Can we limit which tasks the AI does for each team? Yes, with four dials, set in the system rather than left to the prompt.
Short answer. Yes, if the product has roles, groups and per-item access, but enforce the limits in the system and not only in the prompt. There are four dials: which models and assistants a team can use, which document collections it can query, which tools and connectors an assistant may call, and which actions need a person’s approval. OWASP warns that rules written into a system prompt may not be honoured and can be bypassed by prompt injection, and that harm from AI agents comes from excessive functionality, permissions and autonomy. Write the rules down per team first, then check each in the product.
01
Four dials
Dial
Example for a firm
How it is enforced
Models and assistants
The accounts team may use the finance assistant; HR may not use cloud models
Per-model access grants; only administrators add cloud models
Documents
The litigation team’s collection is invisible to marketing
Per-collection access by group, with permission-aware retrieval
Tools and connectors
An assistant may read the calendar but not send email
An approved tool list, read-only scopes, and running in the user’s own context
Actions
A person approves before anything is sent, filed or deleted
Relying on an answer without checking the cited source
Everyone
Summarise their own meetings, format documents
Adding a new connector (administrator only)
Pasting passwords or client identifiers into prompts
03
Enforce it in the system
OWASP’s entry on sensitive information disclosure says that restrictions placed in the system prompt about the data types a model should return may not always be honoured and could be bypassed through prompt injection. Its entry on excessive agency lists the remedies. Minimise the tools an agent can call, what each tool can do and the permissions it holds. Run tools in the user’s context. Require user approval for high-impact actions. Enforce authorisation in downstream systems rather than relying on the model to decide. The GOV.UK AI Playbook states the matching rule for data: make sure the model only has access to the data the user of the model should be able to access.
Open WebUI shows what this looks like in a product: roles, additive permissions and groups, with models, knowledge bases and tools private to their owner until shared, and sharing granted to groups or individuals. A team boundary should be a permission, not a polite instruction.
04
Review and prove it
Use the product’s access-preview view if it has one. Open WebUI offers an admin-only Preview Access view for each user and group.
Test with a real account from each team: ask for something it should not reach.
Review group memberships every quarter.
Log denied attempts and read them.
Revisit the table whenever a tool, connector or model is added.
05
What ANVE says
ANVE has two roles today, administrator and user. Teams, finer roles and single sign-on are not on ANVE yet (see /security). Cloud models are added by an administrator, and their answers are labelled; a policy that blocks confidential data from cloud models is not on ANVE yet. Connectors, for example to Google Workspace or Zoho, work only if an administrator approves one, and every call is logged on the box.
The tools an agent can use must be approved by an administrator, a tool that changes is switched off, and every tool call is logged. An agent can take up to seven tool steps per request, and a claim made without a successful tool call is challenged. Every action is owned and every model call is logged by user.
Whether access can be limited by maker or by model today is not stated here. ANVEAI confirms it in writing before a paid pilot; bring your own team table to the demo call.
06
Questions people ask
Can we stop one team using cloud models?
Only if the product restricts models per group. In Open WebUI a model can be restricted with access grants. Ask any vendor whether a restriction is enforced by the system or only by an instruction to the model.
Will limits make the AI less useful?
They narrow it to the task, which usually helps. Start tight and widen after a review, not the other way round.
Who decides what each team may do?
The person who owns AI in the office, with team leads, and signed off by a partner or director. See who should own AI in a small office.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
