01 · Platform architecture Target architecture

An operating system for the business, not for the agents

The phrase “AI Operating System” is already taken on this market — and almost always it describes a system in which agents live: they get built, tested, published and observed. VOIR is arranged differently. The operating system here is the business itself, and its unit of work is not an agent run but a task the company needed done.

The argument of this page. A system like this can be described in three layers: who is allowed to act, who decides, and what carries the action out. When all three belong to the business rather than to an engineering team, what you have is an operating system for a company. The Control / Agentic / Execution taxonomy is borrowed as a proven three-layer model from enterprise platforms; the content of each layer is ours.

Trace a task Read about the three layers

Where this actually stands. At the time of publication none of the three layers and none of the eight departments is implemented. Everything described below is the target architecture of the product. The step-by-step playback on this page is labelled as a demonstration: it runs on pre-written scenarios, calls no live AI service and changes no data. The project has no customers, testimonials or deployment metrics, so none are shown.

02 · Cross-section

Three layers that every task passes through

These are not stages of software development. They describe how work moves in a running company: first, whether we are allowed to do this at all; then who decides and how; and only then what actually carries the decision out.

L·01 Control — governance and limits

Who is allowed to act

The layer of rules: what an agent may do, which actions need a person to approve them, and how every decision is recorded. It does no work itself — it sets the boundaries of the work.

In a real company this is the charter, the job descriptions and the signing limits. In VOIR the Control layer exists so that a department’s autonomy never turns into unpredictability: anything touching money, customers or external systems passes through a rule the owner set.

  • Rules and prohibitions written in plain language, not in regular expressions
  • Human approval for consequential actions, as a separate and explicit step
  • Access separation: what a department may do, apart from who is using it
  • A decision log: who, when and on what grounds acted that way
  • Financial operations and payments only with the owner’s explicit permission
Analogy Company policy and a manager’s signing limit: what can be done without approval, and what requires a signature.
What this layer does NOT do It makes no decisions about the task and calls no tools. It only permits, forbids, or sends the action for approval.
Target architecture Implemented: no

L·02 Agentic — decision making

Who decides, and how the task is broken down

The layer where the AI CEO and the eight departments live. Here a task becomes a plan: subtasks, owners, sequence, dependencies and the definition of done.

This is the only layer that thinks. The AI CEO takes the task from the owner exactly as it was stated — no prompt engineering, no instruction about which department owns what — and works out the breakdown itself. Departments then operate inside their own function, because marketing, finance and support have different data, different tools and different standards of a good result.

  • AI CEO: takes the goal, decomposes it, assigns owners, watches the deadlines
  • Eight departments as permanent business functions, not a pile of separate chatbots
  • Results handed between departments along a route, not by manual copy-paste
  • A deterministic graph where the process is known, autonomy inside an individual step
  • Escalation to the owner when a department cannot solve the task within its limits
Analogy A managing director and department heads: the request arrives as one sentence and leaves as a plan with named owners.
What this layer does NOT do It does not store access credentials and does not reach into external systems directly — that belongs to the Execution layer.

This layer is described in full on the AI CEO page

Target architecture Implemented: no

L·03 Execution — carrying it out

What actually performs the decision

The layer of tools: connectors to mail, messengers, CRM, spreadsheets and calendars, model calls, deterministic steps, and the result that goes back to the department.

No decisions are made here. This layer executes exactly what the Agentic layer decided and what the Control layer permitted. Its key requirement: credentials must never reach the model. The tool holds the access; the model only sees the value that came back.

  • Native connectors: mail, messengers, CRM, spreadsheets, calendar
  • Credentials stay in the gateway and are never handed to the model
  • Deterministic steps for processes whose sequence is known in advance
  • A choice of several models instead of a lock-in to one vendor
  • Every call recorded: what was done, what it cost, whether the step succeeded
Analogy An employee’s working tools: system access, document templates and the authority to perform an operation.
What this layer does NOT do It does not judge whether an action is needed and does not alter the plan. A failed step is returned to the department, never worked around in silence.
Target architecture Implemented: no

03 · Path of a task

One task, all the way through

Instead of an abstract diagram: the path of a single unit of work, from the owner’s wording to the answer the owner receives. Every step is labelled with the layer it happens in.

