• Staffing table
  • 8 roles
  • Hired: 0

Eight roles on a staffing table that nobody has signed yet.

A VOIR AI employee is not a chat widget on your site and not a “smart assistant” waiting for a question. In the target architecture it is a role with an area of responsibility: duties, a skill set, target metrics and explicitly written limits. Below is the staffing table for all eight roles and the personnel file for each one. Not a single position is filled — no department is implemented, and we say so on every card rather than in a footnote.

Open the personnel files Staffing table Who assigns the work

01 · The table

Staffing table: eight roles

The document is laid out like a real personnel table: role, department, code, autonomy grade, billing unit and status. The “Hired” column reads zero for all eight rows — that is the current state of the product. Rates are given in planned VU and MC billing units, not in money: the operation price list is published, but sales have not started.

Form T-1 (notional) · approved: no · in force: target architecture
№ Role Department Code Autonomy grade Hired Status
01 AI Marketer Marketing MKT L2 — within a campaign plan 0 Target architecture
02 AI Sales Manager Sales SLS L1 — every step approved 0 Target architecture
03 AI Operations Manager Operations OPS L2 — runs the playbook 0 Target architecture
04 AI Financial Analyst Finance FIN L1 — no payment authority 0 Target architecture
05 AI Recruiter People HR L1 — decisions stay human 0 Target architecture
06 AI Research Analyst Research RES L3 — runs the research itself 0 Target architecture
07 AI Developer Development DEV L2 — inside a sandbox 0 Target architecture
08 AI Support Specialist Support SUP L2 — routine cases, rest to humans 0 Target architecture
Σ Headcount total 8 departments — grades set by the owner 0 of 8 all positions open

The table scrolls sideways inside its frame — the page itself never widens

The autonomy grade in the table is a designed default, not a property of a finished employee. The business owner sets the grade; the L1/L2/L3 scale is described in the “Autonomy grades” section below and on the AI CEO page.

02 · Personnel files

Eight cards: duties, skills, KPIs, limits

Every file follows the same form: what the employee does, which skills that requires, which metrics would show the result and — as a mandatory line — what the employee never does. The metrics are named but not measured: there are no numbers in the cards, because there is nothing to measure yet. The cards tilt and light up under the cursor, as if a folder were being pulled off a shelf; on touch devices and with prefers-reduced-motion the tilt is off and the content is unchanged.

FILE · 01 / MKT

Marketing department

AI Marketer

Demand, positioning and acquisition channels: from reading the audience to reporting on the campaign.

Skills

  • market and competitor analysis
  • SEO and content planning
  • social channels
  • ad campaigns
  • image and video generation
  • email marketing
  • lead generation
  • campaign analytics

KPIs · metric names, no values

  • Cost per enquiry — campaign spend divided by incoming enquiries
  • Share of qualified leads — how many enquiries made it through sales qualification
  • Campaign lead time — from brief to finished materials
  • Content plan coverage — share of planned publications that shipped on time

Never does: publish anything externally without human approval, and never spend ad budget on its own.

Target architecture position open
FILE · 02 / SLS

Sales department

AI Sales Manager

Pipeline, deals and lead handling: qualification, proposals, moving a deal through its stages.

Skills

  • prospect discovery
  • target audience analysis
  • proposals
  • CRM work
  • database segmentation
  • personalised outreach
  • pipeline analysis
  • results tracking

KPIs · metric names, no values

  • Stage conversion — share of deals that moved to the next stage
  • Time to first response — from inbound enquiry to first contact
  • CRM record completeness — share of deals with all required fields filled
  • Share of stalled deals — deals with no next step beyond the agreed window

Never does: send emails or messages to customers without review, and never promise discounts or terms that are not in the price list.

Target architecture position open
FILE · 03 / OPS

Operations department

AI Operations Manager

Processes, delivery and deadlines: keeps the playbook, tracks dependencies, works through failures.

Skills

  • process management
  • task delivery control
  • automation of repetitive operations
  • document handling
  • reporting
  • project management
  • quality control

KPIs · metric names, no values

  • On-time delivery — share of tasks closed by the stated date
  • Cycle time — from task creation to result for a standard process
  • Human interventions — how often a process needed a person
  • Rework rate — share of results not accepted on the first pass

Never does: rewrite the playbook after the fact, and never close a task whose result failed the quality check.

