Series voir-pages
page 11 / 26
department 6 of 8
implemented 0 functions

Research is not “find a flattering number”. It is a question, a source with a grade and an action somebody else can verify.

The research department in VOIR AI runs the full loop: from formulating the question to an action card that the owning department accepts. The question is fixed before the search starts — together with the decision it informs and the “enough data” criterion. Every fact carries a source, a date and a reliability grade from A to D; a position with no data is marked “no data” instead of being closed with a plausible-looking estimate. Evidence is synthesised on a board where every thread between cards rests on a source, and every conclusion rests on threads. The insight then leaves not “into the air” but for a specific department as an action card: what to do, on what grounds and what to verify afterwards. Below is an interactive demonstration on fictional data: four stations of the bench and the tables the diagrams are built from. Next to it is the honest picture: what the department should do in the target architecture, what already works inside our own tools, and how the department is wired into the AI CEO cycle.

VOIR AI product · zero functions implemented Marisova ecosystem · research logs and acceptance reviews Demonstration · sources and numbers are fictional

Open the research bench Honest statuses

Department logbook · functions

VOIR-RES/11

QST Question brief and search boundaries not implemented

SRC Source registry with an A–D grade not implemented

INS Evidence board and synthesis into an insight not implemented

ACT Action card for the owning department not implemented

BENCH Research bench · 4 stations demo

HUM Decision on the insight · stays with a human rule

The “not implemented” labels are not decoration: not a single function of the research loop is implemented in the VOIR AI product. The functions that already work as our internal tools are listed separately in the status table.

Demonstration · research bench

Four stations: question, sources, insight, action

Research in VOIR AI is built as a bench on which a question travels from formulation to an action card. The first station is the question brief: what exactly needs to be known, which decision the answer informs, where the search boundaries are and on what sign the research stops. The second is sources: a registry where every fact has a source, a date received and a reliability grade from A to D. The third is the insight: an evidence board where fact cards are linked by threads into conclusions. The fourth is action: the card that leaves for the owning department together with its grounds and a verification sign. Each station is duplicated by a table or a list — the diagram is built from the data, not the other way round. The markets, companies, sources and numbers are invented by us for illustration.

Structural numbers of the page: 4 department functions, 4 bench stations, 4 source grades, 8 departments in the architecture, 0 functions implemented in the product.

Research bench · question → sources → insight → action

Demonstration · data is fictional

Station 1 · Question: the brief is fixed before the search

A question is not a topic “to look at the market”, but a contract: which decision it informs, what counts as an answer, where the search boundaries are and on what sign the research stops. Without a brief the search expands endlessly and its result cannot be checked: it is unclear whether the report answers the customer's question or one the researcher asked themselves. Below is an example brief for the fictional task “Should we enter the B2B subscription segment” of a fictional company.

Question · code Q-241: is there stable demand for a B2B subscription in the “tools for studios” niche — yes or no, and on what grounds
Decision the answer informs The owner decides: open the line, postpone it by two quarters, or close the question. The report must give grounds for all three options, not only for “yes”
Customer The marketing department; the answer then goes to sales and finance as an input for the unit economics
Search boundaries Public sources, open price lists of competitors, industry reviews of the last 24 months. Closed reports and paid databases are not used — we do not have them
“Enough data” criterion At least three independent sources per conclusion, of which at least one graded A or B. If there are none, the conclusion is marked “no data” instead of being filled in by guesswork
Deadline and format A one-page report: question, conclusions, source registry, positions with no data, action cards (demo format)

Demonstration: the question, company, boundaries and criteria are fictional. This is what the brief structure looks like in the target architecture — the research brief is not implemented in the VOIR AI product.

Station 2 · Sources: every fact has its own grade

A fact without a source is an opinion. That is why every fact in the report carries four fields: source, date received, a reliability grade from A to D, and what exactly this source confirms. The grade is assigned before synthesis and is never raised afterwards to make a conclusion “hold together”. A position where no data could be found stays in the registry marked “no data” — visible to the customer rather than drowned in the text. Below is a fictional registry of six sources for the fictional question Q-241; the exact values are in the table under the diagram.

