> ## Documentation Index
> Fetch the complete documentation index at: https://docs.thecontextcompany.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Topic clusters and use cases in production AI agents

> See what users are actually trying to do with your agent, which requests are growing, and where the agent handles them poorly.

Topic clusters answer the question **what are users trying to accomplish?** The Context Company groups production conversations by intent and shows the volume, trend, and quality of each cluster so you can see which use cases are growing, which are failing, and which are missing entirely.

## Why use it

Every AI agent has two kinds of gap: the workflows you built poorly, and the workflows you did not build at all. Both hide in production conversations. Topic clusters make them visible without hand-labeling every session.

You can see:

* The most common things users are asking the agent to do.
* New use cases that appeared this week or this month.
* Topics with high frustration, high escalation, or high failure rates.
* Use cases that are growing quickly and deserve investment.
* Requests for capabilities the agent does not currently support.

## How it works

The Context Company continuously scans conversation content across every run and groups sessions into topic clusters. Each cluster shows:

* A short label describing the intent.
* Volume: number of runs in the cluster and share of overall traffic.
* Trend: stable, growing, declining, or newly emerged over the selected window.
* Example prompts drawn from real sessions in the cluster.
* Rollups for feedback, [Pattern](/analyze/patterns) hits, [failures](/analyze/failures), tools involved, and models used.

Clusters update as new conversations arrive. A topic that first appears this week shows up as a new cluster; a topic whose volume shifts substantially is flagged as trending.

## Discover, explore, track

The topic workflow mirrors the broader [discovery loop](/analyze/discovery):

<Steps>
  <Step title="Discover">
    Open the topics view (or the topic section of the latest [recap](/analyze/recaps)) to see emerging and shifting clusters.
  </Step>

  <Step title="Explore">
    Click into a cluster to see representative runs, feedback, tool usage, and failure rates. Use [Insight Search](/analyze/insight-search) for follow-up questions like "what tool errors show up in the scheduling topic?"
  </Step>

  <Step title="Track">
    Turn important topics into a [custom Pattern](/analyze/patterns#custom-classifiers) if you want ongoing monitoring, alerts, or precise counts per release.
  </Step>
</Steps>

## Example questions

* What are the top ten things users are trying to do with the agent right now?
* What new use cases appeared this month?
* Which topics have the highest thumbs-down rate?
* Which topics are growing fastest?
* Which requests most often escalate to a human, based on session-level signals?
* Which organizations are driving the growth in a specific topic?
* Which requests correspond to capabilities we have not built yet?

Any of these can be asked directly to [Insight Search](/analyze/insight-search), filtered by cluster.

## Segmenting clusters

Because clusters are computed on top of the [normal run model](/concepts), you can filter or segment by any of the standard dimensions:

* [User or organization](/analyze/users-and-organizations) to see whether a topic is concentrated in a few accounts.
* Agent, workflow, or [tool](/analyze/tools-and-models) to attribute cluster behavior to a specific part of the system.
* Custom metadata such as plan tier, region, or [agent version](/analyze/compare-versions) to compare cluster behavior across cohorts.

## Interpreting clusters

Volume alone is not a quality signal. A high-volume cluster with strong feedback and low failure rate means the agent handles that workflow well. A high-volume cluster with repeated frustration, silent failures, or tool errors usually points to one of:

* A wrong or inconsistent answer the agent gives on that topic.
* A missing knowledge base entry or stale content.
* A tool that succeeds technically but produces the wrong result.
* A workflow the agent was never built to handle.

The last case is often the highest-leverage finding: users are asking for a capability that does not exist, and the agent is quietly failing at it.

## Limitations

* Clusters improve with traffic volume. Very low-volume tenants may see broad or unstable groupings.
* Topic labels are generated automatically and are meant to be readable, not authoritative. Two clusters that look similar may still be distinct.
* Clustering runs on conversation content; agents that operate entirely through tools without free-text turns produce fewer meaningful clusters.

## Related

* [Discovery](/analyze/discovery)
* [Insight Search](/analyze/insight-search)
* [Patterns](/analyze/patterns)
* [Recaps](/analyze/recaps)
* [Users and organizations](/analyze/users-and-organizations)