# Cost Tracking & Chargeback via APIM Policies

**Last updated:** 2026-02
**Status:** GA
**Category:** API Management & AI Gateway

---

## Introduksjon

Når organisasjoner skalerer sin bruk av Azure OpenAI og andre AI-tjenester, blir kostnadssynlighet og tildeling av kostnader til riktig avdeling, prosjekt eller team en kritisk utfordring. Azure API Management (APIM) fungerer som et naturlig punkt for å samle inn kostnadsdata fra AI-modeller gjennom policyer som fanger token-bruk, modell-informasjon og forbruker-identitet. Denne informasjonen kan deretter brukes for intern fakturering (chargeback) og kostnadsoptimalisering.

For norsk offentlig sektor med stramme budsjetter og krav om transparens i ressursbruk er APIM-basert kostnadssporing spesielt verdifull. Mange statlige virksomheter deler AI-infrastruktur på tvers av avdelinger og prosjekter, og trenger mekanismer for å fordele kostnader rettferdig. Denne referansen dekker token-telling fra responser, modell-routing-tracking, chargeback-tagging, integrasjon med Azure Cost Management, og egendefinerte metriker.

APIM tilbyr innebygde policyer for å emittere token-metriker (`llm-emit-token-metric`) og logge LLM API-requests med fullstendig token-bruk. Kombinert med Azure Monitor, Application Insights og Cost Management gir dette en komplett pipeline for AI-kostnadssporing fra request til faktura.

---

## Token Counting from Responses

### Azure OpenAI Token Usage

Hver Azure OpenAI-respons inkluderer token-bruk i `usage`-feltet:

```json
{
  "id": "chatcmpl-abc123",
  "object": "chat.completion",
  "usage": {
    "prompt_tokens": 150,
    "completion_tokens": 250,
    "total_tokens": 400
  }
}
```

### llm-emit-token-metric Policy

Den primære policyen for å emittere token-metriker til Azure Monitor:

```xml
<outbound>
    <base />
    <llm-emit-token-metric namespace="ai-cost-metrics">
        <!-- Dimensjoner for kostnadsallokering -->
        <dimension name="Subscription" value="@(context.Subscription.Name)" />
        <dimension name="API" value="@(context.Api.Name)" />
        <dimension name="Product" value="@(context.Product.Name)" />
        <dimension name="ClientIP" value="@(context.Request.IpAddress)" />
        <dimension name="Region" value="@(context.Deployment.Region)" />
        <dimension name="UserId"
                   value="@(context.Request.Headers.GetValueOrDefault("X-User-Id", "unknown"))" />
        <dimension name="Department"
                   value="@(context.Request.Headers.GetValueOrDefault("X-Department", "unassigned"))" />
        <dimension name="CostCenter"
                   value="@(context.Request.Headers.GetValueOrDefault("X-Cost-Center", "default"))" />
    </llm-emit-token-metric>
</outbound>
```

### Token-typer og Kostnader

| Token-type | Beskrivelse | Kostnadsandel |
|-----------|-------------|---------------|
| Prompt tokens | Input-tokens (brukerens melding + system prompt) | Typisk 30-50% av kostnad |
| Completion tokens | Output-tokens (modellens svar) | Typisk 50-70% av kostnad |
| Cached tokens | Tokens fra prompt caching | Rabattert (opptil 50%) |
| Total tokens | Sum av prompt + completion | Grunnlag for fakturering |

### Kostnadsberegning per Request

```xml
<outbound>
    <base />
    <!-- Beregn estimert kostnad per request -->
    <set-variable name="estimated-cost-nok" value="@{
        // Eksempel: GPT-4o priser (tilpasses faktiske priser)
        var promptRate = 0.025m;      // kr per 1000 prompt tokens
        var completionRate = 0.10m;   // kr per 1000 completion tokens

        var body = context.Response.Body.As<JObject>(preserveContent: true);
        var usage = body?["usage"];
        if (usage == null) return "0";

        var promptTokens = usage["prompt_tokens"]?.Value<decimal>() ?? 0;
        var completionTokens = usage["completion_tokens"]?.Value<decimal>() ?? 0;

        var cost = (promptTokens / 1000m * promptRate) +
                   (completionTokens / 1000m * completionRate);

        return cost.ToString("F4");
    }" />

    <!-- Legg til kostnadsinfo i response header -->
    <set-header name="X-Estimated-Cost-NOK" exists-action="override">
        <value>@((string)context.Variables["estimated-cost-nok"])</value>
    </set-header>
</outbound>
```