Target architecture position open
FILE · 04 / FIN

Finance department

AI Financial Analyst

Money, reporting and forecasts: reconciles cash flows, spots gaps early, prepares the data behind decisions.

Skills

  • expense analysis
  • revenue analysis
  • financial reporting
  • budget planning
  • profitability analysis
  • metric control
  • payment calendar

KPIs · metric names, no values

  • Forecast accuracy — the gap between plan and actual inflows and outflows
  • Period close time — how long reconciliation and reporting take
  • Warning lead time — how many days ahead a cash gap becomes visible
  • Reconciliation exceptions — how many operations did not match, and why

Never does: execute payments or any financial operation without explicit permission — this is a requirement of the architecture, not a temporary restriction.

Target architecture position open
FILE · 05 / HR

People department

AI Recruiter

Hiring, team and workload: job specs, first-pass screening of applications, onboarding, internal instructions.

Skills

  • job postings
  • résumé analysis
  • role descriptions
  • employee task management
  • internal instructions
  • onboarding planning

KPIs · metric names, no values

  • Time to fill — from posting to a candidate starting
  • Share screened out by criteria — how many applications were filtered automatically, and by which rule
  • Onboarding completeness — share of the onboarding plan actually completed
  • Team workload — how work is spread across people, without bottlenecks

Never does: make hiring or dismissal decisions, and never assess people by gender, age or other protected characteristics.

Target architecture position open
FILE · 06 / RES

Research department

AI Research Analyst

Market, competitors and data: gathers, cross-checks and hands back a brief where every claim has a source.

Skills

  • market research
  • competitor discovery
  • technology analysis
  • public information gathering
  • analytical reports
  • market monitoring

KPIs · metric names, no values

  • Share of sourced claims — how many statements in the report carry a reference
  • Unverified assertions — hypotheses explicitly marked as unproven
  • Brief turnaround — from question to an answer with sources
  • Data freshness — age of the oldest source in the report

Never does: pass an assumption off as a verified fact, and never bypass paywalls, CAPTCHAs or a source's access restrictions.

Target architecture position open
FILE · 07 / DEV

Development department

AI Developer

Product, integrations and automation: writes and tests code, connects external systems, keeps documentation current.

Skills

  • software development
  • website building
  • API development
  • test automation
  • code review
  • bug hunting
  • technical documentation

KPIs · metric names, no values

  • Share of changes accepted as-is — how much work passed review on the first pass
  • Test coverage — how much of the changed code is verified automatically
  • Regressions — how much previously working behaviour a change broke
  • Time to rollback — how fast a problem was noticed and a working state restored

Never does: ship changes to production unconfirmed, work with live secrets, or delete anyone else's data.

Target architecture position open
FILE · 08 / SUP

Support department

AI Support Specialist

Customers, requests and quality: answers the routine, passes the hard cases to the right department with full context.

Skills

  • inbound request handling
  • standard answers from the knowledge base
  • routing complex cases
  • collecting recurring problems
  • product suggestions

KPIs · metric names, no values

  • Share resolved without a human — and separately, the share that needed one
  • Time to first response — from request to reaction
  • Repeat contacts — how many customers came back with the same problem
  • Routing accuracy — share of requests that reached the right department first time

Never does: pretend to be a human, or promise deadlines and compensation nobody approved.

Target architecture position open

03 · Reporting lines

Who the employee reports to, and where work comes from

This is what separates VOIR AI from “hiring one AI assistant”: an employee here is a member of a department, not a standalone role. The business owner sets the task, the AI CEO breaks it into subtasks and assigns owners, and the result returns to the owner as one answer. So research reaches marketing, customer complaints reach development and money discrepancies reach operations — without a conversation between “tools”. The diagram below is a drawing of the target architecture, not a screenshot of a running system.

Reporting diagram: the business owner hands a task to the AI CEO, which distributes it across eight departments, and results return to the owner as a single report Business owner SETS THE TASK · RECEIVES THE REPORT AI CEO decomposition · assignment · control · escalation Marketing MKT · 01 Sales SLS · 02 Operations OPS · 03 Finance FIN · 04 People HR · 05 Research RES · 06 Development DEV · 07 Support SUP · 08

A drawing of the target architecture. The two arrowed lines at the top are the two-way exchange with the owner: the downward arrow is the task being set, the upward arrow is the report coming back. The finer dashes below are subtasks being distributed and results collected. There is no timeline here: not one of these lines has carried real data yet. How the system is built inside — in “Layers of AI OS”.

