AI Models and Effort
Choose which AI model and how much effort Percy Assistant, Percy Copilot and workflow AI steps use, see what each choice costs, and route topics to a model of their own.
Features Configuration → AI Models (also linked from the AI Assistant tab) is where a business decides what its AI runs on. Perceny ships with a default model for every AI task, priced the way it has always been. This tab lets you pick a different model or a different level of effort for any task, and it shows the price of each choice before you commit to it.
The four tasksCopy a link to this section
Each card on the tab is one task, and each task is set on its own:
- Percy Assistant, text. The receptionist answering texts and web chat.
- Percy Assistant, voice. The receptionist on the phone. Voice is capped at medium effort, because a caller will not wait for a model to think.
- Percy Copilot. The assistant your team talks to inside Perceny.
- Workflow AI steps. The AI actions inside your workflows: classify, extract, write, translate.
A card shows the model the task runs on today, the effort, and whether that is the platform default or your own choice. Change Model opens the catalog.
Reading the catalogCopy a link to this section
The catalog is grouped by provider. Every model shows:
- What one unit costs on that model: a Percy reply, a Copilot message, a workflow step. These are the prices you will be billed, not list prices.
- What your last 30 days would have cost on that model, task by task, worked out from your own recent usage. A business that sends 400 receptionist replies a month sees the monthly figure for 400 replies.
- Which tasks the model is offered for. A model that is not suited to live calls is not offered for voice.
The platform default keeps the price your plan has always carried. Any other model is priced at its own cost plus the published margin, so a cheaper model costs less per unit and a stronger one costs more.
EffortCopy a link to this section
Effort is how long a model may think before it answers. Low is fastest and cheapest, High reasons more and costs more per answer. Not every model supports effort; the catalog says when a choice would be ignored. Voice never goes above Medium.
Copilot included messagesCopy a link to this section
Your plan includes a number of Copilot messages a month. That number is a budget, not a count: on the default model you get the plan's number; on a cheaper model you get more; on a dearer one, fewer. The Copilot pricing panel shows the number for the model you chose, up to four times the plan's count.
Topic rulesCopy a link to this section
Below the cards, Topic Rules route one kind of request to a model or effort of its own without changing the rest. A rule is a short topic in plain words ("pricing questions", "cancellations"), a model and an effort. When Copilot classifies a message as that topic, the rule's model answers it. Up to ten rules; the classifier reads the topic text as data, never as instructions.
What a platform admin controlsCopy a link to this section
A platform admin can narrow the catalog to an allow-list and set the margin a chosen model carries. A model that is removed from the allow-list after you picked it falls back to the default, and the card says so under Ignored.
In the staff appCopy a link to this section
There is no menu row for it. Open Automation & AI → AI Assistant (on the web or in the app's More tab) and pick AI Models under More AI Settings; it opens the same view as the web, for anyone with the features-configuration permission.