Customer Insights skills in Copilot Cowork

What it does

A Cowork plugin that lets a Microsoft 365 Copilot user ask about Customer Insights data and act on journeys and email from inside Cowork chat, rather than switching into the Customer Insights app. It resolves one unified profile, checks consent, drafts and validates marketing email against brand rules, and creates or publishes journeys — all grounded in live org data via the Dataverse MCP server, using the caller's own delegated identity.

Key facts

Ships as three distinct skills — Customer brief and value metrics (read-only profile/segment/prediction lookups), Customer outreach (segment resolution, consent gating, journey and email creation/publishing), and Email content advisor (brand-profile-aware review and drafting of marketing copy).
All Dataverse access goes through the Dataverse MCP server using the user's own permissions — Cowork sees exactly what that user could see directly, no elevated service identity in the loop.
Consent checks run the same logic as the Customer Insights - Journeys runtime (contact point, channel, purpose, business unit, optional topic), so a "can I email this segment" answer in chat matches what would actually happen at send time.
Write operations — creating or changing a record, sending anything — require explicit confirmation; nothing sends or saves silently off the back of a single ambiguous prompt.
When a customer name matches more than one profile, or the target environment is unclear, the plugin asks the user to choose rather than guessing or trying every connected environment in turn.
Requires a Microsoft 365 Copilot licence (for Cowork access), a licensed Customer Insights environment (Data, Journeys or both), and the Dataverse MCP server enabled for that environment.

When to use / skip

Good fit for quick, conversational lookups and routine outreach tasks a marketer or seller would otherwise open the full app for — "who's in this segment", "can I mail this list", "does this draft pass brand review". Less useful for building or tuning a journey's actual logic (branches, wait steps, exit criteria) — that's still a canvas job, Cowork doesn't replace journey authoring, it drives existing journeys and profile lookups from chat.

Configuration decisions

Whether to enable the plugin tenant-wide or pilot it with one team first — since every write still asks for confirmation, the practical risk is lower than it looks, but rollout communication still matters so users understand what "the agent is asking the runtime, not guessing" means.
Which environment(s) to connect by default when a tenant runs more than one Customer Insights environment (dev/test/prod, or regional splits) — the plugin won't guess, but picking a sensible default saves a prompt on every use.
Whether the Dataverse MCP server is already enabled for the target environment, since it's a shared prerequisite this plugin depends on rather than something it turns on itself.

Gotchas

Because access is delegated to the calling user, someone with narrow Customer Insights permissions gets narrow (and sometimes confusingly partial) answers from Cowork — that's correct behaviour, but it reads like a bug to a user who doesn't know their own access is the limiting factor.
The consent check is a snapshot at the moment asked; it doesn't lock the audience, so a segment's consent status between the chat answer and an actual later send can still have moved.
Predictions (churn, model scores) are explicitly labelled as estimates in responses — don't let that framing get lost if answers are copied into a deck or report downstream.

Consultant notes

This is a genuinely good on-ramp for Copilot-sceptical marketing users: the "ask a question, get a governed answer with a citation" pattern is a lower-friction pitch than asking them to trust an agent to run a whole journey unsupervised.
Worth demonstrating the multi-profile disambiguation and the consent-blocking-reason output specifically in early client conversations — it's the detail that answers "how do we know it won't just email someone who opted out" before anyone has to ask.
Because it rides entirely on the user's own Dataverse permissions via MCP, get the target environment's security model right first; this plugin doesn't paper over a messy permission set, it just faithfully reproduces its confusion in chat.
Treat the three skills as separable rollout decisions — a client nervous about AI drafting email content can enable brief/lookup and outreach without turning on the content advisor from day one.

Worth revisiting once usage data shows which of the three skills clients actually reach for.