03b · Place in the cycle

How an employee is wired into the AI CEO cycle

A standalone role with no link to the control centre is just a tool. Below is exactly where the employee appears in the AI CEO cycle, which artifact it receives as input, which it returns as output, and at which points it exchanges work with other departments and with the owner. The cycle itself is described on the AI CEO page; here it is only the employee's part in it. As everywhere on this page, it is target architecture.

1

Task intake

The owner states the task to the AI CEO. The employee takes no part at this stage and does not “see” the request directly — it receives work only after the breakdown.

Employee's part: none

2

Analysis and assignment

The AI CEO breaks the task into subtasks and assigns owners by department. The employee receives a subtask along with its context: what is already known, what counts as a finished result, and which autonomy grade it operates within.

Employee's part: accept the assignment, check the data is complete, ask for what is missing

3

Execution

The employee does the work and records a trace of every action: what was done, on what data, with which tool. When the task needs another department, it hands the artifact over instead of imitating someone else's work.

Employee's part: execute, keep the trace, escalate when data or authority runs out

4

Independent acceptance

The result is checked by someone other than its author: a comparison against the acceptance criteria from the assignment. An employee cannot accept its own work — this is a rule of the architecture, not a setting.

Employee's part: provide the artifact and the trace, fix what was rejected

5

Report to the owner

The AI CEO folds the departments' results into one answer for the owner: what was done, what was not and why, which decisions need a human. The employee contributes its own part.

Employee's part: a summary for its area of responsibility, with an honest status

Artifacts in and out

The same principle for all eight departments: the input is a subtask with context, the output is an artifact that can be checked plus an entry in the trace. Exchange between departments runs on artifacts, not on “forwarding messages”.

Department Receives from the AI CEO Returns (artifact) Exchanges with
Marketing MKT A subtask on demand or a channel, plus research data Positioning, a content plan, materials, a campaign report Research → Marketing; Marketing → Sales
Sales SLS A stream of enquiries, plus qualification criteria Qualified deals, proposals, pipeline state Marketing → Sales; Sales → Finance
Operations OPS A process or task with deadlines and dependencies A playbook, a delivery plan, a report on failures in the chain All departments → Operations; Operations → AI CEO
Finance FIN A period and a line of business to reconcile, plus sales data A revenue and expense summary, a forecast, a cash-gap warning Sales → Finance; Finance → the owner (via the AI CEO)
People HR A request for a role, plus experience requirements A job posting, application screening criteria, an onboarding plan Any department → People; People → Operations
Research RES A question about the market, competitors or data A brief with a source behind every claim, hypotheses marked as unproven Research → Marketing, Sales, Finance, Development
Development DEV A requirement for a feature or an integration A change with tests, technical documentation, a regression report Support → Development; Development → Operations
Support SUP Inbound requests, plus the knowledge base Closed requests, routed cases, a list of recurring problems Support → Development, Operations, Marketing

The table scrolls sideways inside its frame — the page itself never widens

One shared department memory

The employee does not start from scratch every time: playbooks, past decisions and company context live alongside the task and are available to the next one to pick it up. This is what separates a department from a pile of one-off chatbot queries.

Target architecture — not implemented

Escalation instead of guessing

Missing data, conflicting requirements or a step beyond its autonomy grade stops the task and returns it to the AI CEO with a description of exactly what was lacking. The owner sees a question, not a plausibly filled gap.

Target architecture — not implemented

Acceptance by someone else

The result is checked against the acceptance criteria from the assignment by a separate pass. The employee that did the work does not accept it — otherwise the “share accepted without edits” metric means nothing.

Target architecture — not implemented

None of the five stages works today: there is no AI CEO code and no department code. This section describes a contract — how an employee must be wired into the cycle once implementation starts. It can be used as an acceptance criterion: if a future version accepts a task but never hands an artifact to a neighbouring department and never goes through independent acceptance, it does not match this page.

04 · Autonomy grades

How much freedom an employee has is the owner's call

Autonomy is not a property of “smart AI”; it is a setting the employer chooses. For a human employee it lives in the job description; for an AI employee in the VOIR AI target architecture it lives in a grade. The same AI financial analyst on L1 prepares a brief when asked, and on L3 runs the whole process and reports afterwards — the difference is not the model, it is the authority granted.

L1

Execution on request

The employee does exactly what was asked and stops there. Anything beyond the request is approved first. This is the mode for money, legal commitments and any irreversible action.