Demonstration

The three scenarios below are pre-written text, not the system working. They are not executed by a live AI service, they make no network calls and they change no data. Playing the steps exists only to show the order in which a task passes through the layers. The full text of every step is present in the HTML and readable without JavaScript.

Scenario

Waiting: pick a scenario and press Play

You can also walk the steps without the timer: focus the list of steps and use the ↑ ↓ arrow keys.

A customer wants a refund for an order that never arrived

The owner’s wording: “Customer N wants a refund, the delivery fell through. Sort it out and tell me what we are doing.” One unit of work — a customer request — passes through all three layers.

  1. The request is taken in as a task Agentic The owner passes the wording as it is. The AI CEO records the goal: resolve the customer’s request and return a conclusion to the owner — not forward the email to the support inbox.
  2. The goal is broken into subtasks Agentic Three subtasks: reconstruct the facts of the order (Operations), assess the financial consequences of a refund (Finance), draft the reply to the customer (Support). Done means a decision on the refund plus the text of the reply.
  3. Check: is this allowed Control The Control layer compares the plan against the rules. A refund is an action with financial consequences, so it requires the owner’s approval. Gathering the facts and drafting the reply do not.
  4. Departments collect the data Execution Operations pull the delivery statuses for the order, Finance pull the payments and deductions, Support pull the customer’s request history. Credentials stay in the gateway: the model only ever sees the values that came back.
  5. The owner approves Control The AI CEO puts the decision up for approval: the refund amount, the reason, what has already been done and what changes for the customer. Without approval the refund is not made.
  6. One consolidated answer to the owner Agentic The result is written in business language: what happened, who did what, which decision was taken, and what should be fixed in the process so the failed delivery does not repeat. Tool calls do not appear in that answer.

Marketing spend went up but the number of leads did not

The owner’s wording: “We are spending more and getting the same number of leads. Find where it is being lost.” The task moves through the layers in the same order, but the executors are different.

  1. The task is accepted as one sentence Agentic The owner names no channel and writes no prompt. The AI CEO takes the task as a question about two metrics that have diverged and works out which departments hold the data needed to answer it.
  2. Plan: spend, conversion, market comparison Agentic Finance reconcile spend by channel for the period, Marketing pull acquisition numbers and the conversion into a lead, Research pull how the cost of a contact moved across the industry. The order is chosen so the findings do not depend on one another.
  3. Boundary check: read-only Control All three subtasks read and calculate; there is no external action, so no approval is required. Changing an advertising budget would need separate sign-off.
  4. Departments read from their own systems Execution Billing and ad platforms, on-site counters, public market data. Every call is recorded: what was requested, whether it succeeded, what it cost.
  5. The divergence is found and verified Agentic The AI CEO compares what the departments returned and discards the explanations the data does not support. If the divergence stays unexplained, the task goes back to the department with a follow-up question rather than closing on a plausible answer.
  6. A report with options and their consequences Agentic The owner receives an explanation of the divergence and two or three possible actions with the consequence of each. The system does not reallocate the budget — that is the owner’s call.

We need to fill a role in the sales department

The owner’s wording: “We need someone in sales.” This is a long task: it does not end with a single answer and it needs deadline control across stages.

