---
title: "Governing AI Without a Vendor Cloud"
description: "How to govern AI use — identity, fine-grained policy and fleet-wide audit — when the control plane runs on your own machines instead of a vendor cloud."
url: "https://royalsoftworks.com/resources/enterprise-ai-governance/"
source: "https://royalsoftworks.com"
format: "markdown"
note: "Markdown rendering of the HTML page at `url`. Same content, same canonical URL."
---

[All resources](https://royalsoftworks.com/resources/)

# Governing AI Without a Vendor Cloud

## A Policy, Identity, and Audit Plane That Runs on Your Machines, Not Ours

**A Royal SoftWorks whitepaper · AssistantGeneral**

---

## Abstract

Enterprises adopting AI face a control problem: _who may use which model, against which data, with which tools — and what did they actually do?_ Cloud AI vendors answer this with governance over _their_ surface, hosted in _their_ cloud. For regulated and security- conscious organizations that is the wrong shape twice over: the data plane is outside the perimeter, and the controls cover only the vendor's own product. AssistantGeneral inverts the model. The agent runs on the employee's device, and the governance plane — policy, identity, and audit — runs on the **customer's own admin server**. This paper describes that governance plane and the principles behind it.

---

## 1\. The requirement

A defensible enterprise AI deployment needs to answer, per user and per group:

- **Which models** may be used (and which are forbidden)?
- **Which tools, connectors, MCP servers, and messaging bridges** are permitted?
- **Which knowledge bases** can each person reach?
- **Which settings** can the employee change, and which are administrator-locked?
- **What can leave** — can the user export or wipe data; can an administrator reach the machine remotely?
- **What happened** — is there an audit trail?

And it must answer all of this **without a mandatory dependency on the AI vendor's cloud**.

## 2\. The shape: customer-hosted control plane

AssistantGeneral's enterprise build runs an embedded admin server — **the customer's server, on the customer's infrastructure**. Employee installs are managed clients that fetch their policy from it. There is no Royal SoftWorks cloud in the loop: both the data plane (on-device inference and retrieval) and the control plane (this admin server) sit inside the customer's boundary. A license _role_ — administrator versus managed user — decides whether a given install hosts the server or connects to it.

The console speaks only TLS, and each desk pins the console's certificate on first contact. Two things touch our licence server: activating a licence, which needs the internet once, and a desk that signs in with its own licence key, which is checked there. Desks enrolled with an access token issued by your console need nothing from us.

## 3\. The policy engine

A managed user's capabilities are described by a single policy object that gates **every** surface of the product, including:

- **Models** — the allowed providers (e.g. pin everyone to the company-hosted/on-device model), and whether the user may change the endpoint or the system prompt.
- **Tools** — an explicit allow-list, and whether the user may add tool directories.
- **Workflows** — which workflows are allowed; whether the user may create or delete them; whether scheduling and event triggers are permitted; whether the builder is available.
- **MCP servers**, **messaging bridges**, and **connectors** — allow-lists and whether the user may add their own or sync them.
- **Knowledge bases** — which the user may access; whether they may add or delete documents.
- **Settings** — exactly which option groups remain editable, plus a list of administrator-locked keys.
- **UI surfaces** — which gateable regions of the interface the user may even see.
- **Voice** — text-to-speech and speech-to-text on or off.
- **Workspaces and practice packs** — which profession workspaces (law office, judiciary, hospital, clinic, HR, recruitment) and which jurisdiction packs a group receives, and whether it may edit packs.
- **Delegation and verification** — whether helpers (sub-agents) and the legal citation check may run.
- **Engines, downloads, backups and updates** — whether local engines may be installed, models downloaded, backups configured and the application updated.
- **Data** — whether the user may export, import, or delete all data.
- **Action limits** — per connector (read or write), per MCP tool, and per website.
- **Network egress** — open, an allow-list of hosts, or none at all (Offline mode). Every outbound connection the app opens passes this one gate, and the hosts reached are counted in a daily record on the machine.
- **Remote access** — whether an administrator may broker tool execution on the user's machine (off by default). There is deliberately no remote database query.