---

## Model Routing Tracking

### Spore Hvilken Modell som Brukes

Når backend pools med forskjellige modeller brukes, er det viktig å spore hvilken modell og deployment som faktisk betjener requesten:

```xml
<outbound>
    <base />
    <!-- Ekstraher modell-info fra respons -->
    <set-variable name="model-used" value="@{
        var body = context.Response.Body.As<JObject>(preserveContent: true);
        return body?["model"]?.ToString() ?? "unknown";
    }" />

    <set-variable name="deployment-id" value="@{
        return context.Request.MatchedParameters.GetValueOrDefault("deployment-id", "unknown");
    }" />

    <!-- Emit metrikker med modell-dimensjon -->
    <llm-emit-token-metric namespace="ai-model-metrics">
        <dimension name="Model" value="@((string)context.Variables["model-used"])" />
        <dimension name="DeploymentId" value="@((string)context.Variables["deployment-id"])" />
        <dimension name="Backend" value="@(context.Request.Url.Host)" />
        <dimension name="DeploymentType"
                   value="@(context.Request.Url.Host.Contains("ptu") ? "PTU" : "PayGo")" />
    </llm-emit-token-metric>
</outbound>
```

### Modell-pris-mapping

| Modell | Prompt (kr/1K tokens) | Completion (kr/1K tokens) | Type |
|--------|----------------------|--------------------------|------|
| GPT-4o | 0.025 | 0.10 | Standard |
| GPT-4o-mini | 0.0015 | 0.006 | Standard |
| GPT-4 Turbo | 0.10 | 0.30 | Standard |
| text-embedding-ada-002 | 0.001 | N/A | Embedding |
| text-embedding-3-large | 0.0013 | N/A | Embedding |
| PTU (alle modeller) | Fast pris/time | Fast pris/time | Provisioned |

**Merk:** Priser varierer og bør oppdateres jevnlig. PTU faktureres per time uavhengig av faktisk bruk.

---

## Chargeback Tagging

### Implementere Chargeback-modell

En effektiv chargeback-modell krever at hver AI-request tagges med identifiserbar informasjon:

```xml
<inbound>
    <base />

    <!-- Ekstraher chargeback-info fra JWT token -->
    <set-variable name="department" value="@{
        var jwt = context.Request.Headers.GetValueOrDefault("Authorization", "")
                         .Replace("Bearer ", "");
        if (string.IsNullOrEmpty(jwt)) return "unknown";
        var token = jwt.AsJwt();
        return token?.Claims.GetValueOrDefault("department", "unassigned");
    }" />

    <set-variable name="cost-center" value="@{
        var jwt = context.Request.Headers.GetValueOrDefault("Authorization", "")
                         .Replace("Bearer ", "");
        if (string.IsNullOrEmpty(jwt)) return "default";
        var token = jwt.AsJwt();
        return token?.Claims.GetValueOrDefault("cost_center", "default");
    }" />

    <set-variable name="project-code" value="@{
        return context.Request.Headers.GetValueOrDefault("X-Project-Code", "general");
    }" />
</inbound>

<outbound>
    <base />

    <!-- Emit chargeback-metriker -->
    <llm-emit-token-metric namespace="chargeback-metrics">
        <dimension name="Department" value="@((string)context.Variables["department"])" />
        <dimension name="CostCenter" value="@((string)context.Variables["cost-center"])" />
        <dimension name="ProjectCode" value="@((string)context.Variables["project-code"])" />
        <dimension name="Subscription" value="@(context.Subscription.Name)" />
        <dimension name="Model" value="@{
            var body = context.Response.Body.As<JObject>(preserveContent: true);
            return body?["model"]?.ToString() ?? "unknown";
        }" />
    </llm-emit-token-metric>

    <!-- Legg til chargeback-headers i respons -->
    <set-header name="X-Chargeback-Department" exists-action="override">
        <value>@((string)context.Variables["department"])</value>
    </set-header>
    <set-header name="X-Chargeback-Tokens" exists-action="override">
        <value>@{
            var body = context.Response.Body.As<JObject>(preserveContent: true);
            return body?["usage"]?["total_tokens"]?.ToString() ?? "0";
        }</value>
    </set-header>
</outbound>
```

