Digital Virtual Office
An organization of AI employees that plans, delegates, uses EGKits tools, searches approved knowledge and takes part in voice meetings under human authority, with approvals, minutes, budgets, KPIs, audit and tenant isolation.
Digital Virtual Office
The Digital Virtual Office (DVO) gives each company its own organization of AI employees. They sit in departments, report to managers, hold job descriptions and skills, and work through the same EGKits services a person uses. A person gives an instruction; the AI employee plans, may delegate parts to its team, asks when information is missing, and submits a result that the person can review, rate and act on.
Download the complete guide as PDF in English and Arabic from Digital Virtual Office Guide. Status flows for every record are on Digital Virtual Office Workflows.
#What makes it safe for a business
| Control | What it means |
|---|---|
| Human authority | An AI employee can do only what its grants, its autonomy level and the requesting person's live permissions all allow. High-risk and irreversible operations always wait for a person's approval, bound to the exact operation by a SHA-256 fingerprint. |
| Tenant isolation | Every DVO record carries its tenant. The tenant comes from the signed-in workspace, never from the request, and background jobs re-resolve it before touching data. |
| Real work only | Results come from real model and tool calls. In production a route to the stub provider fails the work with a clear error instead of reporting a simulated success. |
| Measured cost | Budgets for the office, a department or an employee are checked before each model call. Usage is recorded per call and summarized in the Cost Center and seven KPIs. |
| Full traceability | Each operation writes an audit event with actor, employee, target, authorization result and correlation id. |
The DVO does not work in the background by itself. Work starts only when a person asks, starts a task, starts plan steps, asks a question in a channel or speaks in a meeting, or when a meeting of AI employees only reaches the time a person scheduled. Every page shows the real status of that work.
#Core concepts
| Concept | What it is |
|---|---|
| AI employee | A named role with a department, a manager, a job description, skills and a versioned configuration. |
| Configuration version | An immutable snapshot of instructions, model, autonomy, data scopes, memory scope and tool categories. |
| Autonomy level | L1 read only, L2 low-risk changes, L3 may request high-risk work through approval. |
| Work request | One execution of an instruction by one AI employee, with steps, usage, result and transcript. |
| Delegation | A child work request given to a member of the employee's team, within depth, size and cost limits. |
| Approval | A person's decision on one exact operation, identified by its fingerprint. Used at most once. |
| Executive plan | Steps proposed by a planning AI employee. Only the requester can start the chosen steps. |
| Task and project | Board work with an owner, status, assignee, dependencies and deliverables, grouped in projects. |
| Knowledge document | A versioned file or text that AI employees may search only when an access rule allows it. |
| Memory entry | A note kept for an employee, department, project or the office. Context, never a business record. |
| Runtime | The tenant's logical AI runtime: a queue, limits, health and lifecycle. |
| Avatar and voice | The 3D figure and the English and Arabic voices of an AI employee. Always labelled as AI, never presented as a person. |
| Meeting | People and AI employees meeting by voice or text in a 3D room, with an agenda, limits and a transcript. |
| Meeting turn | One AI answer in a meeting. Each turn is a real work request with the permissions of the person who spoke. |
| Minutes and action items | Results the AI drafts after a meeting. A person approves the minutes, confirms decisions and approves action items. |
#Getting started
- Platform: enable the
dvoapp in the tenant's subscription. The reconcile job provisions the tenant's runtime within 15 minutes by default. - Tenant administrator: assign the DVO permissions to the right roles. They are seeded into the tenant Admin role.
- Configure AI providers and models in the AI Assistant settings. DVO uses the same model router and never reads provider keys.
- Open Settings: choose the CEO user, the default autonomy, support access, retention and your notifications.
- Open Organization: preview and apply the EGKits standard template, or build departments yourself.
- Review each AI employee, adjust the configuration and activate it. Template hires start as drafts.
- Grant tools on each profile, set tool policies on Tools and create budgets on Budgets and usage.
- Start the first work request, or ask the Chief of Staff for a plan from the Command Center.
- For voice, open AI Agent › Settings › Providers, and on the Azure AI Speech card paste the key, set the endpoint to
https://<region>.api.cognitive.microsoft.com, turn it on and press Test connection. Then use the speech test on the Voice card in Settings. Set the meeting rules on the Meetings card, then schedule the first meeting.
