Organization policies
In Configuration → Assistant, Policies tab, you decide how the assistant works for the whole organization. The settings you leave on Platform default follow the platform’s value. Click Save policies when you finish.
Model and effort
| Setting | What it decides |
|---|---|
| Default model | Which model the assistant works with for anyone who hasn’t chosen one in My assistant. |
| Default effort | How much it thinks before acting, from low to max. More effort gives better results on difficult requests, but it takes longer and uses more. |
| Models members can choose | The models shown in the panel’s model selector and in My assistant. None ticked = all. The default model has to be among the allowed ones. A model marked “not available in this assistant version” isn’t offered. |
| Fallback models | An ordered list (Fallback 1, Fallback 2…). If a member’s model is overloaded or unavailable, the turn is finished by the first model in this list that answers, and the reply says so in one line. Their own model is tried again on the next message. By default, No fallback. |
| Background model | The one that names conversations and writes the summaries and the compactions. By default, Platform default (Haiku 4.5). |
| Subagent model | For subagents that do not set their own model. By default, Same as the main model. |
Each person can choose a different model and effort for themselves, among the allowed ones.
Tools
The assistant works with seven tools: searching the catalog, managing credentials, opening and saving flows, applying changes to a flow, inspecting it, probing and running, and managing the account (schedules, environments, versions, skills, files, team…).
Tools that always require confirmation
The tools you tick ask for permission before acting. By default, only the one for managing the account. Nobody can remove these confirmations; each person can only add their own.
Disabled tools
The tools you tick can’t be used in your organization. If the assistant tries, it says so: it explains that the tool is disabled in the organization and that its administrator can enable it in Configuration → Assistant → Policies.
Allow the assistant to delete
It is off by default. That way, the assistant doesn’t delete anything: not flows, schedules, environments, skills, files or versions. When someone asks it to, it tells them where to delete it by hand.
If you turn it on, the assistant will be able to delete, with the person’s permission (which in Automatic mode is granted automatically) and within what their role allows. The assistant can’t change this setting.
Working on its own
| Setting | What it decides | Default |
|---|---|---|
| Recurring tasks enabled | Whether people can give it recurring tasks. Off, they can’t be created and existing ones don’t run. | On |
| Max. recurring tasks per user | How many tasks each person can have. | 5 |
| Automatic review of failed runs for the whole organization | Whether the assistant reviews failing flows and leaves fix proposals. Off, nobody can turn it on. On, each person can turn it off for themselves. | On |
| Max. automatic reviews per user per day | The maximum number of failures it reviews per day per person. | 10 |
| Cost-saving suggestions for the flows of the whole organization | Whether the assistant leaves suggestions to make flows cheaper in the bell. Off, nobody gets them. On, each person can turn them off for themselves. They don’t use up usage. | On |
Tasks and reviews count towards each person’s usage.
Rate limits
They stop one person or one flow from monopolizing the assistant.
| Setting | Default |
|---|---|
| Assistant messages per user (per minute) | 40 |
| Assistant messages per flow (per minute) | 20 |
| Tool calls per user (per minute) | 60 |
| Tool calls per conversation (per minute) | 120 |
You can lower them, but not raise them above the platform’s maximum. When someone goes over them, they have to wait a few seconds to carry on.
Heads up / Good to know notes
An observer reads each turn and leaves one line above the box with what the person may be missing: a decision the assistant made for them, a doubtful result… See “Good to know” notes. Each check is one more model call, which counts towards that person’s usage.
| Setting | What it decides |
|---|---|
| ”Heads up / Good to know” notes | Whether there are notes in your organization. Off, nobody can turn them on. On, each person can turn them off for themselves in My assistant. |
| Check every N steps | Every how many assistant tool calls it checks. Empty = 6. |
| Check when the turn ends | Whether it also checks at the end of each reply. |
| Observer model | Which model does the checks. By default, Platform default (Claude Opus 5.5). |
Compatible gateway
Allow the compatible gateway lets the assistant’s credential be an Anthropic-compatible gateway (DeepSeek, LiteLLM…). It’s experimental and not supported by Anthropic. Off, it cannot be saved and a saved one stops being used.
Identity
Identity guard (does not reveal what the assistant is built with): if you turn it on, when someone asks which model or company the assistant is built with, it answers that it is the NappAI assistant. By default, it follows the platform’s setting.
Organization memory
The conventions can also be changed in Assistant memory, under Settings, which is always visible.
The Memory tab (next to Policies) keeps your organization’s conventions as free text, up to 4,000 characters: how you name flows, tone, preferred models, your glossary… The assistant receives them at the start of every conversation of any member, before each person’s own instructions (in My assistant). Every member can read them; only an admin can change them.
They are preferences, not permissions: they don’t override the policies on this page or the confirmations. The usual folder and credential belong to each person, and even then the assistant always asks.