  1. A task with a horizon of weeks is accepted Agentic The AI CEO records a process with stages and deadlines rather than an answer: role profile, job description, budget, flow of applicants, first day. The task stays open between stages.
  2. Stages distributed across departments Agentic Sales describe what the role does, HR write the job description and the screening criteria, Finance confirm the budget, Operations draft the onboarding plan.
  3. External actions go out for approval Control Drafting the text needs no approval. Publishing it on external job boards does, and so does any offer made to a candidate, because both are actions taken outside in the company’s name.
  4. The stage runs and applications come in Execution The tools publish the vacancy once approved and collect the applications. The outcome of the stage is stored, so a week later the system resumes from the same point instead of starting the analysis over.
  5. Deadline control and escalation Agentic If a stage sits past its planned date, or too few applications arrive, the AI CEO tells the owner which stage the task is stuck at and what could be changed — instead of waiting quietly.
  6. A single hiring report Agentic Role profile, job description, screening criteria, confirmed budget and the plan for the first tasks — as one chain, stating which department prepared what and where the decision stayed with the owner.

04 · Composition

Eight departments, seen through the layers

For every department we state which layer owns its decisions, what it receives as input and what it must return. The individual department pages are not published yet — they come later, which is why there are no links to them here.

DEPT·01 / MKT

Marketing

department page: coming soon
Input A demand goal, product data, constraints on budget and on the tone of communication
Output Materials and acquisition channels, campaign performance, a verdict on what is working
Decision layer Agentic — the department picks the channel mix itself within the limits it was given
Needs approval Launching and paying for ad campaigns, any public communication in the company’s name
Target architecture Implemented: no
DEPT·02 / SLS

Sales

department page: coming soon
Input Requests and leads from Marketing, product terms, limits on discounts
Output Qualified requests, drafted proposals, the state of the pipeline
Decision layer Agentic — prioritising requests and estimating how likely a deal is to close
Needs approval Sending a message to a customer; terms or discounts outside the approved range
Target architecture Implemented: no
DEPT·03 / OPS

Operations

department page: coming soon
Input Orders, commitments, procedures and deadlines, process documents
Output The actual state of delivery, bottlenecks, a work plan with named owners
Decision layer Hybrid — known processes run as a deterministic graph; deviations are worked out by the department
Needs approval Changing a date promised to a customer, and anything that touches company commitments
Target architecture Implemented: no
DEPT·04 / FIN

Finance

department page: coming soon
Input Payment and expense data, the budget, questions about the profitability of a line of business
Output Calculations, reports by business line, the consequences of each option, a cash forecast
Decision layer Agentic — weighing options and their consequences, not merely recomputing numbers
Needs approval Every financial operation and payment, without exception — only with the owner’s explicit permission
Target architecture Payments only with the owner’s permission
DEPT·05 / HR

HR

department page: coming soon
Input A need for a role, what the position does, constraints on budget and on the type of employment
Output The job description, screening criteria, an onboarding plan, internal instructions
Decision layer Agentic — screening applications against the stated criteria, with the reasoning shown
Needs approval Publishing a vacancy, any message to a candidate, any decision about a person
Target architecture Implemented: no
DEPT·06 / RES

Research

department page: coming soon
Input A question about the market, competitors or data; the scope of the search and the date an answer is needed
Output A summary naming its sources and the date of the snapshot, with an explicit “no data” where there is none
Decision layer Agentic — choosing sources and judging how reliable what was found actually is
Needs approval Nothing: the department works only with public sources and performs no external actions
Target architecture Implemented: no
DEPT·07 / DEV

Development

department page: coming soon
Input A product requirement, a defect report, a task to integrate with an external system
Output A verified change, technical documentation, an estimate of effort and of risk
Decision layer Agentic — choosing how to meet the requirement
Needs approval Releasing a change into the working environment and any access to production data
Target architecture Implemented: no
DEPT·08 / SUP

Support

department page: coming soon
Input Customer requests, the knowledge base, tone rules and the limits of what may be promised
Output Replies for routine cases, classified requests, systemic causes passed on to departments
Decision layer Agentic — judging a case and choosing: answer, ask for details, or hand it to another department
Needs approval Promises to a customer, compensation, anything outside the approved reply rules
Target architecture Eighth department — added to the list from the specification

05 · Positioning

“Operating system” is already in use. The difference is elsewhere

Below is how the companies we researched name their platforms, and what their unit of work turns out to be. The wording comes from public pages and materials collected during the research. The comparison describes the subject of each platform, not its quality.

