How does an AI organisation brain work?
An organisation brain is not a chatbot or a note-taking app. It is the layer between your assignment and the AI that carries it out: it selects the context, passes the work to the right role, holds back the result until a human approves it, and only then writes it back to memory. This page walks through that cycle, including its boundaries.
One assignment, five phases
Everything below on this page is connected to this cycle
An assignment comes in, receives context, is delegated, passes by a human and only then becomes part of the memory
The four visible steps are shown on the Jarvis hub; here, writing back counts as a fifth phase, because that is precisely where the difference lies between a smart chat and a brain that will still know tomorrow what was decided today. The example below is schematic: it shows the form of the cycle, not a real customer assignment.
What an organisation brain is, and what it is not
In short, such an organisation brain is a shared, controlled memory for your company, plus the orchestration of who and what is allowed to draw from it. It stores the things currently held in people's heads, chats and separate documents: how you work, what you agree, which version applies, who approved something and why an earlier choice was reversed. And it passes that knowledge to the AI tools you already use, instead of each tool maintaining its own incomplete version of the story. Jarvis is our implementation of this: a knowledge and orchestration layer between your people, your sources and ChatGPT, Claude and Codex.
It is useful to first say what it is not, because the term borders on three things that closely resemble it
It is not a note-taking app. A second brain in the personal sense, like those people build in note-taking tools, belongs to one person and has no roles, no approval and no demonstrable audit trail. An organisation brain is shared by definition, and within a month, shared memory without permissions and provenance becomes a cluttered attic.
It is not a new chatbot.
Your AI tools remain in place and you keep your own subscriptions; the brain sits alongside them and supplies them with information. And it is not a search index of your files. A search index finds documents. A brain knows which document applies, since when, to whom, and who determined that.
The difference lies in that final word: validity
Almost any AI setup can retrieve text. Almost none can explain why that text is true now. That difference is precisely what an organisation brain adds, and it is also why the rest of this page is about metadata, statuses and approval rather than smart answers. The intelligence comes from the language model you already have. The reliability must come from somewhere else.
If you want to test the idea without software first, you can
The free company-brain template is the manual version: a single fillable file that tells your AI who your company is, what you provide and how you work. That works surprisingly well until several people start writing in it and no one knows which version applies any more. From that point onwards, you no longer need a file but a system, and that is where this page begins.
Why every AI session starts empty
Anyone who works with AI every day knows the pattern. This morning, you explain in ChatGPT how your quotations are structured, what tone you use with customers and which margin you never reveal. This afternoon, you open Claude for another piece of work and that conversation starts from zero. Tomorrow, your colleague uses the same tool and explains the same things again, but slightly differently. Two different truths do not arise because someone is careless, but because there is no place where the truth belongs.
There are three causes, and they compound one another
The first is the session boundary: by design, a chat is a conversation, not an archive. What it contains is context, not knowledge. The second is the tool boundary: memory features offered by the tools themselves reside in that single account and that single tool, so as soon as you switch tools or a colleague joins in, they disappear. The third is the human boundary: most of what a company knows has never been written down. It is in the head of the person who devised it, and that works until that person goes on holiday.
Skills solve part of that, and we share them for free: a skill is an instruction file that gives your AI a consistent working method for one task
But a skill teaches the AI how to do something, and that is different from remembering what applies in your organisation. The skill is neither the engine nor the memory: it says how a quotation should be structured, not which rates you charge or which customer received an exception last month. Those two layers need each other. The skill teaches the AI, the brain is where the memory lives.
The consequence of that empty start is not just lost time
The main problem is that you can never say with certainty what an AI based something on when it wrote it on your behalf. As long as that remains the case, you can use AI for drafts, but not for work for which someone is accountable. The entire structure below exists to restore that certainty, and this comes at the expense of speed in exactly one place: approval. That trade-off was deliberate. The theory underpinning this design, from stateless models and context rot to retrieval and MCP, is explained in detail in how an AI brain forgets nothing between chats.
One sentence, six fields around it
This is the brain's smallest building block and, at the same time, its most important one
A memory item is not the sentence itself, but the sentence plus everything needed to assess whether you may trust it
Without those six fields, a memory item is a claim; with those fields, it is a verifiable fact. The content below is a schematic example, not a real client rule.
We use this system ourselves every day. Every agent session, every task and every decision is logged in it and can be reviewed. Our own product documentation, pricing files and release status are stored in that brain and published from it, so this explanation has the same source as the one we work with ourselves. If a page differs from that source, the page is wrong. The part that is not yet finished appears below, explicitly identified.