A · primary source, verifiable
B · industry review with a method
C · publication without a method
D · no data, position open

Source registry — the data source for the diagram above. All sources, companies and numbers are fictional. The grade was assigned before synthesis and was never changed afterwards.
ID	What kind of source (fictional)	Grade	What it confirms
S-01	Open price list of the competitor “Studio Plan”, pricing page, snapshot of 12.09	A	Three subscription tiers, prices, availability of an annual plan — can be re-checked at any moment
S-02	Public filings of “Studio Plan” in the corporate register, extract of 03.10	A	Revenue dynamics over two years, share of the subscription model in the revenue structure
S-03	Industry review “Creative Tools 2026”, method described in the appendix	B	Niche size estimate, distribution by buyer type, 24-month forecast
S-04	Interview with a studio founder on an industry podcast, transcript of 21.08	C	Reasons the studio moved to a subscription; a single voice, non-representative sample
S-05	Discussions on two public studio forums, a sample of 40 threads over 6 months	C	Recurring complaints about annual prepaid plans; self-selection, grade no higher than C
S-06	Churn after the first year — no public data exists	no data	Five registry positions stayed open; the churn conclusion in the report is marked “no data”

Demonstration: the sources are fictional. The rule of the target architecture: a grade is assigned before synthesis and is never raised for the sake of a “prettier” conclusion; a “no data” position stays visible in the registry.

Station 3 · Insight: an evidence board where every thread rests on a source

Synthesis is not “write the conclusion more beautifully”, but showing what the conclusion is assembled from. On the evidence board each card is one piece of evidence with the grade of its source, and each thread between cards is a link with a type: confirms, contradicts or requires verification. A conclusion counts as established when threads from at least three independent cards lead to it and at least one of them is graded A or B. If there is a single thread, that is a hypothesis, and it is marked as a hypothesis. Below is a fictional board for question Q-241; the list of links and their types is under the diagram.

thread “confirms”
thread “contradicts”
thread “requires verification”

E-01 → conclusion · confirms The niche leader has kept the subscription model for two years running: the demand is not a one-off experiment. Grounds — S-01 (grade A).
E-02 → conclusion · confirms The subscription share of revenue is growing according to public filings. Grounds — S-02 (grade A).
E-03 → conclusion · confirms The niche size is estimated by a review with a described method; the estimate is consistent with the order of the leader's revenue. Grounds — S-03 (grade B).
E-04 → conclusion · contradicts Complaints about annual prepaid plans show that the shape of the subscription matters more than the fact of one. The contradiction is not hidden: it enters the conclusion as a condition — “monthly billing only”. Grounds — S-05 (grade C).
E-05, E-06 → conclusion · require verification Churn after the first year and acquisition cost remain without data. The conclusion is marked incomplete: an “open the line” decision without these two numbers would be taken blind.

Demonstration: the board, cards and links are fictional. The rule of the target architecture: a single thread is a hypothesis, not a conclusion; a contradicting piece of evidence enters the text of the conclusion instead of being removed.

Station 4 · Action: the insight leaves as a card

A report nobody read is worth as much as no report at all. That is why the result of research leaves not as text “to nowhere” but as a set of action cards: each names the receiving department, the action itself, the grounds with a reference to the source and the sign by which the action is later verified. The action card is accepted by the owning department, not by the researcher: research has no authority to decide for marketing, sales or finance. Below are fictional cards for question Q-241.

Card A-1 · recipient: marketing Test a monthly plan against the annual prepaid one on your own audience. Grounds: E-01, E-02, E-04 (S-01, S-02 grade A; S-05 grade C). Verification sign: second-month renewal rate compared with the competitor's annual plan.
Card A-2 · recipient: finance Build the unit economics of the subscription under monthly billing: acquisition cost, churn, payback period. Grounds: positions E-05 and E-06 are marked “no data” — these numbers must come from our own figures, there are no external ones. Verification sign: the model balances without assumptions that cannot be named.
Card A-3 · recipient: sales Prepare a subscription conversation script for the studios that complained about prepaid. Grounds: E-04 (S-05, grade C — self-selection, so the script is tested on the first ten conversations rather than rolled out at once). Verification sign: first-touch conversion no lower than the current one-off sales model.
Card A-4 · recipient: the owner The “open / postpone / close” decision stays with a human. The research department passes grounds for all three options and explicitly names what is missing for a confident “yes”. Grounds: the whole Q-241 report. Verification sign: the decision is recorded with its reason, not “after a discussion”.

Demonstration: the cards are fictional. The action hand-off loop is not implemented in the VOIR AI product; in the Marisova ecosystem a similar principle already works in our internal tools — see the status table.

Source scale

Reliability grade: A, B, C and an honest “no data”

A grade is not a “do I like this source” rating, but an answer to one question: can this fact be re-checked and who is accountable for it. The scale is declared in advance and applied the same way to every report of the department. A grade cannot be raised afterwards: if a conclusion does not hold on the available sources, it is marked as a hypothesis and the sources as insufficient. All four grades below describe the target architecture: the scale is not implemented in the VOIR AI product.

