Case topics and AI topic clustering

What it does

AI topic clustering read the text on incoming cases, grouped semantically similar ones and generated a named topic for each cluster, which then drove the Topics report in Customer Service historical analytics. Microsoft deprecated the legacy version and removed it on 31 August 2025, so on a current environment this is a capability you have to plan around rather than configure.

Key facts

Microsoft announced deprecation on 20 June 2025 and ended support on 31 August 2025, at which point the legacy AI topic clustering for cases was removed. The same date applies to the legacy conversation topic clustering.
The documentation states plainly that users can no longer use it. There is no in-product migration path and no named replacement offered on the deprecation notice.
While it existed it was enabled by default whenever Customer Service historical analytics was enabled — which is why some older environments and older docs still describe topics appearing without anyone asking for them.
Settings lived in the Copilot Service admin center under Insights, on a Topics clustering for cases management page.
Clustering ran over one chosen text field. Data mapping offered Case title, Description or Serial number, with Case title as the default.
Data cleaning let you strip tags and named phrases from titles before clustering, which mattered enormously for tenants whose case titles were prefixed by an email subject line or a ticket reference.
The admin page reported Status, Topics generated, Cases associated to a topic (as a percentage), Last successful run and Run frequency. That percentage was the honest measure of whether the feature was working.
The natural language model supported English, French, German, Italian, Japanese, Portuguese, Simplified Chinese and Spanish. Other languages would still cluster, with worse results.
Topic automation for Copilot agents was not supported in Government Community Cloud.
The same deprecation wave took out historical analytics for unified routing (end of support 31 August 2025) and, earlier, the Customer Service Analytics and Omnichannel Insights Power BI template apps (no longer downloadable from 30 April 2024).

When to use / skip

You cannot use it, so the question is really what to do when a client asks for case theme analysis — and they will, because it is a genuinely useful thing to want and it has been demoed to them at some point in the last five years.

The honest answer for most projects is that theme analysis now belongs outside the out-of-box reports. If the client needs it, the routes are structured categorisation (case subject tree, case type, a well-governed root-cause field) reported through the standard analytics, or unstructured text analysis built in Fabric over the Dataverse case data. The first is boring, cheap and reliable, and on most service desks it is what the client actually needed. The second is a real project with a real cost.

Skip the whole conversation if the client's case volumes are low. Clustering is only interesting above a few thousand cases a month; below that a manager reading the case list learns more, faster.

Configuration decisions

Whether the client's demand for theme analysis is met by structured categorisation instead. Decide this in discovery, before anyone builds a Fabric pipeline.
If you go structured: which field carries the theme, whether it is mandatory, and who governs the option set. An ungoverned "Other" option destroys the reporting within a quarter.
Whether case title quality is good enough to support any future text analysis. If titles are auto-generated from email subjects, that is a channel configuration decision with reporting consequences.
Whether any downstream artefacts — saved views, dashboards, Power Automate flows, integrations — reference the old topic entities or the Topics report, and what replaces them.
If the client needs unstructured analysis, whether it lands in the Fabric-based custom data model route (so it appears alongside the standard reports) or in a standalone Power BI workspace.

Gotchas

Older Microsoft Learn pages, blog posts and partner decks still describe the Topics dashboard as a live part of Customer Service historical analytics. Verify against the actual environment before you demo it, because it is a memorable thing to get wrong in front of a client.
The enablement toggle for historical analytics used to switch clustering on silently. Environments provisioned before the deprecation may still carry old topic data and old configuration remnants, which look like a working feature until someone checks the last successful run date.
Deprecation of the unified routing historical reports landed at the same time. If a client's reporting pack quietly depended on both, they lost two things in one weekend and will describe it as one incident.
The Power BI template apps that people reach for as the fallback are also gone — deprecated 6 November 2023, unavailable to download since 30 April 2024. Do not offer them as a workaround.
Rebuilding topic analysis in Fabric is not a like-for-like swap. The old feature was pre-trained and free at the point of use; anything you build now carries capacity cost and model governance.
Language coverage matters if you do rebuild. A UK client with a European shared service centre needs the language question answered before, not after, someone picks an approach.

Consultant notes

Get ahead of this in discovery. If case theme reporting appears in a requirements document, flag the deprecation in writing at that point rather than discovering it during UAT.
Push structured categorisation hard. Most clients asking for AI-discovered topics are really asking "why can I not see what people are contacting us about", and a governed subject tree answers that for a fraction of the cost.
If the environment predates the removal, check for orphaned topic data and stale report references as part of your pre-go-live housekeeping.
When you quote a Fabric-based alternative, quote the capacity and the ongoing ownership, not just the build. Clients accept the build cost and are surprised by the rest.

Worth another look if Microsoft ships a successor to topic clustering — the deprecation notice currently names no replacement.