- Read
- Only the enquiry history allowed for this employee and organisation
- Draft
- Prepare a response without external change
- Send
- A separate permission; check recipient, content and required approval
- Verify
- Confirm the actual change in the external system
- Revoke or reconcile
- Revoked access blocks the action; reconcile an unknown outcome before retrying
What the team gains
Access to a specific function.
AI sees only the necessary source and permitted fields.
A useful draft.
An answer or proposed change can be checked before action.
Control over sending and writing.
Sensitive operations are performed after clear authorisation.
A verifiable result.
We check the system's actual state, not just a successful API response.
A tool connection is bounded by permissions
User and purpose
Identify the task, source and responsible owner.
AI proposal
Select an agreed function, not unrestricted access.
Permitted tool
Check allowed actions and effective user rights.
Business system
Verify the actual result and record errors.
Read authority
For example, search approved documents only.
Write authority
Sending, booking and record changes need separate rights and confirmation.
Example: first allow search in approved documents. Sending, booking or changing records needs separate rights and confirmation; a connected tool does not authorize every action.
From a useful answer to a controlled action
An assistant can find an order and prepare a suggested next step. But permission to read a record does not imply permission to change a price, send a message or refund a payment. Each action needs an appropriate interface, authorisation and a check of the result.
We start with one scenario and a limited set of data. An API or MCP alone does not prove that a process is ready or grant permission to write.
What pNet does
Checks compatibility
We select the AI client, system and specific function. We check the supported interface, version, plan restrictions, permissions and available fields. We do not promise a universal connector just because two products have APIs.
Defines permitted access
We associate each action with a user, organisation and role. We limit the objects and fields that can be read or changed. The system checks permissions before returning data; telling a model “do not show extra information” is not a substitute for access control.
Separates a proposal from execution
An assistant can prepare changes and explain them. For a significant action, we define the necessary confirmations, input checks and permitted limits. A document or an external service response does not become a new instruction with elevated authority.
Verifies the actual result
After a permitted action, we check the target system's state. A “done” message must be based on the result, not merely on a sent request. If the response is lost, the action cannot automatically be assumed to have failed and retried without a check.
An assistant can find an Odoo record or prepare a proposal to change it. These are different actions. An academy may let staff read a programme without granting permission to enrol a participant; a supplier may look up a specification, but that does not permit changes to price, order or service status; a Beauty salon cannot access a customer's private history merely from service information.
pNet limits access by user, company, necessary fields and the duration of permission. For writes, we define input validation, approval, result verification and how to investigate an unknown outcome. An API or MCP provides a technical interface; the customer's permission and acceptance remain separate.
It is sensible to start with permitted search and a draft. Business integrations provide reliable exchange; a model is connected only where its work is needed.
Important: a model's ability to propose an action is not permission to perform it. The AI provider, level of access and data shared are agreed separately.
A permitted-action record for a service employee
| Level | Read |
|---|---|
| What AI can do | Find a permitted document or status |
| Who confirms the result | User permissions and the source |
| Level | Propose |
|---|---|
| What AI can do | Prepare a draft answer or change |
| Who confirms the result | An authorised employee |
| Level | Change |
|---|---|
| What AI can do | Perform only an approved, limited action |
| Who confirms the result | The system checks access and logs the effect |
| Level | Verify |
|---|---|
| What AI can do | Read the state again after the operation |
| Who confirms the result | An employee resolves exceptions and discrepancies |
Examples of useful connections
A supplier's manager sees the relevant order and prepares a change request. Price, reservation, dispatch and refunds do not receive a single “do everything” button without their own permissions and conditions.
An academy administrator finds the status of a specific participant. Access to another course group's or organisation's materials is not disclosed through a general assistant answer.
How we check the integration
We agree examples of permitted reading, denied access and a selected action. We test the wrong object, another organisation, incomplete information, retries and service unavailability. Sensitive and financial operations receive a separate assessment; real effects are tested only with permission.
We hand over a description of permissions, limits, result verification and access revocation. After launch, pNet helps maintain the connection as interfaces change and develop only agreed capabilities.
Frequently asked questions
Compatibility is checked for the specific client and system. We choose a supported method rather than promise identical functions in every application.
Not by default. Permissions should match the selected users and actions; broad access is not a convenient substitute for design.
No. Important limits must be enforced by the tool and target system, and user confirmation must have a clear scope.
Yes. Reading and preparing a proposal let us test usefulness before enabling external actions.
The terms of the chosen service, account and function are checked during design. Zero retention or the absence of all transfers cannot be promised universally.
Connected services
Name one function you need
Tell us which assistant is in use, which system it should work with and what action is needed. Do not include passwords, keys or customer documents in an ordinary message.