  • Finance: reconciles and explains, never pays
  • Sales: drafts the email, a human sends it
  • People: screens applications by criteria, the manager decides
Target architecture authority granted by the owner
L2

Running to a plan

The employee runs a process against an approved plan and chooses how to execute within it. A deviation from the plan, going over budget or genuine ambiguity is a reason to stop and ask — not to make something up.

  • Marketing: runs the content plan, a human approves publishing
  • Operations: holds the playbook and deadlines, escalates failures
  • Support: closes routine cases, passes the rest on with context
Target architecture the plan is set by the owner
L3

Owning the whole process

The employee sets its own subtasks, picks the order and reports after the fact. Even on L3 the prohibitions stand: money, external publishing and irreversible operations still require a human approval.

  • Research: runs market monitoring and produces sourced briefs
  • Development: implements inside a sandbox; production only after review
  • Any department: only by the owner's explicit decision
Target architecture a grade does not lift a prohibition

The scale matches the autonomy ladder on the AI CEO page: that page describes how the AI CEO distributes authority, this one shows what it looks like in an employee's job description.

05 · Onboarding

How an employee joins the company

Onboarding an AI employee in the target architecture mirrors onboarding a person: role, access, instructions, probation, review. One thing differs — “probation” here is not a period of time but a level of supervision, which the owner relaxes as the results earn trust. The order of the steps below is a design; it has not been tested on real deployments, because there have been none.

Open the position

The owner picks a department and states the outcome needed: not “do marketing”, but “four publications a week and a report on cost per enquiry”. That statement becomes the employee's duties — the way a job description does.

in: task statement · out: job description

Grant access and set the grade

The employee receives only the connections its tasks require, plus an autonomy grade of L1, L2 or L3. Secrets never enter the model's context: the employee works through a connection, not through the text of a token. This is an architectural requirement, not an optional setting.

in: list of systems · out: permission set + grade

Write the rules and the limits

What counts as a finished result, when to stop and ask, and what is never done at all. The limits in the personnel files above are not marketing copy but the content of this step: they go into the employee's instructions as written.

in: acceptance criteria · out: execution rules

Run a probationary mode

Early tasks run at L1: every action is visible and confirmed by a human. Once the share of results accepted without edits holds steady, the grade goes up. Raising it is the owner's decision, not a property of the system.

in: stream of tasks · out: grade decision

Measure against the named KPIs

The metrics in an employee's card start counting from the first real execution. That is exactly why the cards name metrics without values: a number without a measurement is fiction. Target values are set by the owner, based on their own economics.

in: actual execution · out: measurable KPIs

06 · Limits

What an AI employee is not

A page promising “an employee that replaces a department” sells distrust along with it. Below is what a VOIR AI employee does not do and will not do in the target architecture. These limits matter more than the skill list: they decide whether you can trust the system with money, customers and reputation.

Not a person, and not a replacement for one

An employee has no legal status, bears no liability and signs nothing. Responsibility for the outcome always stays with the business owner. In support, the employee must say that a system is answering, not a human.

Does not pay and does not touch money

Financial operations and payments are not performed without explicit permission — a direct requirement of the architecture. The AI financial analyst reconciles, explains and warns; a human makes the payment.

Publishes nothing externally unconfirmed

Posts, customer emails, changes on a public site and replies in the company's name go out only after a human approves. A mistake in a draft costs nothing; a mistake in a publication costs reputation.

Does not guess when data is missing

Missing data, conflicting requirements or a step beyond its authority is a reason to stop the task and escalate to a human — not to “plausibly fill the gap”. The escalation logic is described on the AI CEO page.

Does not hide what it did

Every action leaves a trace: what was done, on what data, with which tool, and what came out. The trace is rendered as a business narrative for the owner — see “Explainability” for how that works.

Does not work today

As of publication, zero of the eight roles are implemented. This is not “coming soon” and not “in private beta”: the employees' code does not exist. The page describes what will be built, and can be used as a commitment to check against.

07 · Comparison

Employee, agent, chatbot, human

“AI employee”, “AI agent” and “chatbot” get used as synonyms, which makes it impossible to tell what you are actually buying. The table below separates four roles by six criteria. The VOIR AI column describes the target architecture — what will be, not what is.

