VOIR AI is an AI operating system for business: the owner states a task, AI CEO breaks it into subtasks and assigns them to eight specialised departments. Today that is a target architecture, not a working product, and we say so on every page. Below — why we are building it, who is building it, what has been done so far, and the rules we work by.
01 · History
VOIR AI has no tidy legend about founders meeting in a garage. What it has is a date, a list of requirements and twenty competitor teardowns — and we prefer to tell it that way.
On 30 October 2026 the project began with the thing people usually do last: checking the infrastructure. Before writing a single line of HTML we confirmed that the domain, the certificate, the web server and the publishing path exist and do not touch the neighbouring sites. Only then did the market research start — ten platforms from the specification and ten more found beyond it: Zapier, Lindy, Relevance AI, Gumloop, CrewAI, Workato, Bardeen, Pipedream, Relay and UiPath, plus the players in the “AI employee” segment.
The research produced an uncomfortable but useful conclusion: “an AI agent you assemble from a prompt” is something practically everyone has. It is a hygiene requirement, not an advantage. The real gap in the market is elsewhere. Platforms sell tools — builders, triggers, integrations. The business owner is left working as the dispatcher: choosing which service to hand what to, copying results from one window into another, and merging nine separate answers into one. That is exactly the work we want to take away. Not another builder, but a structure where a task can simply be delegated.
This is where the metaphor of a company rather than a pile of chatbots came from: AI CEO in the centre and eight departments around it — marketing, sales, operations, finance, people, research, development and support. The eighth department, support, was not in the original list, which ended at development. We added it deliberately, because without handling customer requests the chain breaks at the point most visible to the owner.
On 31 October the site went to production, and on the same day the priorities changed. The project lead imposed a moratorium on mass page creation and set a requirement that shaped everything after it: the site must present a real product, not an imitation of one. First the architecture, an inventory of what already exists in our systems, a prototype plan and its review by an independent role. Then — pages about what actually works. On 1 November the moratorium was lifted and publication of the sections resumed in that order.
This history is short because the project is young. We are not going to pad it with invented “years of experience” or a “team of 40 specialists”: one honest page of biography is more useful than a good-looking legend nobody can verify.
02 · Mission
Small and medium businesses drown in operations not because the owner is lazy, but because there is nobody to delegate to: the functions exist, the people for them do not. The mission of VOIR AI is to make delegation available where hiring a department is impossible.
We are building a system where a task is stated the way an owner would state it to a human manager: “launch a new product”, “find the cause of the cash gap”, “work through the wave of complaints”. The system then decides which departments get which parts of the task, in what order, what counts as a result, and when to return the consolidated answer. The owner sees the route, not a black box.
The second half of the mission is about trust. An AI that is handed control of a business has to be auditable: what exactly was done, on what data, by whom, and what went wrong. That is why explainability and an execution log are not “a feature from a list” for us — they are a condition for the product existing at all. A system that cannot explain its result has no business returning one.
03 · Principles
Every rule has not only a formulation but proof — a place where you can see it right now. A principle with nothing behind it counts as a declaration.
Everything that is not implemented carries a “Target architecture” badge. Demonstrations are labelled as demonstrations. If we have no customers, we write “0 customers” rather than putting up placeholder logos.
Proof: the “Customers: 0” stamp in the mission file above, and the badges on every department description
The owner states the result they need, not an instruction for a model. Breaking the task into subtasks, choosing the executors and ordering the work is AI CEO’s job — not the user’s.
Proof: the delegation demo on the home page — four tasks broken down across departments
At every step you can see which department is working, what it received and what it returned. A black box has no place in running a business: a decision nobody can explain can be neither challenged nor fixed.
Proof: tracing one task through the three layers on the AI Operating System page
The finance department prepares data, calculations and options. Payments, outbound messages to customers and other actions with external consequences are never performed without explicit human confirmation.
Proof: the Control layer — rules, permissions and the signing limit in the architecture
Tasks, workflows, agent memory and the execution log are exported as open JSON. The portability commitment is published in advance — before any of that data exists.
Proof: the commitments section on the pricing page
No page and no module counts as finished until an independent role has checked it: automated checks, visual review at three viewport widths, and verification against the live site.
Proof: the flight log below — milestones carry verification results, not just the fact of publication
04 · Architecture
We publish the design of the product before implementing it on purpose: that way it can be argued with early, and a finished module can later be checked against the promise. All three compartments below are target architecture.
Who is allowed to act, within what limits, and what requires a human’s confirmation. The analogy from an ordinary company is the internal regulation and the manager’s signing limit. This layer exists so that automation cannot take an irreversible action behind the owner’s back.
The layer that makes decisions: breaking a task into subtasks, assigning executors, tracking deadlines, checking results and folding the parts into one answer for the owner. This is where the company structure lives — the reason the project was started at all.
Tools, integrations and steps whose outcome must be predictable: send an email, read a spreadsheet, call an external service. The model does not “improvise” here — it calls a specific tool, and the fact of that call lands in the execution log.
Why publish it early: an architecture written before the code is a commitment, not a description of something finished. Once a module is implemented it can be compared against this page. If the implementation diverges from the plan, we will update the description and state the reason for the divergence — that is recorded in the log.
05 · Crew
We do not publish faces, names and job titles “like a large company”, because there is no large company. Instead — an honest list of roles with the area each one is accountable for. The internal rule of the project is simple: “built it and accepted it yourself” does not count as work.