### APIM Products for Chargeback

Bruk APIM Products for å gruppere API-tilgang per avdeling:

| Product | Beskrivelse | Rate Limit | Chargeback |
|---------|-------------|-----------|-----------|
| AI-Standard | Standard AI-tilgang | 10K TPM | Avdelingsbudsjett |
| AI-Premium | Utvidet AI-tilgang | 50K TPM | Prosjektbudsjett |
| AI-Unlimited | Full tilgang (admin) | Ubegrenset | Sentralt budsjett |

```xml
<!-- Product-basert rate limiting -->
<inbound>
    <base />
    <choose>
        <when condition="@(context.Product.Name == "AI-Standard")">
            <llm-token-limit counter-key="@(context.Subscription.Id)"
                             tokens-per-minute="10000" />
        </when>
        <when condition="@(context.Product.Name == "AI-Premium")">
            <llm-token-limit counter-key="@(context.Subscription.Id)"
                             tokens-per-minute="50000" />
        </when>
    </choose>
</inbound>
```

---

## Azure Cost Management Integration

### Log Analytics for Kostnadsdata

Token-bruk logges til Azure Monitor via LLM API-logging:

```
APIM → Diagnostic Settings → "Logs related to generative AI gateway"
     → Log Analytics Workspace → ApiManagementGatewayLlmLog
```

### KQL Query: Daglig Kostnad per Avdeling

```kusto
ApiManagementGatewayLlmLog
| where TimeGenerated > ago(30d)
| extend Department = tostring(CustomDimensions["Department"])
| extend Model = tostring(ModelDeployment)
| extend PromptTokens = toint(PromptTokens)
| extend CompletionTokens = toint(CompletionTokens)
// Pris-mapping (oppdater etter faktiske priser)
| extend PromptCostNOK = case(
    Model contains "gpt-4o-mini", PromptTokens * 0.0000015,
    Model contains "gpt-4o", PromptTokens * 0.000025,
    Model contains "gpt-4", PromptTokens * 0.0001,
    PromptTokens * 0.00001  // Default
  )
| extend CompletionCostNOK = case(
    Model contains "gpt-4o-mini", CompletionTokens * 0.000006,
    Model contains "gpt-4o", CompletionTokens * 0.0001,
    Model contains "gpt-4", CompletionTokens * 0.0003,
    CompletionTokens * 0.00003  // Default
  )
| extend TotalCostNOK = PromptCostNOK + CompletionCostNOK
| summarize
    DailyCostNOK = sum(TotalCostNOK),
    TotalTokens = sum(toint(TotalTokens)),
    RequestCount = count()
  by Department, bin(TimeGenerated, 1d)
| order by TimeGenerated desc, DailyCostNOK desc
```

### KQL Query: Månedlig Chargeback-rapport

```kusto
ApiManagementGatewayLlmLog
| where TimeGenerated > startofmonth(ago(0d))
| extend CostCenter = tostring(CustomDimensions["CostCenter"])
| extend Department = tostring(CustomDimensions["Department"])
| extend Model = tostring(ModelDeployment)
| summarize
    TotalPromptTokens = sum(toint(PromptTokens)),
    TotalCompletionTokens = sum(toint(CompletionTokens)),
    TotalTokens = sum(toint(TotalTokens)),
    RequestCount = count(),
    UniqueUsers = dcount(tostring(CustomDimensions["UserId"]))
  by CostCenter, Department, Model
| extend EstimatedCostNOK =
    TotalPromptTokens * 0.000025 + TotalCompletionTokens * 0.0001
| order by EstimatedCostNOK desc
```

### Azure Workbook for Kostnadsdashboard

APIM tilbyr et innebygd Analytics-dashboard for LLM-APIer:

```
1. APIM → Monitoring → Analytics → Language models
2. Viser: Token consumption, Request count, Modell-fordeling
3. Filtrer etter tidsperiode og API
```

For egendefinert dashboard:

```
1. Azure Monitor → Workbooks → New
2. Legg til KQL-queries for chargeback
3. Visualiser med tabeller, grafer, kart
4. Del med stakeholders via Azure RBAC
```

---

## Custom Metrics

### Emit Custom Metriker med Policy Expressions

```xml
<outbound>
    <base />

    <!-- Emit custom metrikk for estimert kostnad -->
    <emit-metric name="ai-estimated-cost"
                 value="@{
        var body = context.Response.Body.As<JObject>(preserveContent: true);
        var usage = body?["usage"];
        if (usage == null) return 0d;

        var prompt = usage["prompt_tokens"]?.Value<double>() ?? 0;
        var completion = usage["completion_tokens"]?.Value<double>() ?? 0;

        return (prompt * 0.000025) + (completion * 0.0001);
    }"
                 namespace="ai-cost">
        <dimension name="Department"
                   value="@((string)context.Variables.GetValueOrDefault("department", "unknown"))" />
        <dimension name="CostCenter"
                   value="@((string)context.Variables.GetValueOrDefault("cost-center", "default"))" />
    </emit-metric>

    <!-- Standard token-metriker -->
    <llm-emit-token-metric namespace="ai-tokens">
        <dimension name="Department"
                   value="@((string)context.Variables.GetValueOrDefault("department", "unknown"))" />
        <dimension name="Model" value="@{
            var body = context.Response.Body.As<JObject>(preserveContent: true);
            return body?["model"]?.ToString() ?? "unknown";
        }" />
    </llm-emit-token-metric>
</outbound>
```