How knowledge enters
Knowledge enters through four routes, and the difference between those routes determines how strict the subsequent gate is. The first route is the person who records something: someone writes down an agreement, a rate or a working method and submits it to the brain. The second is a source that you connect: a document, a system or a file that the owner designates as the authoritative location.
The third is what emerges from the work
During an assignment, an AI role discovers something that is apparently true, for example that a client prefers to be updated by telephone, and proposes it as a memory item. The fourth is a proposal from outside your own brain: we can offer generic product knowledge or an improved working method.
That fourth route warrants a separate sentence, because it contains a boundary that many systems do not have
Our superbrain contains generic product knowledge, policies and proven patterns. A client brain contains the sources, people, rules and context of one organisation. Improvements can be proposed from the top down, but client content does not automatically flow in the other direction and is not used to train anything. The precise intent behind that separation is explained in the in-depth article on the superbrain and client brains.
What all four routes have in common: nothing enters as ready-made truth
A submitted item receives proposal status. It is immediately supplied with its six fields, because a proposal without a source cannot be assessed and is therefore returned rather than stored. Its status changes only after approval, and only then does the item count when the brain provides context for a subsequent assignment. That is why the memory grows slowly, and that is precisely the intention: a brain that swallows everything it encounters becomes unusable within a quarter.
There is a second filter as well, and it may be the least intuitive part of the entire design
When the brain retrieves context for an assignment, relevance is not the first criterion. It first filters by organisation, role, classification and purpose, and only then ranks by content relevance. A source may match the question perfectly and still not be permitted for the agent, the person or the task. Being able to find something and being allowed to use it are two different things, and applying them in that order is the difference between a useful assistant and a leak.
From proposal to active
Between entering and counting, there is one gate, and that gate is a person. In this system, an approval is not a button but a recorded event: an authorised person approves an exact, immutable version, with a consequence that can occur only once. The five requirements at the bottom of the figure are the properties that make an approval usable as evidence.
There is a pitfall that we explicitly identify because it can break any approval system: approval fatigue. If the system asks for a click at every small step, people start clicking blindly and the gate becomes a formality. The solution is not less control but better grouping: define low-risk steps in advance as policy, show only decisions with a real consequence, and bring exceptions to the fore. The detailed reasoning appears in human approval.
Source provenance: why knowing where something comes from is not a luxury
Take the sentence from FIG.02 again: a quotation remains valid for thirty days. On its own, that is a perfectly sound rule. But as soon as someone uses it, questions arise. Does it apply to all services or only to projects? Since when? Is it in the terms and conditions, in a price list or in the notes from a meeting last year? Who may change it? An AI that neatly retrieves the sentence and confidently quotes it does exactly the wrong thing: it gives you an answer you cannot verify.
Provenance is therefore not an additional field but the core of the memory.
If you do not retain the metadata, you retain only text, and text without context is why people start double-checking AI answers. The aim is not for the brain always to be right. The aim is for you to be able to establish within thirty seconds what an answer is based on, so that you can decide for yourself whether to adopt it. That is a more modest promise than infallibility and a far more useful one.
Conflicts are product data, not an error.
If the CRM, the price list and the contract state different amounts, their average is not an answer but a new falsehood. The brain should show that conflict together with the three sources and turn it into a review task. The conflict status and the eventual resolution themselves also become part of the trail, so that you can later see why that choice was made. In practice, this is one of the most useful features: it exposes contradictions that had gone unnoticed in the organisation for months.
Maintenance is also needed, and that may be partly automated
A scheduled session can review sources that have expired, check links and flag inconsistencies. What such a session may not do is independently replace pricing policy, contract text or customer knowledge. The result of automated maintenance is therefore always a review task, never a publication. The full details, including the minimum fields per item, are set out in source provenance and knowledge architecture; the broader design choices are set out in the knowledge section.
Enough evidence, as little content as possible
Logging is what enables you to establish afterwards what happened on your behalf
At the same time, a log that copies everything is a second copy of your trade secrets in a place no one thinks about
The principle is therefore clear-cut: record who did what and when, and which version was involved, but do not duplicate the content. The lines below are a schematic example of what such a trail looks like, not a real log file.