Target architecture A

Primary source

A document or data you can open and verify: public filings, an official price list, a register, your own export from a system. The party that published it is accountable for the fact.

A conclusion may rest on such a fact alone
The date received is mandatory: primary sources go stale
If it cannot be verified, it is no longer grade A

Target architecture B

Review with a method

An industry study, analytical report or statistics that describe how the data was collected: sample, period, counting method. You can read the method and disagree with it — that is exactly what makes it grade B.

Can hold a conclusion together with one grade-A source
A discrepancy between the methods of two reviews is recorded, not averaged away
Without a described method, a review falls to grade C

Target architecture C

Publication without a method

An article, a post, an interview, a forum discussion, a practitioner's comment. Useful as a signal and as raw material for hypotheses, but not as the ground of a conclusion: it is unknown who counted it and how, and the sample is almost always self-selected.

A conclusion never rests on grade C alone
Three independent Cs can point a direction but do not replace an A or a B
Such a fact enters the report marked “signal, requires verification”

Target architecture D

No data

A brief position for which no sources could be found. This is not “low reliability” but a separate state: the fact does not exist, and the report must say so plainly. Substituting a plausible-looking estimate is the most frequent and most expensive violation in research.

The position stays visible in the registry instead of being deleted
A conclusion that lacks data is marked incomplete
“No data” is a normal result of the department's work

How the grade is applied in practice

Conclusion threshold

At least three independent sources per conclusion, of which at least one A or B. Independence is checked by the origin of the data, not by the number of links: five publications that reprinted one news item are one source, not five.

No upgrades

The grade is assigned when the fact enters the registry and never changes afterwards. If a conclusion feels weak, strengthen the search, not the grade. Every attempt to raise a grade after the fact is logged as an incident.

Shelf life

Every fact has a date received. Market data goes stale faster than the structure of the market does: a report older than the horizon declared in the brief returns for reassembly instead of being served as current.

Contradiction is part of the result

If two grade-A sources say different things, that is not “pick the convenient one”. Both values enter the report, the discrepancy is named, and the conclusion is either limited by a condition or marked incomplete.

Department roles · target architecture

The role contract: input, work, output, artefact

The research department is three roles, each with its own contract. The split is deliberate: whoever formulates the question should not be the one to declare it closed, and whoever synthesises conclusions should not be the only one searching for sources. The contract describes what the role takes as input, what it does, what it must hand over and which artefact stays behind. A role without a contract is a “smart assistant” that helps one day and invents things the next; a role with a contract can be checked. All three roles below describe the target architecture: not one of them is implemented in the VOIR AI product.

Target architecture SCT

Source scout

InputThe question brief: search boundaries, data horizon, “enough” criterion
WorkSearches sources within the declared boundaries, assigns the grade on entry, records the date received and what exactly the source confirms
OutputThe source registry and the list of positions with no data
ArtefactRegistry: ID, source description, grade, date, confirmed fact, independence flag

Target architecture ANL

Synthesis analyst

InputThe source registry and the sufficiency criterion from the brief
WorkAssembles the evidence board: cards, threads, contradictions. Checks the conclusion threshold and the independence of sources, not their count
OutputConclusions with their grounds, hypotheses, and positions marked incomplete
ArtefactEvidence board: card, grade, thread type, linked cards, conclusion status

Target architecture EDT

Report editor

InputThe analyst's conclusions, the source registry, the question brief
WorkBrings the report to the customer's format and breaks it into action cards: recipient, action, grounds, verification sign
OutputA one-page report and a set of action cards per department
ArtefactReport: question, conclusions, registry, positions with no data, action cards, version and author

Statuses

Two status columns: the blueprint and our own workshop

Research is a field where a blueprint is especially easy to pass off as a working product: “AI studies the market itself” sounds plausible on its own, and any text answer of a model looks like a finished report. That is why the table is split into two columns. The left one is what is implemented in the VOIR AI product (right now, not a single function of the loop). The right one is what already works inside the Marisova ecosystem as our own tools. The right column must not be read as a promise to a customer: our internal work is not sold and not transferred.

