Copilot segment creation in Customer Insights

What it does

You describe the segment you want in ordinary English and Copilot turns it into rules in the normal segment builder. It's a front end on the rule builder, not a separate segment engine — what comes out is an ordinary rule-based segment you can edit, save and run.

Key facts

Preview. Microsoft's own page says it isn't meant for production use and may have restricted functionality. Treat anything you build on it as subject to change.
Path is Insights > Segments > New > Build your own, then the Copilot icon to open the Copilot pane. There's no separate "create with Copilot" entry point to bookmark.
Documented region availability is the United States and Switzerland only — the two geographies with a local Azure OpenAI deployment for this workload. Everywhere else depends on the cross-geography consent setting.
Gated by Enable Copilot features powered by Azure OpenAI in Settings > System > Consent. Default is On. Only a Customer Insights - Data admin can change it.
The pane offers suggested prompts generated from your environment's own data, so they change per environment. Handy for demos, less handy when the tenant is half-populated.
Selecting Use applies the generated result to a rule. You then use Edit details to set the segment name and the output table name, and optionally a description and tags.
Save leaves it as a draft; Run saves it and creates the segment. A Copilot-generated segment sitting in draft has no membership and no downstream effect.
Where more than one relationship path exists between tables, Copilot takes the shortest path by default and you can edit it. This is the single biggest source of quietly wrong results.
Thumbs up/down feedback on the generated rule feeds Microsoft's model quality work. Copilot does not share data across environments.
The general help-pane Copilot is suppressed on segment create and edit pages — this dedicated experience replaces it there, so don't expect the usual pane to appear.

When to use / skip

Use it where the person defining the segment knows the business question but not the data model. That's the honest use case: a marketing manager who can say "customers in Manchester who bought in the last 90 days and aren't in the loyalty scheme" but would never find their way through five related tables. It's also a very good demo — it lands well in a steering group and costs you nothing to show, provided the environment is in a supported region and has real unified data behind it.

Skip it for production segment authoring on a delivery you're accountable for. Segments that feed activation or measures need reviewing as rules anyway, and once you're reviewing the rules you may as well have written them. Preview status also means you can't put it in a signed scope document without a caveat. And skip it entirely if the environment is outside the US or Switzerland and the client's compliance position won't allow cross-geography data flow — you'll spend more time on the consent conversation than the feature is worth at this maturity.

Configuration decisions

Whether the global Copilot consent stays On for this environment at all, and who owns that decision — it's tenant-wide and it's an admin-only switch, not a per-user preference.
For non-US/Switzerland environments, whether to turn on cross-geography data flow. That is a compliance decision, not a technical one, and it usually needs the client's data protection lead in the room.
Whether Copilot-generated segments are allowed to reach activation and export, or whether they're confined to exploration and a human rewrites anything that goes live.
Naming and output table conventions for segments created this way. Copilot doesn't impose one, and you'll end up with a soup of similar-sounding segments if nobody does.
How well your relationships are defined and named. Copilot's path selection is only as good as the relationship model beneath it, so relationship curation becomes a Copilot quality task as well as a unification one.
Which roles get Contributor access to the segment area, given that natural language lowers the skill floor for creating segments considerably.

Gotchas

Shortest relationship path is a heuristic, not an answer. If two paths exist between the customer table and a transaction table, the short one may be the wrong one, and the segment will still run and still produce a plausible count. Nothing flags this.
The cross-geography setting doesn't appear at all in US and Switzerland environments, because it's already On and can't be changed. Consultants who saw it in one tenant and can't find it in another usually assume the feature is broken.
Draft versus Run catches people out constantly. Someone builds a segment with Copilot, sees it in the list, and assumes it has members. It doesn't until you run it.
The separate Allow data sharing for Copilot features consent is Off by default and can only be changed in the Power Platform admin center, not in Customer Insights. If the client asks "is Microsoft reading our prompts", that's the setting to point at.
Copilot only knows the tables and columns that exist in your unified data, and it reads their names literally. Cryptic source column names produce cryptic suggestions and poor prompt suggestions.
Preview features change without a release note reaching your client. Don't build training material or runbook screenshots around this yet.

Consultant notes

Demo it, but demo it with a rule expansion visible. Showing the generated rule alongside the prose is what turns "AI magic" into something the client's data team will actually trust.
Tell the client plainly that the reviewable artefact is the rule, not the sentence they typed. Sign-off happens on the rule.
Check the consent tab and the environment region the day before any demo. Copilot silently absent because of geography is an embarrassing five minutes to spend in front of a sponsor.
Push back on any request to make this the primary segment authoring route for a regulated client while it's still preview. Offer it as an assist for analysts instead.
Before go-live, spot-check a handful of Copilot-created segments against hand-written equivalents. Path selection differences show up as membership counts that are close but not equal, which is exactly the failure mode nobody notices.

Worth another look once this leaves preview or the region list grows beyond the US and Switzerland.