Criterion Chatbot Standalone AI agent VOIR AI employee Human employee
What it holds A conversation: answers a message, waits for the next one One task or one tool A role: an area of responsibility with duties, skills and KPIs A role and an area of responsibility
Who assigns work The user, every time from scratch The owner or a trigger, with no distribution across roles The AI CEO: takes the owner's task and distributes it across departments A manager
Sees others' work No No, unless a separate graph is built Yes: one department's output becomes another's context Yes, through shared processes
How it is bounded By the conversation script By the set of available tools By an autonomy grade L1/L2/L3 and an explicit list of prohibitions By the job description and the law
What is left afterwards Chat history The result of one task A trace of every action, rendered as a report for the owner A result, a report and experience
Cost of a mistake A wrong answer in a conversation A wrong result for one task A wrong decision on a process — which is why money and publishing go through a human A wrong decision on a process

The table scrolls sideways inside its frame — the page itself never widens

The “VOIR AI employee” column describes the target architecture. The comparison with chatbots, standalone agents and human employees is made by criteria, not by measurement: no VOIR AI deployment exists, so there are no numbers to compare. How the execution layers of the system are built — in “Layers of AI OS”.

08 · Questions

What people ask about AI employees

These answers repeat what is written above, deliberately, so you do not have to hunt for it. None of them promises anything that is not in the target architecture.

Can I hire a VOIR AI employee today?
No. None of the eight roles is implemented: zero employees have been hired as of publication. Everything on this page is target architecture, and every card carries the corresponding stamp. The website itself and the project documentation are what actually work.
Why are there no KPI numbers in the personnel files?
Because there are none to show — there is nothing to measure yet. We publish the metric names and how each would be calculated, so that once a role ships, the result can be checked against the promise instead of the definition being rewritten to fit the code. A number such as “34% conversion lift” would be invented. See “Personnel files”.
How is an AI employee different from a chatbot?
A chatbot replies to a message and waits for the next question. In the VOIR AI target architecture an employee owns the whole task: it accepts the work, breaks it into steps, does it, checks the result and reports back — and when data or authority runs out it escalates to a human instead of guessing. A full comparison is in the table above.
How is an AI employee different from a department?
A department is an area of responsibility; an employee is the one doing the work inside it. At VOIR AI each of the eight employees belongs to a department and receives work through the AI CEO, so the output of one is visible to another: market research reaches marketing, and customer complaints reach development. The department list is on the home page.
Do the employees work autonomously?
In the target architecture, within the autonomy grade the owner sets: from acting only on an explicit request (L1) to running a whole process and reporting afterwards (L3). Anything involving money, publishing externally or irreversible operations always requires a human approval. The grades are described in “Autonomy grades”.
What will an employee cost?
The planned model is published in advance on the pricing page: two billing units — VU for the result of an operation and MC for model cost at cost price — plus a line-by-line price list of operations. Sales have not started, no payment provider is connected, and there are deliberately no checkout buttons on the site.
How do I contact the team?
By email at robert@voir.su — the only working contact channel at the moment. Feedback forms are not published until their server-side handler exists: a form that does nothing looks like a lie.

09 · Next in the file

What to read after the job descriptions

Three sections that answer the questions a staffing table raises: who assigns work to the employees, how the system is built inside, and what it will cost.

AI CEO · Functions Who assigns work to the employees Taking a task from the owner, decomposition, assignment, deadline control and folding the results into one answer. AI Operating System · Layers How the system is built inside Control, Agentic and Execution: what each layer is responsible for, and where a department employee sits in it. Pricing · Operations What an employee's work will cost A line-by-line price list of operations in VU and MC units. This is a planned model: sales have not started and there is no checkout. Home · Demonstration What delegating a task looks like Four business tasks broken down across departments. This is a demonstration on prepared scenarios, not a live AI service. About · Architecture Three bays of the system and what is done The whole target architecture and the project's log book with dates — including what is still not built. AI OS · Task trace What is left after an employee works A trace of execution broken down on three examples: how each action turns into a report the business owner can read.

The staffing table is open — and still empty

Eight roles are described, the limits are written down, the metrics are named. One thing is missing: the implementation. We publish the table before launch on purpose — it can be checked and argued with in advance, and after launch the result can be compared against the promise instead of the description being rewritten to fit the code. If you want to hear when the first role actually works, write to us by email: a signup form will appear only together with a working handler behind it.

Write to robert@voir.su See the delegation demo What we commit not to do

The delegation demo on the home page uses prepared scenarios and is not connected to a live AI service. The commitments on the pricing page describe planned rules, not a contract in force: sales have not started.