- request-id, time, actor, organisation and route
- changes to authentication, roles, connectors and subscription entitlements
- versions of sources, memory items, agents and approvals
- tool action, result status, duration and error category
- support access, export, deletion and incident status
- full prompts and full answers
- the content of documents
- API keys and tokens
- payment card details
- special categories of personal data
If content logging is nevertheless temporarily required for an incident, that is not a silent exception: that logging is assigned a purpose, an owner, a time limit, an access policy and proof of deletion. Critical audit events are intended to be append-only and time-synchronised, only authorised roles are given access, and an export masks secrets and excludes data from other organisations. The full protocol, including the retention periods, is set out in the logging protocol, and what happens during an incident is set out in the incident process.
How to read that trail
A log that no one reads is a reassuring sound, not a control. The practical question is therefore: which questions can you answer with it, and where do you look. In the interface, this runs through three entry points. From a task, you see that task's entire progression: which context was provided, which role worked on it, which draft resulted, who gave approval and what happened afterwards. From a memory item, you see the reverse path: which versions have existed, why it was changed, who approved it and which source underlies it. And from a person or role, you see what happened under that authority.
The questions this answers are perfectly ordinary questions
Why did that message to that customer state a period of thirty days? Who approved our use of that wording? When was that rate changed and on what basis? Has anyone from outside the team ever looked at this file? With a standalone AI chat, the answer to those questions is by definition: we no longer know. With a brain that has an audit trail, the answer is a line containing a time, an actor and a version.
Two things are important to state honestly here
The first: a trail proves what the system did, not that a decision was wise. Logging is not a quality mark. It makes assessment possible, nothing more, and that is already a great deal. The second: how robust this is for your organisation depends on the release gates described later. The design is fixed and visible in the preview; the production controls that enforce it technically are still partly a work in progress.
For those who want to see it for themselves: the interactive demo is the real interface with fictitious sample data and works without an account
You can click through it, approve a draft, assess a knowledge proposal and see where the payment gate remains closed. The guided tour with an explanation for each screen is on the platform tour, and the register showing which documents are valid and which are still drafts is in the documentation.
What Jarvis deliberately does not do
The boundaries below are product features, not small print. They get in the way at precisely those moments when a system without boundaries seems faster, and that is the point. Each rule is a request you could make, followed by what happens.
What is not yet finished
Jarvis is a preview, and that is not modesty but a status. What is there is a working interface containing the core interactions: assignments, approvals, memory proposals, knowledge reviews, inbox, task statuses, subscription changes and the five-step onboarding process. And we identify what is not yet there just as precisely, because a product page that shows only the attractive part is exactly the kind of source we would reject in the brain itself.
Not yet connected or not yet demonstrated:
production authentication, the negative test set demonstrating that data can never leak between organisations through a client parameter, the OAuth connections to external tools, the automated payment flow, the provisioning of client brains and production observability. This also includes exercises that have yet to take place: an incident, recovery, export and deletion exercise, and an external security review. That list is why we currently present Jarvis as a preview rather than a finished product: we do not put a promise on display when its evidence is still sitting on the shelf. Precisely what is still missing for each item is stated on the security page.
The pricing reflects the same honesty.
The plans you see below are the current proposition, undergoing validation. Final amounts, VAT treatment and the payment provider integration will still be formally approved and tested end to end before launch. Automatic payment is therefore still disabled: you will not see a payment screen for a process that has not yet been completed behind the scenes. Choosing a plan and starting independently is the route we are building, and we will only open it once the payment, invoicing and withdrawal chain works correctly.
What is already correct: the demo is the real interface with fictional sample data and works without an account, the pricing model without a usage meter is a verifiable property rather than a promise, and the working method shown on this page is the one we ourselves use every day. The current status of each gate is available in machine-readable form in the documentation, and the underlying materials are in the trust centre, including documents that are still labelled as drafts there.
Try it yourself first, then start independently
You can start using Jarvis yourself, yet there is still an order below. An organisation brain that does not suit your situation is costly in a way that has nothing to do with the price. The sequence below helps with that: first the free template, then the demo, and only then a conversation.
available today The company brain template: the manual version A single fillable file that teaches your AI who your company is, what you provide and how you work. The free template is the paper-based predecessor of everything on this page: the same idea, without roles, without approval and without a log. Works in Claude, ChatGPT, Codex and Cursor, and you do not need anything from us to use it.
without an account The interactive demo: the real interface If the template becomes unmanageable because several people are writing in it, you will want to see what this looks like as a system. The demo is the working Jarvis interface with fictional sample data: approve a draft, assess a knowledge proposal, see where the payment gate remains closed. The explanation for each screen is in the platform tour.
independently or together Start independently, or first tailor it together You choose a plan and start independently. If business agreements or customisation are involved, we go through four things together: the use case, individual, team or multiple brains, the plan, the sources and AI accounts required, and the agreements concerning terms, privacy role and price. If it is not a good fit, we will say so, and the template from rung 1 is then a perfectly good final destination.