### Fail-closed, versioned, and most-restrictive-merge

The policy engine is designed to be safe by default:

- **Fail-closed normalization** — a missing or malformed field becomes the most restrictive value (deny), so a partial or corrupt policy payload can never silently _widen_ access.
- **Schema versioning** — every policy record states the schema it was written against. A capability introduced in a later release is granted to an older policy rather than read as a deny, so upgrading the fleet never silently strips features from existing groups — while a malformed field in a current policy still fails closed.
- **Most-restrictive merge** — when a user belongs to multiple groups, capability flags are AND-ed, allow-lists are intersected, locked keys are unioned, limits take the smaller value, and the earliest expiry wins. A "deny" in any group denies. (Knowledge- base _data_ grants are resolved separately and remain additive — capabilities tighten, data access widens — which is the correct, intuitive behavior.)
- **A usable managed default** — when an administrator hasn't authored a specific policy, managed employees get a restrictive-but-workable baseline: pinned to the company model, core/security/data settings locked, bulk export and remote control off — while everyday work (tools, workflows, knowledge bases, voice) keeps functioning.

### Always in force

A managed desk stores the policy it was last given, sealed against tampering with a key held by the operating system's credential store, and applies it at every start — whether or not the console is reachable. A desk that has never received one runs the managed default; a stored policy that has been deleted or altered locks the desk down to deny-all. Each desk reports which policy it is running, so the console shows who is compliant.

## 4\. Identity: SSO and SCIM

The admin server integrates with the organization's identity provider through **SSO (OpenID Connect)** for authentication and **SCIM** for user and group provisioning — so joiners, movers, and leavers flow from the IdP into AssistantGeneral's users, groups, and therefore policies automatically.

During the beta both are enabled per organisation as part of a pilot, configured with us, while the self-service configuration screen is completed; they are off in a default install. SAML 2.0 is not offered until XML signature validation is in place — the endpoint refuses rather than trust an unverified assertion. There is no LDAP or Active Directory sign-in yet.

## 5\. Shared knowledge with access control

Administrators can publish shared knowledge bases and control, per group, who may reach each one. Capability policies tighten what a user can _do_; knowledge-base grants govern what data they can _see_ — managed independently and correctly.

## 6\. Audit, and nothing behind the employee's back

Actions are recorded to an audit trail and aggregated for review, giving the organization the "what did they actually do?" answer that governance requires — and that regulated environments mandate.

- **Always on for organisation installs**, and not switchable off there.
- **Content-free** — it records actions, not prompts or documents.
- **Tamper-evident** — each entry is chained to the one before it, and each desk's chain is forwarded to the console with a continuity check, so a missing stretch shows.
- **The agent's own actions** are recorded alongside the people's.
- **A read-only auditor role** can review it without being able to change anything.

The same rule applies in the other direction. Anything an administrator does on an employee's computer — running a tool, pushing or fetching a file, changing settings or policy — is recorded on that computer and shown to the person who uses it, including the attempts that were refused. Nothing the administrator does there is silent.

## 7\. Tailored interfaces: workspaces and surfaces

Beyond security gating, the product shapes itself to the job. Each chat can open in a **profession workspace** — law office, judiciary, hospital, clinic, HR or recruitment — with its own object, document scope, everyday actions, evidence rail and produced document, and a **practice pack** for the jurisdiction. Administrators decide which workspaces and packs each group receives, and separately which interface **surfaces** a user may see at all. Presentation (what a workspace puts on screen) and entitlement (what a user may use) are deliberately different levers: a workspace is never a permission, and a policy denial always wins.

Three of the six — judiciary, HR and recruitment — are high-risk uses under Annex III of the EU AI Act. Their workspaces carry a non-dismissible notice that a person decides, and their practice tools run through the ordinary chat so every request is in the audit trail.

## 8\. Why this matters

For the buyers who most need governed AI — legal, healthcare, finance, government — the governance plane is not a checkbox; it is the precondition for deploying at all. By making that plane fail-closed, group-aware, identity-integrated, audited, and — critically — **hosted on the customer's own server with no vendor cloud in the loop**, AssistantGeneral offers a control story that cloud-only AI products structurally cannot match.

---

_To see the policy engine, audit and role-tailored UI running on your own infrastructure — and company sign-in against your identity provider in a pilot — contact Royal SoftWorks._

## Evaluate AssistantGeneral

See the product this paper describes — a private, local-first AI agent with an enterprise control plane that runs on your own infrastructure.

[Explore AssistantGeneral](https://royalsoftworks.com/products/assistant-general/) [Talk to us](https://royalsoftworks.com/contact/)

<!-- verse:royalsoftworks
     Mateju 6:5
     I kad se moliš Bogu, ne budi kao licemeri, koji rado po zbornicama i na raskršću po ulicama stoje i mole se da ih vide ljudi. Zaista vam kažem da su primili platu svoju.
     https://www.wordproject.org/bibles/verses/serbian/02_prayer.htm
-->