### Azure Monitor Alerts for Kostnadsoverskridelse

```bicep
resource costAlert 'Microsoft.Insights/metricAlerts@2018-03-01' = {
  name: 'ai-cost-threshold-alert'
  location: 'global'
  properties: {
    severity: 2
    evaluationFrequency: 'PT1H'
    windowSize: 'PT24H'
    criteria: {
      'odata.type': 'Microsoft.Azure.Monitor.SingleResourceMultipleMetricCriteria'
      allOf: [
        {
          name: 'DailyTokenBudgetExceeded'
          metricNamespace: 'ai-tokens'
          metricName: 'Total Tokens'
          operator: 'GreaterThan'
          threshold: 1000000  // 1M tokens per dag
          timeAggregation: 'Total'
        }
      ]
    }
    actions: [
      { actionGroupId: actionGroup.id }
    ]
  }
}
```

### Eksport til Power BI for Rapportering

```kusto
// Eksporter data til Power BI via Log Analytics
ApiManagementGatewayLlmLog
| where TimeGenerated > ago(90d)
| project
    Timestamp = TimeGenerated,
    Department = tostring(CustomDimensions["Department"]),
    CostCenter = tostring(CustomDimensions["CostCenter"]),
    Model = ModelDeployment,
    PromptTokens = toint(PromptTokens),
    CompletionTokens = toint(CompletionTokens),
    TotalTokens = toint(TotalTokens),
    SubscriptionName = tostring(CustomDimensions["Subscription"])
```

---

## FinOps Integrasjon

### Azure Cost Management Tags

Kombiner APIM-metriker med Azure resource tags for helhetlig kostnadsbilde:

| Tag | Formål | Eksempel |
|-----|--------|---------|
| `Department` | Avdelingstilhørighet | "IT-seksjonen" |
| `CostCenter` | Kostnadssenter-kode | "KS-4210" |
| `Environment` | Miljø | "production" |
| `Project` | Prosjektkode | "AI-chatbot-2026" |

### Kostnadsmodeller for AI

| Modell | Fordeler | Ulemper |
|--------|---------|--------|
| Per-token chargeback | Presis, rettferdig | Kompleks å administrere |
| Flat rate per avdeling | Enkelt, forutsigbart | Urettferdig for lavbrukere |
| Tier-basert (freemium) | Balansert, insentiverer effektivitet | Krever grensehåndtering |
| PTU-allokering | Fast kostnad, forutsigbart | Ingen fleksibilitet |

---

## Referanser

- [AI gateway in Azure API Management - Observability and governance](https://learn.microsoft.com/en-us/azure/api-management/genai-gateway-capabilities#observability-and-governance) — Oversikt over token-metriker
- [llm-emit-token-metric policy](https://learn.microsoft.com/en-us/azure/api-management/llm-emit-token-metric-policy) — Policy-referanse for token-metriker
- [Log token usage, prompts, and completions](https://learn.microsoft.com/en-us/azure/api-management/api-management-howto-llm-logs) — LLM API-logging
- [Plan and manage costs for API Management](https://learn.microsoft.com/en-us/azure/api-management/plan-manage-costs) — APIM-kostnader i Cost Management
- [Azure Cost Management overview](https://learn.microsoft.com/en-us/azure/cost-management-billing/costs/overview-cost-management) — Helhetlig kostnadsadministrasjon

---

## For Cosmo

- **Bruk denne referansen** når kunder trenger å implementere kostnadssporing og intern fakturering for AI-tjenester gjennom APIM.
- Start med `llm-emit-token-metric` policy med Department- og CostCenter-dimensjoner — dette gir umiddelbar synlighet i token-bruk per avdeling.
- For norsk offentlig sektor: Anbefal tier-basert chargeback-modell med APIM Products som mapper til avdelingsbudsjetter. Flat rate er for enkelt, per-token er for komplekst for de fleste.
- Husk at PTU-kostnader er faste per time — chargeback for PTU bør baseres på allokert kapasitet, ikke faktisk bruk.
- Kombiner APIM-metriker med Azure resource tags for å gi et helhetlig bilde i Cost Management. APIM-metriker alene viser kun token-bruk, ikke infrastrukturkostnader.