Status of the research department functions as of 3 November 2026. Verified against the code, configuration and live URLs — not against plans.
Function	In the VOIR AI product	In the Marisova ecosystem (internal)
Question brief (QST)	not implemented Target architecture: question, the decision it informs, search boundaries, sufficiency criterion	partially works Internal tasks for agents are formulated with boundaries and acceptance criteria in the task card — our working process, not a product. Not provided to customers
Source registry and grades (SRC)	not implemented Target architecture: the A–D scale, date received, independence flag, no grade upgrades	partially works In internal research tasks, sources are recorded in reports with a date and a link. The grade scale as a formal mechanism does not exist. Not provided as a service
Synthesis into an insight (INS)	not implemented Target architecture: evidence board, thread types, conclusion threshold, contradiction as part of the result	partially works Independent acceptance of research against criteria exists as the ecosystem's internal QA loop; there is no evidence board. Not provided as a service
Hand-off of action (ACT)	not implemented Target architecture: action card with recipient, grounds and verification sign	partially works Research results go into tasks with a named recipient and criteria — the internal task log. Not provided to customers
Research bench (the demonstration on this page)	demonstration Four stations with fictional markets, sources and numbers; not connected to a live AI service	—

Link to the AI CEO

The research department does not choose what to research

Research in VOIR AI is the execution part of the AI CEO cycle. The AI CEO takes the task: study a niche, test a hypothesis, compile a report on a competitor, assess whether there is enough data for a decision. It also distributes the work across the department's roles and folds the result into a report for the owner. The department's artefacts pass an independent acceptance by a role that did not do the work. The list below shows exactly what the research department does at each of the five steps of the cycle.

01 Task set The department receives a question brief and a sufficiency criterion fixed before the start: what needs to be known, which decision the answer informs, within which boundaries to search, which data horizon counts as current. The department does not pick its own questions — otherwise it starts researching what interests it, not what the owner's decision needs.
02 Analysis and assignment The AI CEO decides which part of the work goes to research and distributes it across the department's roles: search and grading — to the source scout, the evidence board and the conclusion threshold — to the synthesis analyst, the report format and action cards — to the report editor. The autonomy grade is fixed: what the role does itself and what always goes to the owner for confirmation — any decision that changes the plans of other departments.
03 Execution The roles run the task through the functions: source registry, evidence board, report, action cards. Every role leaves a trace of its actions and hands the artefact further together with its sources or an explicit “no data” mark. Where data is missing, the step escalates instead of being padded with a plausible-looking market estimate.
04 Independent check A QA role that did not run this cycle checks the artefacts against the criteria of step 1: does every fact have a source, a date and a grade; is the conclusion threshold met; were any grades raised after synthesis; are the “no data” positions visible; did contradictions enter the text of the conclusion. A false-green gate is checked separately: an acceptance that signs off any report verifies nothing.
05 Acceptance and report The AI CEO folds the checked artefacts into one report for the owner: what is established and on which sources, what remains a hypothesis, which data is missing, which action cards went to which departments, which decisions stayed with a human. Contradictions and no-data positions enter the report as they are, without smoothing — otherwise the owner decides based on a picture that does not exist.

Artefacts the department leaves behind

Question briefquestion, the decision it informs, search boundaries, data horizon, sufficiency criterion
Source registryID, description, grade A–D, date received, what it confirms, independence flag
“No data” listbrief positions for which no sources could be found, with the reason
Evidence boardfact cards, thread types, contradictions, status of every conclusion
Reportquestion, conclusions with grounds, hypotheses, limitations, version and author
Action cardsrecipient, action, grounds with a source reference, verification sign
Incident logcases of grades raised after the fact, of “no data” replaced by an estimate, of expired reports

Touchpoints with neighbouring departments

AI CEO → research The task arrives with a question brief and a sufficiency criterion. Without a brief the search has no stopping rule: research turns into an endless market overview and its result cannot be accepted.
Research → marketing Action cards on demand, segments and positioning go to AI Marketing with grounds and a verification sign. Marketing accepts the card or returns it with an objection — both are recorded.
Research → sales The report on competitors and buyer objections goes to AI Sales: conversation scripts are built on sources, not on the department's hunch.
Research → finance Market data and niche size estimates go to AI Finance as the external input of the unit economics. “No data” positions are passed as they are: finance must know which figures of the model came from outside and which simply do not exist.
Research → development Connecting external data sources, register APIs and industry databases is a task for the development department. Research formulates the interface requirement: source, date, idempotent re-upload, grade on entry.
Research → owner and QA All artefacts go to independent acceptance, and any conclusion that changes the plans of other departments escalates to the owner. The department never declares a decision made on its own.

How a single executor is built — on the AI agents page; who exactly works in the departments — on the AI employees page; how a staff assembles several departments under one task — on the AI teams page.

Boundaries

What the research department never does

Research is a field where the temptation to “complete the picture” is at its maximum: an incomplete answer looks worse than an invented one, and the customer rarely sees the difference. That is why the limits are built into the function contracts, not into settings: no role of the department can bypass them, including the role responsible for research.