#Permissions
| Permission resource | Protects |
|---|---|
Dvo |
Overview of the office |
DvoOrganization |
Organization, departments, reporting lines and templates |
DvoEmployees |
AI employees, profiles, configuration versions, avatars and the Virtual Office floor |
DvoExecutions |
Work requests, the Command Center and executive plans |
DvoApprovals |
The Approval Center. Approve and Reject are separate actions |
DvoTasks |
Tasks, the board and projects |
DvoCommunications |
Channels, messages and voice conversations |
DvoMeetings |
Meetings, the meeting room, minutes, decisions, action items and Meeting Analytics. View, Create, Update, Approve and Export are separate actions |
DvoKnowledge |
The Knowledge Center and memory |
DvoTools |
Tool catalog, policies and grants |
DvoBudgets |
Budgets and usage, budget alerts and the Cost Center |
DvoReports |
The KPI Center and DVO reports |
DvoRuntime |
Runtime status, queue, suspend, resume and requeue |
DvoSettings |
Office settings |
DvoAudit |
The audit log and full transcripts of other people's work |
DvoPlatform |
Root workspace only: AI Runtime Operations for all tenants |
Subscription limits (employees, departments, concurrency, monthly tokens and cost, providers and models, knowledge storage, retention, monthly speech minutes, concurrent meetings and meeting participants) come from the plan, then an optional per-tenant override set only by the platform. Reaching a quota never deletes data and never stops running work.
#The organization
Applying the EGKits standard template creates the Executive Office and seven departments with twenty AI employees, in English and Arabic. Department heads report to the CEO (a person chosen in Settings) and have a dotted line to the Chief of Staff for coordination.
| Department | Head | Members |
|---|---|---|
| Executive Office | Chief of Staff (DVO-01) | — |
| Technology & EGKits Development | Chief Technology Officer | Backend & Architecture, Frontend & UX, QA & Testing, DevOps & Infrastructure |
| Finance & Accounting | Chief Financial Officer | Accounting & Billing Specialist |
| Sales & Business Development | Chief Commercial Officer | Business Development Researcher, Sales & Proposal Specialist |
| Marketing & SEO | Chief Marketing Officer | SEO & Content Specialist |
| Customer Support & Operations | Chief Operating Officer | Customer Support Specialist, Implementation & Project Coordinator |
| Product Management & Innovation | Chief Product Officer | Business Analyst & Product Researcher |
| HR & Administration | Chief Human Resources Officer | HR & Administrative Specialist |
- Every AI employee has exactly one solid-line manager. Reporting cycles are refused; up to five dotted lines add coordination without authority.
- Dragging a card onto a new manager in the chart is a real move that the server re-checks. The chart supports arrow keys and Enter and mirrors in right-to-left.
- Applying a template adds only what is missing, keeps renamed and archived employees and checks limits for the whole set before saving once.
#AI employees and autonomy
The profile holds the job description, responsibilities, restrictions and skills, the configuration with its versions, the granted tools, performance, recent work and memory. Only people edit configuration; any request running as an AI employee is refused.
| Level | Read tools | Low-risk changes | High-risk and irreversible |
|---|---|---|---|
| L1 Read only | Yes | No | No |
| L2 Low-risk changes | Yes | Yes, when explicitly granted | No |
| L3 Request high-risk work | Yes | Yes, when granted | Only through a person's approval |
Before every tool call the server checks, in order: the tenant's tool policy, the employee's configuration, an active grant, the tool's declared permissions, the app entitlement and data scope, the requesting person's live permissions, and the autonomy level. Writes in restricted areas such as money, external communication and deployment always need L3 and approval.
#Tools
AI employees act only through tools: existing EGKits commands and queries marked as AI tools, so the mediator runs the same permission, validation, audit and tenant checks as for a person.
| Built-in tool | What it does |
|---|---|
dvo_delegate |
Gives part of the work to a member of the employee's team |
dvo_request_info |
Asks the requesting person a question; the work waits for the reply |
dvo_escalate |
Raises a risk or decision to a person with facts and a recommendation |
dvo_submit_result |
Submits the final result |
dvo_create_task |
Creates a task owned by the human requester; it never starts by itself |
dvo_search_knowledge |
Searches the Knowledge Center within the employee's access rules |
dvo_remember / dvo_recall |
Writes or recalls memory within the configured memory scope |
Module tools cover CRM (including creating a follow-up activity), Sales, General Ledger, HR, Newsletter and CMS. Each requires the module's own permission, and list results are capped at 50 rows per call.
