Flex-Web · Voglans, Savoie, France
Internal AI assistants: documents, permissions and testing
An internal AI assistant helps a team find information or prepare drafts from approved material. Start with one task and a controlled set of documents. The assistant must respect access rights, identify its sources and pass unsupported questions to a person.
Discuss your project ↗Which first use cases are useful?
These are illustrative examples. They do not prove a measured productivity gain or a deployment for a named client.
- Find the relevant paragraph in an approved internal procedure.
- Prepare a reply based on a validated service description.
- Summarise a request for the person who will handle it.
How should the documents be prepared?
Choose current, approved documents with an owner and a clear update process. Remove obsolete copies and separate information with different access rights. Identify which document takes precedence if two sources disagree.
If the required information does not exist, document the process first. An assistant cannot reliably recover a business rule that has never been written down.
How should access and sources work?
Only retrieve documents the user is authorised to read. Restrict the assistant’s tools and actions to its purpose. Treat text inside retrieved documents as source material, not as permission to change its instructions or access other systems.
Show the source behind an answer so a person can verify it. If sources are missing, contradictory or outdated, the assistant should say so and route the question for review. Source citations help checking; they do not make every generated sentence correct.
What should you test before release?
| Question or situation | Expected behaviour |
|---|---|
| Answer exists in an approved document | Relevant answer with a verifiable source. |
| Answer is not in the documents | No invented answer; request for clarification or review. |
| User requests a restricted document | No disclosure of inaccessible content. |
| Sources disagree | Explain the uncertainty and request a decision. |
| Document contains instructions to bypass rules | Keep access restrictions and task boundaries. |
| An important customer commitment is proposed | Require the agreed human approval. |
Can the assistant send messages or take actions?
Reading documents and taking actions are different permissions. Start with consultation or drafts. Sending a customer message, changing a price or updating a business record requires specific approval rules, traceability and failure handling.
Review real errors after launch and update the documents or workflow. Human oversight and maintenance remain necessary.
What budget and schedule should be planned?
The internal assistant can be included in Flex-Web’s €990 including French taxes CRM and automation package, within the scope defined in the quote. Optional support is €99/month including French taxes. Advanced integrations, document preparation, licences and model usage are scoped separately.
The schedule depends on the documents, permissions, connections and acceptance tests. It is agreed before ordering.
Questions before you start
Is an internal assistant the same as a public chatbot?
No. Its scope concerns internal users and authorised business information. A public chatbot has a different audience and disclosure boundary, which requires separate scoping.
Can it guarantee that every answer is correct?
No. Sources, access controls, test cases and human review reduce specific risks. Unsupported answers and consequential decisions still need a clear escalation path.