Paid Embed token volume on Cohere's API fell 18% month over month while paying accounts stayed flat. How would you investigate?

Introduction

Cohere's Embed models convert text and images into vectors that power semantic search and retrieval for enterprise AI systems. Embed is sold per token through Cohere's API, through cloud marketplaces, and inside private deployments. The finance team flags that paid Embed token volume on the public API fell 18% month over month, while the number of paying accounts was flat. Revenue from Embed fell by a similar amount.

This is a hard root cause question because Embed usage is "lumpy" by nature. Customers embed a large document collection once (a backfill), then embed only new and changed documents afterwards. A drop could be healthy, harmful, or a measurement artefact. I'll work through it step by step.

Step 1

Clarifying Questions (3 minutes)

• Is the drop in tokens, requests, or both?

Why it matters: Fewer tokens per request suggests behaviour change (shorter chunks, compression); fewer requests suggests lower activity. Expected answer: Requests fell 6%; tokens fell 18%. Impact on approach: Tokens per request dropped sharply, so we look at how customers call the API, not just whether they call it.

• Was there a pricing, model or API change in the period?

Why it matters: Internal changes are the most common cause. Expected answer: A new Embed model version became the default two months ago; the previous version's deprecation date was announced last month. Pricing was unchanged. Impact on approach: Migration behaviour is a key suspect.

• Is the drop broad or concentrated?

Why it matters: Separates a market shift from a few large customers. Expected answer: The top 20 accounts generate 65% of Embed tokens; we haven't split the drop yet. Impact on approach: Analyse top accounts individually.

• Did marketplace or private deployment volume change?

Why it matters: Customers may be moving channels rather than leaving. Expected answer: Embed usage via cloud marketplaces rose 9% in the same month. Impact on approach: Channel shift is a candidate.