#Work requests
A work request is one AI employee working on one instruction, with the authority of the person who started it. Every model call, tool call, delegation, question and result is a checkpointed step, so work resumes from the last completed step after a restart. Each request is bounded by model calls, steps, tokens, estimated cost and a 30-minute deadline.
| Status | What you can do |
|---|---|
| Queued | Wait, or cancel |
| Running | Watch the steps; cancel if needed |
| Waiting for delegates / input / approval | Open the delegations, reply on the request page, or decide in the Approval Center |
| Blocked | Fix the budget, quota or configuration cause, then Resume |
| Succeeded | Review, rate and act on the result |
| Failed | Read the coded reason, such as provider.failed; start again |
| Cancelled | Nothing; history is kept |
#The Approval Center
When a tool needs approval, the work pauses and a request appears in the Approval Center. A person who holds DvoApprovals Approve or Reject and every permission the operation uses decides with a comment. After approval the call goes through the full authority check again and runs only if the fingerprint still matches. An AI execution can never decide, irreversible operations need someone other than the requester, and pending approvals expire after 72 hours by default.
#Plans, tasks and communications
- Executive plans: a person describes a goal; the planner may only propose steps or ask a question. Only the requester picks and starts steps, and each started step becomes a work request and a board task.
- Command Center: live counts of running, queued, waiting and blocked work, today's results, monthly cost against the budget, approvals and questions waiting for the viewer, plans and activity.
- Tasks: a task runs only when a person starts it, now or when its dependencies complete. Its status then follows the real execution. Board moves use the same rules as buttons and offer only valid drop targets.
- Projects group tasks with members, a status and a shared channel.
- Communications: general, department, project, task and direct channels. Asking an AI employee in a channel starts real work; the answer is posted only when that work finishes.
#Knowledge and memory
The Knowledge Center holds documents AI employees may consult. Text, Markdown, CSV, JSON, XML, YAML, HTML and Word files are indexed; only the current version is searched. An AI employee retrieves a document only when an access rule names it, its department or all employees, and the person it works for may read the document. The search tester shows exactly what a chosen employee would receive.
Memory keeps short notes that help AI employees work consistently. It is context, never an authoritative business record. Entries written by an AI employee stay unverified until a person checks them; people can inspect, correct, verify and delete entries, and deletes are audited. Entries expire, and each AI employee has a cap on active entries; at the cap it asks a person to review its memory and nothing is deleted automatically.
#Avatars and voice
Each AI employee has a 3D avatar and a voice in English and Arabic. The avatar is always labelled as an AI employee and is never presented as a person.
- The catalog has 22 Microsoft Rocketbox characters (MIT licence) with facial expressions and shared animations. Each of the twenty standard roles has its own avatar and default voices.
- The Avatar and voice tab of the profile shows a live preview. A person may choose another catalog avatar or voice; the change is versioned and audited. A request running as an AI employee cannot change any employee's appearance or voice.
- The portrait is shown instead of 3D when the device has no WebGL, the file fails to load, the person asks for reduced motion or the page is hidden.
- Talk by voice on the profile or in the Virtual Office opens a voice conversation. It is the same private conversation as in Communications: each question starts real work with the asker's permissions, and the answer is spoken, with moving lips, only when that work finishes.
| Part | How it works |
|---|---|
| Speech provider | Azure AI Speech. The key is saved with the other AI provider keys and stays on the server. |
| Recognition | Runs in the browser only after the person presses Speak and allows the microphone, for one question. The browser gets a short-lived token, never the key. Audio is not recorded. |
| Speech | Text to speech runs on the server and returns audio with a lip-sync timeline. Markdown, links and code are not read aloud. |
| Languages | English (en-US) and Arabic (ar-EG), detected automatically or chosen by the person. |
| Quota | Monthly speech minutes, metered once per spoken answer. When the quota runs out, speech stops and the conversation continues in text. |
Without an Azure AI Speech key, a development host uses a silent demo voice: the lips move, nothing is spoken or recognised, every page says so and it is never billed.
#Meetings
People meet AI employees by voice or text in a 3D room: one to one, with a few of them, or with all of them. Each AI answer is a real work request started with the permissions of the person who spoke, and each AI employee answers from its own role.
| Meeting | What happens |
|---|---|
| One to one | A person and one AI employee. The employee answers every turn. |
| A few AI employees | A person names or picks who should answer; otherwise the most relevant ones answer, within the responder cap. |
| All AI employees | Invite all AI employees. Only the most relevant ones answer each question. |
| People together | People in the room can start a voice call among themselves through the platform call service. |
| AI employees only | Scheduled by a person and run by an AI moderator in rounds, with an agenda and hard limits. Needs the workspace policy. |
- Scheduling: Meetings shows Live now, Upcoming, My meetings and Past. New meeting sets the type, objective, start and duration, AI employees, people, moderator, agenda with AI turns per topic, privacy, language, transcript, recording, limits and repetition. A choice the workspace policy forbids is disabled with the reason.