Does not decide for the owning department

Research prepares the grounds for a decision: conclusions, sources, positions with no data, action cards. Deciding “enter the segment”, “change the price”, “close the line” belongs to whoever owns the result — a human, or a department with the authority. A report is not an order.

Does not replace “no data” with an estimate

The most expensive violation in research is a plausible-looking number in place of a missing fact. A position without sources stays in the registry as “no data”, visible to the customer. A conclusion that lacks data is marked incomplete instead of being polished into a confident-looking one.

Does not raise a grade after the fact

The grade is assigned when the fact enters the registry and never changes afterwards. If a conclusion does not hold on the available sources, strengthen the search — do not reclassify a publication without a method into an “industry report”. Every upgrade attempt is logged as an incident.

Does not pass five reprints off as five sources

Independence is checked by the origin of the data. Five publications that quoted one news item give one source, not five. Without this check the “three independent sources” threshold turns into a formality that creates the appearance of confirmation.

Does not hide contradictions

If a source speaks against the conclusion, it enters the report with the thread type “contradicts”. The conclusion is then either limited by a condition or marked incomplete. An evidence board where every thread is one colour is a sign of a biased search, not of a world that agrees.

Does not draw metrics that do not exist

No “improved forecast accuracy”, no “research time cut by X percent”, no customer reviews or logos. The sources, markets and numbers in the demonstration on this page are invented and marked as a demonstration: they show how the stations are built, not the results of a rollout.

Questions

What people ask about AI Research

Are the markets and sources in the demonstration real?

No. All markets, companies, sources, grades and numbers in the demonstration are invented by us to show how the stations are built: what a question brief looks like, how a fact gets a grade, how evidence is synthesised on a board and how the insight leaves as an action card. The project has no customers, so there are no real studies we could show either.

Does the research department already work in the product?

As a VOIR AI product — no. All four functions on this page describe the target architecture: not one of them is implemented in the product code. Separately, there is the Marisova ecosystem, where setting research tasks with boundaries and criteria, recording sources with dates and independent acceptance of results exist as our internal tools. They are not part of the VOIR AI product and are not provided to customers.

Where does the system get market data from?

In the target architecture — from the sources declared in the question brief: public documents, open price lists, industry reviews with a described method, public discussions. There are no connections to paid databases, closed reports or analytics services in the VOIR AI product, and we do not write that they “will appear soon”. Every fact carries a source and a date; a position without a source is marked “no data”.

What is a source grade and why is it needed?

It answers one question: can the fact be re-checked and who is accountable for it. Grade A — a primary source you can open and verify. Grade B — a review with a described method. Grade C — a publication without a method: a signal and material for hypotheses, but not the ground of a conclusion. A separate “no data” state — a position for which no sources could be found. The scale is declared in advance and applied the same way to all reports; a grade cannot be raised after synthesis.

How is an insight different from a “whatever came out” conclusion?

An insight on this page is a conclusion whose ground is visible: threads to specific fact cards, each of which has a source with a grade. A conclusion has a threshold — at least three independent sources, of which at least one A or B. If there is a single thread, that is a hypothesis, and it is marked as one. Contradicting evidence is not deleted: it enters the conclusion as a condition.

What efficiency metrics does the page show?

No rollout metrics at all. There is no “improved forecast accuracy”, no “research time cut”, no SLA, no customer reviews or logos. The numbers of the demonstration are fictional; the counters in the text are structural facts: four department functions, four bench stations, four source grades, eight departments in the architecture, zero functions implemented in the product. Anything else would be invention, and we will not invent metrics for a product that is not launched.

How is the research department linked to the AI CEO?

The department does not choose what to research. The task and the sufficiency criterion are taken by the AI CEO; it also distributes the work between the department's roles — the source scout, the synthesis analyst and the report editor, — folds the result into a report for the owner and decides what to change in the next cycle. The department's artefacts pass an independent acceptance by a QA role that did not run this cycle. In detail — on the AI CEO and AI CEO Cycle pages.

Can we connect our databases and analytics?

Right now — no: there are no connections to external databases, BI systems or analytics services in the VOIR AI product. Integration is a function of the development department, described as target architecture. What really exists today is listed in the status table on this page: it has two columns — “in the VOIR AI product” and “in the Marisova ecosystem” — and the second column is not a promise to a customer.

What to read next after the research loop

A report does not exist apart from the system: it enters the AI CEO cycle, leaves as action cards to the owning departments and becomes an input for the team assembled under a task. The three pages below show these links from different sides.

Process

See pricing Ask a question about the platform