See what your AI is costing you
AI is usually the only part of an automation that costs money per run. A flow that drafts one reply is pennies; the same flow pointed at a busy inbox is a line on an invoice. The AI Call Log is where you see that as it happens — every model call your workspace made, what it cost, and which feature spent it.
Where to find it
Section titled “Where to find it”Go to Settings → AI, then the Log tab.
The AI settings screen has five tabs — Providers, Models, Routing, Builder History
and Log — and they all live on one page. The Log tab’s address is
/settings/ai?tab=log.
Reading the page
Section titled “Reading the page”Three summary tiles sit across the top — Today, This week and This period — each showing the number of calls, the total cost, the average response time, and any errors. The period follows the day filter in the corner: Today, 7 d, 30 d or 90 d.
Below that is the call table:
| Column | What it tells you |
|---|---|
| Time | When the call ran |
| Model | Which model answered, with the tokens in ↑ and out ↓ |
| Feature | What asked for it — the drafter, an agent, memory, a flow step |
| Status | OK, or Error with the reason |
| Latency | How long the model took |
| Cost | What that single call cost |
Feature is the column that finds the money. Cost per call is tiny by design, so a single row never looks alarming. Group your eye by Feature instead: if one feature name fills the table, that is where your spend is going, whatever any individual row says.
Costs are written in the unit that keeps them readable — 0.029¢ below a cent,
1.37¢ up to ten, 42.5¢ up to a dollar, and $1.23 beyond it.
Why some costs start with a tilde
Section titled “Why some costs start with a tilde”A cost like ~0.203¢ is an estimate, not a billed figure. The two look
similar, so the tilde is there to keep them apart. Hover it and the page tells you
which you’re looking at.
The difference comes from the model provider. When Routario buys AI through a broker, the broker returns the exact amount it just charged, and that number is reproduced as-is with no tilde. Models called directly on Azure return only token counts — no cost field exists in the reply at all — so Routario works the cost out itself: tokens multiplied by Azure’s own published rate.
Those rates are not typed into Routario by hand. They are read from Microsoft’s public price list and refreshed nightly, so a price change on Microsoft’s side arrives on its own.
An estimate is close, but it is a list price, and it can differ from your invoice in two ways worth knowing:
- Discounts are invisible to it. If your Azure agreement includes committed-use or enterprise pricing, your real cost is lower than what’s shown.
- It assumes a standard global deployment. A data-zone deployment bills around 10% higher, and that setting isn’t visible to Routario, so it isn’t reflected.
Treat the estimate as a reliable guide to relative spend — which feature is expensive, whether a change made things better or worse — and your Azure invoice as the final word on the absolute number.
Why a cost is sometimes blank
Section titled “Why a cost is sometimes blank”A dash in the Cost column is a statement, not a gap. Hover it and the page names one of four reasons:
| What you see | What it means |
|---|---|
| This call reported no token usage | The call did no billable work — most failed calls land here. |
| No published rate matches this deployment name | The model ran and cost you something, but Microsoft publishes no standard rate under that name, so it can’t be worked out from the tokens. |
| Rates haven’t been loaded yet | The nightly price refresh hasn’t run on this workspace yet. It resolves on its own. |
| This provider doesn’t return a cost | The model came from somewhere Routario can’t price — a locally-run model, for example. |
The second one is the one to watch: those calls did cost money, they just can’t be shown. If a model you use a lot sits permanently on a dash, your visible total is understating real spend.
Worked example — finding an expensive step
Section titled “Worked example — finding an expensive step”A weekly bill looks higher than expected.
- Open Settings → AI → Log and set the range to 30 d. The This period tile gives you the total to explain.
- Scan the Feature column. One name — say the reply drafter — fills most of the table.
- Filter to it with the Feature dropdown. Now the tiles show that feature alone.
- Look at the token counts in the Model column. A prompt going in at 13,000 tokens costs roughly ten times one going in at 1,300 — and long prompts are usually accidental: a whole document pasted into a step that needed one paragraph. The exception is an Ask AI step with Search the web on: the pages it finds are part of what the model reads, so tens of thousands of tokens in is what a searched answer costs, not a mistake.
- Trim what that step sends, or point it at a smaller model on the Routing tab, then compare the same window a week later.
Checking a specific failure
Section titled “Checking a specific failure”Filter Status to Error to see only calls that failed, with the provider’s reason in the row. Failed calls usually cost nothing, which is why most of them show a dash rather than a number.
Where to go next
Section titled “Where to go next”- See what happened — the Logs page — the workspace-wide record, including an AI lens across every run
- Keeping automations healthy — spotting a flow that has started misbehaving
- Ask AI — the step most of these calls come from