- Meeting profile: the Meetings tab of each AI employee sets whether it may take part and moderate, its longest answer (30 to 400 words), who may invite it and its preparation instructions.
- The room is laid out face to face, at a small table, in a large room or as an auditorium. A person speaks only after joining with consent, through the microphone or by typing, and may pick who should answer. Up to 3 AI employees answer each turn by default (at most 5).
- An answer appears only when its work has really finished; when it fails, the transcript says the AI employee did not answer. Read answers aloud plays each answer in the employee's voice, and Interrupt stops answers still being prepared.
- Meeting state is kept in the database, so a restarted server or a reconnecting browser continues the meeting. A job runs every minute, recovers stranded turns and ends a meeting after 30 idle minutes by default.
- AI employees only: needs Allow meetings with AI employees only in Settings, an AI moderator and an agenda. It starts at its time, runs with the permissions of the person who scheduled it and stops at the workspace limits (12 AI turns and 200,000 tokens by default).
What is said in a meeting is a question, not an authorization: it cannot grant permissions, approve requests or change an AI employee's configuration. In meetings AI employees use only read tools and propose actions instead of doing them, and the moderator never assumes they agree. Live meetings are never ended because of a quota.
#Minutes, decisions and action items
The meeting record turns what was said into reviewed results. AI employees only propose; a person decides.
| Result | Rules | Statuses |
|---|---|---|
| Transcript | Append-only, searchable and exportable as plain text. When the workspace does not keep transcripts, the lines are removed when the meeting ends and no minutes are drafted. | — |
| Recording | Off unless the workspace allows it. Needs a kept transcript and everyone's consent on joining; the room shows the indicator. Only metadata is stored, never audio. | On, Off |
| Minutes | Summary, topics, each participant's own position, proposed decisions, open questions and risks, drafted by one AI participant as real work and checked against a schema. Draft again or write them by hand. An approved version never changes. | Draft, Approved, Replaced |
| Decision | Exists only when a person confirms it, from a transcript line, a proposal or their own words. Withdrawn decisions stay in the history. | Confirmed, Withdrawn |
| Action item | Proposed by AI or added by a person, then reviewed. Approving needs an assignee, an AI employee or a person. An approved item for an AI employee becomes exactly one task linked to the meeting: right away when the workspace creates meeting tasks automatically, otherwise when the reviewer chooses Create task. Restricted items (financial, external, production or destructive) create tasks that wait in the Approval Center. | To review, Approved, Rejected, Task created |
The Follow-up tab tracks the tasks and suggests a follow-up meeting while items are open. Invited people get a meeting invitation; the owner and the human moderator are told when minutes wait for review; the owner is told when a meeting task's real work completes.
Decisions are never inferred from AI suggestions, and nothing in the minutes is a decision until a person confirms it.
#The Virtual Office and Meeting Analytics
The Virtual Office, second in the DVO menu, is a 3D floor with one zone per department and each AI employee at a desk, next to a list that can do everything the floor can. Under each avatar it shows real activity: Available, Waiting to start, Working, Waiting for a person, Blocked, In a meeting or Not active.
- Selecting someone opens a card with Open profile, Talk by voice, Open the work and Enter room. The list can be searched and filtered and serves keyboard and screen-reader users.
- Activity comes only from real executions and live meetings the viewer may see. It is never simulated, and the page says when something is hidden. The floor mirrors in right-to-left languages.
Meeting Analytics, opened from Meetings, shows for a chosen period: meetings completed, cancelled and AI-only with their average length; AI answers given and failed and their tokens; minutes approvals, confirmed decisions and action items; tasks created and completed; and each AI employee's meetings and answers. It counts only meetings the viewer may see.
The seven meeting reports in the Reporting Center read only meeting turns, which hold no words. Transcripts, minutes, decisions, action items and participants are never reportable.
#Budgets, costs and KPIs
| Area | What it does |
|---|---|
| Budgets | Monthly token or USD limits for the office, a department or an employee, with a warning threshold. A used-up budget blocks new model calls; the work resumes when the limit is raised. |
| Budget alerts | Every 15 minutes, usage is compared with budgets and subscription quotas. Alerts inform only; nothing is stopped or deleted. |
| Cost Center | Estimated cost from configured model rates next to the invoiced amount a person enters, with the variance. |
| KPI Center | Success rate, on-time completion, rework rate, human intervention rate, approval rejection rate, quality score and cost per successful result, from recorded work only. |
| Reporting Center | Forty DVO reports and two dashboards, Digital Virtual Office – Cost and Usage and Digital Virtual Office – Workforce Performance. They cover the workspace (cost trend with a forecast, tokens, model calls, provider invoices, budgets, alerts and the KPI scorecard), each department (cost, usage, KPI results and status, staffing), each AI employee (usage and cost, top spenders, KPI results, quality ratings, meeting answers) and meetings (outcomes, answer rate, weekly trend, reasons). Names replace identifiers, KPI tiles compare a date range with the period before, and every report needs DvoReports. |
#Runtime, audit and retention
- Each tenant has one logical runtime with its own queue partition, limits and health. Work runs in process on Hangfire workers of the isolated
dvoqueue; no operating-system process is started per task. - A fair scheduler gives each tenant a bounded share, leases are extended by heartbeat, and expired leases are retried or dead-lettered. A person requeues dead-lettered work from the Runtime page.
