AI Economics · Wedge 2

One row per inference. Six providers, one schema.

OpenAI, Anthropic, Bedrock, Vertex, self-hosted clusters, and any framework that emits OpenTelemetry GenAI spans — normalised into a canonical event that carries cost, tokens, latency, the workload, the customer, and the trace. Chargeback-ready without a migration.

Request pilot access

Canonical event shape

40 columns

Unattributed SLI target

<3%

Provider sources

6 live

Where your data comes from

Six ingest sources, one normalised shape. Drop in your keys or point your OTel collector at our receiver — you are importing spend inside an hour.

OpenAI

Organization usage API · project / user / api-key / model / batch

Anthropic

Organization usage API · workspace / api-key / model / service tier

AWS Bedrock

CloudWatch metrics + optional invocation logs · request attributes preserved

Google Vertex

BigQuery billing export · SKU → model inference · labels for attribution

Self-hosted inference

Prometheus scrape · vLLM / TGI / Ollama / Triton · tokens + GPU-seconds

OpenTelemetry GenAI

OTLP/HTTP receiver at /api/v1/otel/genai · gen_ai.* semconv

What makes this different from a cost Explorer filter

A cost explorer lets you filter on tags your vendor already sent. AI Economics fills the gaps, preserves provenance, and surfaces attribution quality as a first-class number.

Every row is a provider row

We never estimate when the provider reports actual cost. The `source` column tells you exactly where the number came from — provider API, OTel trace, or Prometheus scrape.

Unattributed share is the SLI

A live <3% target on the dashboard. If a customer or feature tag is missing, we show it. No hidden slippage into a mystery bucket.

Chargeback without a migration

Priority-ordered allocation rules fill missing customer tags from workload_id, model, regex, header, or OTel trace attrs. Provider-reported attribution is never overridden.

Self-hosted inference is not a footnote

GPU-seconds flow from Prometheus; the dollar joins against your CUR. You see $/1K-tokens for your vLLM cluster the same way you see it for GPT-4.5.

Ingest via OpenTelemetry in 30 seconds

Point your OTel collector at the receiver. Any SDK emitting the gen_ai.* semantic conventions is picked up automatically — OpenLIT, OpenLLMetry, or a hand-rolled tracer.

# OTel collector config (exporters: otlphttp/finesse)
exporters:
  otlphttp/finesse:
    endpoint: https://api.finesse.cloud/api/v1/otel/genai
    headers:
      x-api-key: fns_live_...

service:
  pipelines:
    traces:
      receivers: [otlp]
      processors: [batch]
      exporters: [otlphttp/finesse]

The receiver only persists spans carrying gen_ai.* attributes — the rest are silently dropped so you can fan out a single collector to Finesse without double ingest.

See every inference your business paid for.

Pilot access is reviewed with your team. Your canonical AI cost row lives with you even if you leave — a single JSONL export lists every event we ever stored.

Request pilot access

Product AI Economics

AI Economics · Wedge 2

One row per inference. Six providers, one schema.

Request pilot access

Canonical event shape

40 columns

Unattributed SLI target

<3%

Provider sources

6 live

Where your data comes from

Six ingest sources, one normalised shape. Drop in your keys or point your OTel collector at our receiver — you are importing spend inside an hour.

OpenAI

Organization usage API · project / user / api-key / model / batch

Anthropic

Organization usage API · workspace / api-key / model / service tier

AWS Bedrock

CloudWatch metrics + optional invocation logs · request attributes preserved

Google Vertex

BigQuery billing export · SKU → model inference · labels for attribution

Self-hosted inference

Prometheus scrape · vLLM / TGI / Ollama / Triton · tokens + GPU-seconds

OpenTelemetry GenAI

OTLP/HTTP receiver at /api/v1/otel/genai · gen_ai.* semconv

What makes this different from a cost Explorer filter

A cost explorer lets you filter on tags your vendor already sent. AI Economics fills the gaps, preserves provenance, and surfaces attribution quality as a first-class number.

Every row is a provider row

We never estimate when the provider reports actual cost. The `source` column tells you exactly where the number came from — provider API, OTel trace, or Prometheus scrape.

Unattributed share is the SLI

A live <3% target on the dashboard. If a customer or feature tag is missing, we show it. No hidden slippage into a mystery bucket.

Chargeback without a migration

Priority-ordered allocation rules fill missing customer tags from workload_id, model, regex, header, or OTel trace attrs. Provider-reported attribution is never overridden.

Self-hosted inference is not a footnote

GPU-seconds flow from Prometheus; the dollar joins against your CUR. You see $/1K-tokens for your vLLM cluster the same way you see it for GPT-4.5.

Ingest via OpenTelemetry in 30 seconds

Point your OTel collector at the receiver. Any SDK emitting the gen_ai.* semantic conventions is picked up automatically — OpenLIT, OpenLLMetry, or a hand-rolled tracer.

# OTel collector config (exporters: otlphttp/finesse)
exporters:
  otlphttp/finesse:
    endpoint: https://api.finesse.cloud/api/v1/otel/genai
    headers:
      x-api-key: fns_live_...

service:
  pipelines:
    traces:
      receivers: [otlp]
      processors: [batch]
      exporters: [otlphttp/finesse]

The receiver only persists spans carrying gen_ai.* attributes — the rest are silently dropped so you can fan out a single collector to Finesse without double ingest.

See every inference your business paid for.

Pilot access is reviewed with your team. Your canonical AI cost row lives with you even if you leave — a single JSONL export lists every event we ever stored.

Request pilot access