Owner · final decisions · priorities
Sets the task, approves the architecture and any commercial promise, changes priorities. This is the role that on 31 October imposed the moratorium on mass pages and demanded that we build the product rather than its presentation — and on 1 November lifted it, once the order of work was settled. Public promises and pricing are not published without sign-off.
Code · architecture · publishing
Writes the pages and modules, owns the design system, bilingual support, accessibility and deployment with a backup. Has no right to declare their own work accepted — only to hand it over for independent review.
Verification · evidence · statuses
Checks every artefact against the facts: files exist, commands reproduce, numbers add up, the live site responds. Issues a VERIFIED or REJECT status and keeps the evidence — screenshots and reports the developer cannot rewrite.
Usability · WCAG · three viewport widths
Audits published pages against Nielsen’s heuristics and AA accessibility requirements: contrast, tap target size, behaviour without JavaScript and with reduced motion. Findings are recorded with coordinates, not in general terms.
Competitors · pricing · unclaimed niches
The twenty platform reports and the comparison matrix are this role’s work. One rule applies: data comes only from public sources, and anything unverifiable is marked “n/a” rather than filled in by guesswork. Financial figures for private companies are not invented.
Marketing · sales · operations · finance · people · research · development · support
The eight departments and AI CEO are the future executors inside the product, not members of our team today. They will appear as they are implemented and will operate under the Control layer: with permission boundaries, an action log and human confirmation for irreversible operations.
Why no names and photos: publishing a “team” of invented faces is exactly the kind of dishonesty that principle 01 was written against. When people appear on the project who are ready to answer for it publicly, their names will be published here — with a real area of responsibility.
06 · Flight log
Entries are ordered by date and contain only what can be verified: page publications, automated check results and decisions to change priorities. There are no interim victories like “secured an investor” here — there were none.
The project was created from scratch: domain, certificate, web server and publishing path were checked so that neighbouring sites would not be affected. The design system was assembled (dark theme, two accents — amber and cyan), the bilingual runtime was written and the first market research reports were connected.
Ten reports from the required list and ten more beyond it, a comparison matrix and an opportunity map with priorities. The main finding: everyone has an agent builder; the gap is in delegating a whole task and in making the route visible.
The first honest cycle: automated checks found 41 defects, all of them were fixed, and the following run gave 203 checks passed with no failures. A later visual review lowered the status to NEEDS_FIX — and that stayed in the log too, rather than being deleted.
Three pages in two languages: an architectural cross-section of the system, the AI CEO card and a price list with billing units. Every capability description carries a “Target architecture” badge, and no sales are announced on the pricing page.
On the morning of 30 October the DNS record did not exist; by night the blocker was removed by the owner. A Let’s Encrypt certificate was issued, the site answers over HTTPS in two languages, and publishing was additive with a backup — without editing the configuration of neighbouring sites.
The lead stopped mass page creation and demanded that the product be built: architecture, an inventory of existing systems, a prototype plan and integrations. The rule “DONE ≠ VERIFIED” was introduced — work does not count as finished without independent verification.
The premium animation layer was connected additively and respects reduced-motion. Median cumulative layout shift on the live site dropped from 0.053–0.20 to 0.002–0.005, and tap targets were brought up to 44 pixels.
Pages are being published again — one at a time, in priority order, with the honesty requirements intact. This page, “About the platform”, is part of the same series. The product work has not stopped: the prototype is being built in parallel.
The next milestone is not a page but a working loop: accept a task, break it into subtasks and assign them to executors in an isolated environment. As soon as it starts passing independent verification, an entry with a result — not a promise — will appear here.
07 · Roadmap
The order of waves comes from the opportunity map assembled after reviewing twenty platforms. We deliberately do not put dates on it: promising “launch in March” with no working code is the same dishonesty, only with a calendar attached.
Public sections, a description of the architecture and of the pricing model before sales start, bilingual support and automated checks on every page. The goal of this wave is that any statement on the site can be verified.
An isolated loop in which a task is accepted, broken into subtasks and assigned to executors. We start with the research → marketing → sales chain, because that is the one that shows the owner a visible result fastest.
Access rules, limits and human confirmation for actions with external consequences. Connectors to external systems are built only after the execution log records every call to them.
Eight departments with memory and their own quality criteria, billing according to the published price list, and the first real deployments. Only at that point will customers and metrics appear on the site — because until then there honestly are none.
The order of waves may change — and if it does, an entry about it will appear in the flight log above with the reason stated. Quietly rewriting the plan after the fact is what we consider the worst of the options.
08 · Questions
The questions this page is most often searched for. The answers repeat what is written above — deliberately, so you do not have to go looking.
09 · Where to look next
Everything listed here is published and answers over HTTPS. Sections that do not exist yet are deliberately not shown — a broken link is worse than no link.
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.
AI Operating System · LayersControl, Agentic, Execution
A breakdown of the three layers of the system and what each one is responsible for. Target architecture.
AI CEO · FunctionsWhat the decision centre does
Task intake, decomposition, assigning executors, tracking deadlines and consolidating the result.
Pricing · Price listWhat a single agent operation costs
The planned model: VU and MC units, a line-by-line price list and data portability commitments.
10 · Next step
Open the delegation demo on the home page and compare it with what is written here. If you find a mismatch between a promise and the actual state of things, write to us: we will fix the page, not the wording in a report.
The demo runs on pre-written scenarios and is not connected to a live AI service. That is stated on the demo page itself as well.