- Every operation writes an audit event; the Audit page links each target to its record. Metrics and traces carry no tenant id, user id or text.
- A daily job applies each tenant's execution and audit retention (7 to 3650 days). Usage, tasks, projects, plans, messages, KPIs, knowledge and memory are kept regardless of age.
#Page reference
| Page | Address | Permission |
|---|---|---|
| Overview | /digital-virtual-office |
Dvo |
| Virtual Office | /digital-virtual-office/office |
DvoEmployees |
| Command Center | /digital-virtual-office/command-center |
DvoExecutions |
| Approval Center | /digital-virtual-office/approvals |
DvoApprovals |
| Plans | /digital-virtual-office/plans |
DvoExecutions |
| Tasks | /digital-virtual-office/tasks |
DvoTasks |
| Projects | /digital-virtual-office/projects |
DvoTasks |
| Communications | /digital-virtual-office/communications |
DvoCommunications |
| Voice conversation | /digital-virtual-office/voice/{id} |
DvoCommunications |
| Meetings | /digital-virtual-office/meetings |
DvoMeetings |
| Meeting record | /digital-virtual-office/meetings/{id} |
DvoMeetings |
| Meeting room | /digital-virtual-office/meetings/{id}/room |
DvoMeetings |
| Meeting Analytics | /digital-virtual-office/meetings/analytics |
DvoMeetings |
| Knowledge Center | /digital-virtual-office/knowledge |
DvoKnowledge |
| Memory | /digital-virtual-office/memory |
DvoKnowledge |
| Organization | /digital-virtual-office/organization |
DvoOrganization |
| AI employees | /digital-virtual-office/employees |
DvoEmployees |
| Work requests | /digital-virtual-office/executions |
DvoExecutions |
| Tools | /digital-virtual-office/tools |
DvoTools |
| Budgets and usage | /digital-virtual-office/budgets |
DvoBudgets |
| Cost Center | /digital-virtual-office/costs |
DvoBudgets |
| KPI Center | /digital-virtual-office/kpis |
DvoReports |
| Runtime | /digital-virtual-office/runtime |
DvoRuntime |
| Settings | /digital-virtual-office/settings |
DvoSettings |
| Audit | /digital-virtual-office/audit |
DvoAudit |
| AI Runtime Operations (platform) | /admin/dvo-platform |
DvoPlatform |
#Troubleshooting
| Symptom | What to do |
|---|---|
| "The Digital Virtual Office is not available." | Ask the platform to enable the dvo app; the runtime is provisioned on the next reconcile. |
| Work stays Queued | Check the Runtime page counts, whether the runtime is suspended, and that a node runs the dvo queue. |
| Work shows Blocked | Raise the budget or ask for a quota override, then Resume. |
Failed with provider.failed |
Check the provider status and the AI provider settings, then start again. |
| A step shows Not granted or Permission missing | Grant the tool on the profile, or ask a person who holds the permission to start the work. |
| I cannot approve | The approval page states the reason. Ask another approver. |
| A document is never found | Use the search tester with that employee and fix the access rule or file format. |
| Dead-lettered work | Fix the cause, then requeue from the Runtime page. |
| A portrait shows instead of the 3D avatar | The device has no WebGL, reduced motion is on or the file did not load. Everything still works with the portrait. |
| Voice shows Not set up or Demo voice, and the microphone is greyed out | On the Azure AI Speech card in AI Agent › Settings › Providers, save the key and the regional endpoint and turn it on, then run the speech test on the Voice card in Settings. |
| Nobody answers in a meeting | Check that the AI employees are active and allowed to take part, and that the meeting's AI turn and token limits are not used up. |
| No minutes after a meeting | The workspace does not keep transcripts, or the draft failed the schema. Choose Draft again or write the minutes by hand. |
Work waiting for a person, an approval or a delegate is parked by design and resumes when that happens. It is not stuck.