A workspace crossing most of its monthly AI budget is information, not a verdict — it might mean healthy usage growth, a runaway loop in one agent surface, or a workspace that was never right-sized for its actual usage. The card exists so a human looks before the quota starts gating features on that workspace mid-cycle, not to demand an immediate cap increase.

2. 2.Attribute spend to a surface before touching a number

   'Which agent surfaces drive the most inference spend' is the real question underneath most cost spikes — triage, summarization, and skill-distillation calls are not interchangeable, and the fix for one (route to a cheaper model tier) is different from the fix for another (cap call frequency, tighten the prompt). Raising a budget without this step just pays for whatever the actual problem is, on a longer leash.

3. 3.Distinguish a workspace-level fix from a system-level one

   Some investigations end with 'this one workspace needs a bigger or smaller cap.' Others end with 'this prompt is generating more tokens than the task requires, across every workspace that uses it' — and only the second one is worth fixing once instead of quota-adjusting forever. A review that stops at the cap number misses the compounding fix sitting one level down.

4. 4.Decide, then record the rationale, not just the number

   'Approved' should carry why — the investigation's finding, not just a new cap. A budget log that is only numbers cannot answer, six months later, whether a given increase was a one-time exception or should have become the new baseline. The rationale is what makes the next quarter's planning card faster instead of a repeat of the same investigation.

5. 5.Set next quarter's budget from actuals, not the plan tier's sticker number

   The recurring 'decide the AI budget for next quarter' card should pull real usage from the quarter that just ended, not restart from a generic per-tier default. A workspace that consistently sits well under its cap is over-provisioned; one that consistently brushes the ceiling under normal use, not a spike, is under-provisioned — both are visible only if you look at actuals first.

This board encodes one opinion about how the work should run. Here is where that opinion is wrong and something else fits better.

• You have real FinOps tooling that already covers this. Datadog Cloud Cost, CloudZero, and similar platforms do granular cost attribution and anomaly detection across your entire cloud bill, of which AI inference is one line item — if that is already wired up, this board duplicates a slice of what it already does, worse.
• Your AI spend is small and flat. A recurring review board for a cost that barely moves month to month is process for its own sake — check in quarterly from a spreadsheet until the spend is large or volatile enough to justify a standing board.
• You need real-time budget enforcement, not a review cadence. This board is where a human decides what to do about spend signals; the actual hard-stop and soft-cap enforcement has to live in the metering system itself (the credit ledger, the spend breaker), not in whether someone got to a kanban card this week.
• You have one workspace, not many. Per-workspace attribution is the whole value proposition here — a single-tenant product with one bill to watch gets little from a board built around comparing spend across workspaces.

Honest comparisons, including where the other tool wins. Planoda is pre-launch, so nothing below is a benchmark — it is a description of how each product approaches this job.

• Datadog Cloud Cost / CloudZero

  Purpose-built cloud cost platforms ingest billing data across every vendor — compute, storage, and AI inference alike — and do automatic anomaly detection, tagging, and chargeback at a depth no kanban board approaches. For total cloud spend visibility, they are simply the more capable tool, full stop.

  Where Datadog Cloud Cost / CloudZero is better: Breadth and automation: if the goal is 'catch any cost anomaly anywhere in the stack automatically,' a dedicated FinOps platform does that continuously and across far more categories of spend than AI alone. This board is narrower and more manual by design — a human-in-the-loop review of one specific budget category.

• OpenAI / Anthropic usage dashboards

  Model-vendor consoles show raw token and dollar usage per API key with good vendor-side detail. What they cannot show is which of YOUR customer-facing workspaces or product surfaces that usage maps to — an API key doesn't know it is being called by your triage feature for one tenant and your summarizer for another.

  Where OpenAI / Anthropic usage dashboards is better: For raw, vendor-verified usage numbers and billing reconciliation, the vendor's own console is the authoritative source — this board's per-workspace attribution is a layer built on top of exactly that data, not a replacement for checking it.

• A monthly spend spreadsheet

  Many teams start AI cost oversight as a sheet someone updates at the end of the month. It is fast to start and requires no new tooling, and for a single flat cost line it is often enough. It loses the linkage between a spend decision and the investigation that produced it — six months in, the sheet has numbers but not the reasoning behind the changes.

  Where A monthly spend spreadsheet is better: Quick to stand up with zero new process, and perfectly adequate for a small, stable spend. If AI cost isn't yet large or volatile enough to need investigation-and-decision discipline, the spreadsheet is the right amount of process.