Power Automate licensing decision guide

What it does

Takes a client from "we want to automate this" to a defensible licence mix. Power Automate splits into user licences, which entitle a person, and capacity licences, which entitle an automation — and almost every argument about Power Automate cost comes from someone applying the wrong one of those two models to their scenario.

Key facts

Two families. User licences (Power Automate Premium, Trial, Free) attach to a person. Capacity licences (Process, Hosted Process) attach to an automation: a cloud flow, a machine, a hosted machine or a hosted machine group.
Seeded rights come with Microsoft 365, Dynamics 365 and other Power Platform licences. They cover standard connectors and, in general, use within the context of the licensed app. Premium connectors are not part of that.
Power Automate Premium gives a user unlimited cloud flows within action limits, standard, premium and custom connectors, one attended bot on their registered machine, process mining, and an AI Builder credit allowance. The published daily action allowance is 40,000 per user.
A Process licence is allocated to a cloud flow or to a standard machine. On a cloud flow it grants premium and custom connector use regardless of the owning or triggering user's licence, plus a published 250,000 actions per day. The flow must be in a solution for a Process licence to be assigned.
Up to 10 Process licences can be stacked on one cloud flow, each adding another 250,000 daily actions.
Allocated to a machine, a Process licence becomes one unattended bot running one desktop flow at a time. Parallel unattended runs need one Process licence per concurrent run. Up to one cloud flow inherits the Process plan from a related machine or machine group with Process capacity.
Hosted Process is a superset of Process — it covers hosted machines and hosted machine groups, and each hosted bot can also act as an unattended bot.
Capacity licences don't replace user licences. Registering a machine still requires a Power Automate Premium user, triggering a desktop flow from a cloud flow still needs the connection user to hold desktop flow entitlement, and several monitoring pages in the portal only render for Premium users.
Each licence type carries its own Dataverse database and file storage entitlement, and Premium carries a process mining data storage allowance capped at tenant level.

When to use / skip

Start from the automation, not the person. Ask who or what triggers it, what it touches, and whether a human is present. A flow triggered by a person, using only standard connectors, inside the app they're already licensed for, is usually already covered by seeded rights and you should say so rather than selling something. A flow that touches a premium connector needs premium entitlement from somewhere — the triggering user's Premium licence, or a Process licence on the flow itself.

The Process licence is the one consultants under-use. Where a single high-volume flow serves hundreds of people, licensing the flow is dramatically simpler than licensing everyone who might trigger it, and it survives the flow owner leaving. Where a handful of power users each build and run their own premium automations, per-user Premium is the cleaner answer. When a client insists on capacity licences alone to avoid buying user licences, that doesn't work — creation, machine registration and attended running all need real users.

Configuration decisions

Which existing licences the client already holds, and what those seed. This is the first question and it frequently changes the answer entirely.
Per-flow or per-user for each premium scenario: does the entitlement need to follow a person, or sit permanently with an automation?
Expected daily action volume per flow — actions per run multiplied by runs per day — measured against the relevant allowance, and whether stacking is needed.
Attended, unattended or hosted RPA, and the peak number of parallel desktop flow runs. That number is the number of bots.
Whether to turn on licence auto-claim so premium licences are assigned when someone actually needs one rather than pre-allocated on a guess.
Whether pay-as-you-go suits any workload with unpredictable or bursty volume better than a fixed licence.

Gotchas

Seeded rights are scoped. A flow that starts inside the licensed app and then writes to a third-party system through a premium connector has left the seeded boundary, and clients rarely see the line being crossed.
Process licences only attach to flows in a solution. Discovering this after a client has built forty premium flows in the default environment is a genuinely painful week.
Action allowances are per licence, not per tenant, and they're daily. A single badly written loop can consume a day's allowance before lunch.
Trigger frequency drives cost more than flow count. Nobody counts the polling trigger checking every minute, forever.
Unattended RPA still depends on a Premium user having registered the machine. Buying Process licences without any Premium users leaves you unable to set anything up.
The published request limits sit under transition-period arrangements that Microsoft has adjusted more than once. Always check the current allocations page before committing numbers to a proposal.

Consultant notes

Build the licence model from a table of automations, not a headcount. One row per flow with trigger, connectors, volume and a human-present flag gets you to an answer in an hour and gives the client something to challenge.
Never quote prices from memory or from a KB. Get current figures from the pricing page or the client's licensing partner, and put the date on the estimate.
Flag the storage entitlements. Dataverse capacity arrives bundled with these licences and is quietly consumed by logging tables and attachments; it's a recurring source of unexpected top-up purchases.
Where a client is already deep into Dynamics 365 or Microsoft 365 E-series, do the seeded rights analysis properly before proposing anything. You will sometimes find they're most of the way there already, and saying so buys a lot of credibility.
Revisit the model after three months in production. Real action counts are always different from the estimate, usually because of retries and polling.

Check the current licensing pages before every proposal — this is the fastest-moving thing in the product and the most expensive to get wrong