Beam AI

“The Operating System for AI Transformation”

Unit: a process

The platform maps the processes of a department, estimates their volume and cost, and puts agents on the most expensive ones. The layer is called AgentOS. This is a system for turning a company’s processes into ones that run themselves.

Source: research/extra-10.md, Beam AI section (public home page)

Sierra

“Agent OS”

Unit: a customer-facing agent

An operating system for customer-experience teams and for the developers who build agents: assembly, observability, long-running improvement plans. The subject of the system is the lifecycle of an agent that talks to customers.

Source: research/extra-10.md, official Sierra blog announcement

Dust

“Multiplayer AI for human-agent collaboration”

Unit: a workspace

People and agents work in shared spaces with shared context; the site has a Departments section. Coordination happens through the shared workspace and mentions — there is no separate central manager.

Source: research/extra-10.md, dust.tt (home page, Series B blog post)

Workato

“Neutral control & execution platform for Enterprise AI”

Unit: an integration and a recipe

The origin of the Control / Agentic / Execution taxonomy we use as our frame. The subject of the platform is governing enterprise integrations, with a layer of agents on top of them.

Source: research/06-workato.md, Workato pricing page, 30.10.2026

VOIR AI

“Your entire business. One AI operating system.”

Unit: a business task

The subject of the system is not an agent, not a process and not a workspace, but a company task — from the owner’s wording to a consolidated answer. That is why the centre is an AI CEO who takes the goal and distributes it across eight departments, rather than a catalogue of agents the owner has to assemble.

Target architecture, product in development

Why this difference matters

If the unit of work is an agent or a process, the business owner still has to work out which agents to assemble, in what order to run them and how to merge what comes back. If the unit of work is a business task, then the breakdown, the assignment, the sequence and the consolidation become the system’s job. That difference decides who does the management work: a person holding an agent builder, or a platform with a manager at its centre.

The role of the central manager is described in full on the AI CEO page

06 · Limits

What this system does not do

The limits are described in as much detail as the capabilities: they mark exactly where a department’s autonomy ends and the owner’s decision begins.

It does not make payments or financial operations

The finance department prepares calculations, reports and assessments of consequences. Payments, transfers and any operation involving money are not performed without the owner’s explicit permission — enforced at the level of the Control layer rules, not as a default setting.

It does not take the owner’s place on decisions about people and customers

The departments prepare the material: a job description, screening criteria, the text of a reply to a customer, the terms of an offer. The decision about a person and the promise to a customer stay with the owner or an authorised employee.

It does not pass off a guess as a result

If there is not enough data for a conclusion, the task goes back to the department with a follow-up question or escalates to the owner. An answer of “there is no data” counts as a correct outcome of the department’s work, not as a failure.

It does not bypass site protections or break service terms

Information is gathered from external sources legitimately: robots.txt and the terms of the service are respected, and the agent identifies itself as automated. Circumventing blocks and impersonating a human are not part of the target architecture.

It does not work today

This is a target architecture. At the time of publication none of the three layers and none of the eight departments is implemented. The task-path demonstration on this page uses pre-written text and is not connected to a live AI service.

07 · What next

Where to go from this page

The Agentic layer · in detail AI CEO Who accepts the task, how it is broken into subtasks, how deadlines are watched and how the owner is reported to. The L1/L2/L3 autonomy scale lives there too. Open the page Home · composition The eight departments The full task list of every department: marketing, sales, operations, finance, HR, research, development and support. To the department list
Not published yet Pricing, About, Contact These pages are being written. Until they are published none of them is a link — so you are never sent to an empty page. coming soon