Journey creation agent

What it does

You describe the outreach you want in plain English and the agent assembles a real-time journey around it — trigger or segment entry, the sequence of message steps, waits, and if/then branches on behaviour like opens and clicks. It builds the skeleton and hands it to you for review; nothing goes live without a person publishing it. It is the replacement for the older journey Copilot.

Key facts

Preview. Not supported for production, and Microsoft reserves the right to change the shape of it.
Requires version 1.1.65002.153 or later. Check under Settings > Versions before you promise anything in a scoping call.
Rolled to the NAM and OCE regions with the June 2026 update. Other geographies follow later — treat EMEA as unavailable until you have seen it in the target tenant.
English only at present. It will not parse a French or German prompt into a journey.
Turned on per environment in AI Hub, not via the ordinary Feature switches page.
Consumes Copilot credits. Capacity comes from Microsoft Copilot Studio — either prepaid message packs or pay-as-you-go — and has to be allocated to the specific environment in the Power Platform admin centre.
Tenant must be opted in to allow Anthropic as a subprocessor (Microsoft 365 admin center or Power Platform admin centre) before the agent will run.
It orchestrates across email, SMS, push and custom channels, and can key journeys off Dynamics 365 Sales events (opportunity created, deal stalled) or Customer Service events (case created, case resolved) where those apps are in the same deployment and connected.
It does not create segments and it does not write content. Segments must already exist, and messages must exist at least as drafts. The agent wires up assets, it does not author them.
Users need Marketing Manager or Marketing Professional (or equivalent rights that include journey creation).

When to use / skip

This is worth switching on where a marketing team is fluent in what they want but slow at the canvas — the classic "we know the flow, it just takes three days to build and QA it" problem. It shortens the gap between a brief and a first draft journey, and the draft is usually structurally correct, which is the boring part.

Skip it on any project that is still fighting the fundamentals. If segments are half-built, content is not approved, or the trigger model is unsettled, the agent has nothing to assemble and you will spend the workshop explaining why it produced an empty scaffold. Skip it too if the client is outside NAM or OCE, or works in a language other than English — that is a hard stop today, not a tuning problem. And be honest that preview status means you cannot put it on the critical path for a go-live. Most first-phase deployments do not need it; it earns its place in a second wave once the asset library is real.

Configuration decisions

Whether to enable it at all in the production environment while it is preview, or confine it to a sandbox for the first release and revisit at GA.
Copilot credit model — prepaid message packs versus pay-as-you-go — and how many credits get allocated to this environment versus other agents competing for the same pool.
Whether the tenant will accept the Move data across regions term and the Anthropic subprocessor opt-in. This is usually a legal or data-protection conversation, not an IT one, and it can take longer than the build.
Which roles get the agent. Giving it to everyone who can open the app produces a lot of half-finished draft journeys.
Whether Sales and Customer Service event triggers are in scope, which means those connections have to be configured first.
Your review gate: who signs off an agent-generated journey before publish, and whether that is captured anywhere other than a verbal nod.

Gotchas

Refreshing the browser mid-generation throws away the proposal. If the user has not already pushed it through Review and refine, the work is gone with no recovery.
The credit meter is easy to miss during a proof of concept. Nobody notices consumption until the pack runs dry and the agent silently stops being useful.
The Anthropic subprocessor opt-in is tenant-wide. It is not something you toggle for one environment, so a cautious central IT team can block the whole feature.
Region availability is by service region, not by where the client's office is. A UK client on a NAM-hosted tenant behaves differently from a UK client on an EMEA tenant.
The agent grounds on what exists in the environment. In a freshly provisioned sandbox with three demo segments it will look weak, and clients read that as the product being weak.
It is not a content generator. Expect at least one stakeholder to assume "AI builds the journey" includes writing the emails, and to be disappointed when it does not.

Consultant notes

Demo it against a realistic environment with genuine segments and drafted content. A sparse demo environment makes this feature look far worse than it is.
Get the tenant-level consents raised with the client's data protection lead at the start of discovery, not the week before the demo. That is the item that slips.
Push back on any suggestion of building the go-live journeys with a preview agent. Use it to accelerate drafting, then treat the output as a normal journey subject to the usual review and testing.
Before go-live, check the environment version is still at or above the minimum and that credit allocation survived any environment copy — a restore from backup will not carry Copilot Studio capacity assignment with it.
Set the expectation early that this is English-only. For multilingual marketing teams the agent is a tool for one person, not the whole team.

Worth revisiting when it moves past preview or lands outside NAM and OCE — both would change the advice quite a lot.