These are the current plans and the proposition is undergoing validation: the amounts have not yet been finalised. What counts towards capacity under each plan is stated on the pricing page, together with the reason why there is no meter for prompts, tokens or calls. Anyone who first wants to know how isolation, roles and release gates are arranged can go to security and privacy.
Where the evidence is
This page describes how it works. Every promise made here has a document or screen somewhere that explains it in detail. Below you can see where it is, sorted by what you want to know.
The choices behind the design
in-depth informationWhy it works this way and not another, explained by topic.
Source provenanceThe minimum fields for each memory item, conflicts as product data and why retrieval is also authorisation.DOC Human approvalThe five characteristics of a useful approval, and how to prevent approval fatigue.DOC Superbrain and client brainsWhat may flow from the top down and what never goes the other way.DOC Why there is no usage-based billingThe reasoning behind a fixed monthly amount instead of metering your usage.DOCThe evidence and documents
trustThe materials you may hold us to, including what is still in draft form.
Security and privacyIsolation, minimum permissions, support with consent and the outstanding points that still need to be demonstrated.SEC Logging protocolWhat is and is not recorded, integrity, access and retention periods.TRUST Privacy statementWhich data we process, for what purposes, for how long and what you may request.DOC Evidence fileThe register of what has been demonstrated and what is still a release gate.TRUST AI RegulationHow we view the European rules and Jarvis's role within them.TRUSTSee for yourself and get started
the three-step approachIn this order: free, take a look, and only then have a conversation.
The free business brain templateThe manual version of this entire story, in a single fillable file.FREE The interactive demoThe real interface with fictional sample data, without an account.DEMO The platform tourExplained screen by screen, with the corresponding route.TOUR Get started with JarvisChoose a plan yourself, or first go through it together in four steps.STARTThe surrounding layer
contextWhere Jarvis sits within our own offering, and what exists alongside it.
Jarvis in briefThe hub: what the product is, who it is for and what you are buying.HUB Plans and capacityThe four plans side by side, including what does and does not count.PRICE The AI employeeWhen you do not want to carry out the work yourself but want it prepared for you, with a human providing approval. From 950 euros per month, through Mansotti, the business of which TheSEO is the trading name.SERVICE The free skills libraryInstruction files that give your AI a consistent way of working, separate from the brain.FREE AI and automationThe service behind it: from individual skills to working automation.SRVWe provide the AI employee through Mansotti, the company of which TheSEO is the trading name, with the same core rule as above: output remains a draft until a person approves it. That is a different proposition from Jarvis. Jarvis manages the memory and coordination of the AIs you already have; the AI employee is for anyone who wants to hand over a recurring task completely.
Frequently asked questions
How does an AI organisational brain work, in brief?
In five phases. A task is recorded with its objective and owner. The brain determines which sources are permitted for that role and scope. The work goes to the appropriate AI role, which produces a draft with evidence. A human approves the exact version before anything is released externally. Only then does approved knowledge become part of the memory, together with its source, scope, version, validity, status and approval. Each phase leaves an entry in the log.
What is the difference compared with ChatGPT's or Claude's memory?
An AI tool's memory resides in that tool and that account. If you switch tools, or a colleague joins in, it starts again. An organisational brain sits alongside them and passes the same core knowledge and agreements to every connected AI, together with roles, provenance and approval. You retain your own subscriptions with your own provider; the brain does not replace them, but supplements them.
What exactly is contained in a memory item?
The text itself plus six fields. Source, with the file, system, owner and a traceable reference. Scope, meaning who it applies to. Version, with a content hash, reason for the change and the previous version. Validity, with the effective date, expiry date and review point. Status, one of five states: proposed, under review, active, conflicting or superseded. And approval, with the authorised actor, the time and the exact version that was approved.
What is logged and what specifically is not?
Logged: request ID, time, actor, organisation and route, changes to authentication, roles, connectors and subscription entitlements, the versions of sources, memory items, agents and approvals, the tool action with result status, duration and error category, and support access, export, deletion and incident status. Not logged by default: full prompts and answers, document contents, API keys and tokens, payment card details and special categories of personal data. If content logging is temporarily required for an incident, it is assigned a purpose, an owner, a time limit, access controls and proof of deletion.
Can the brain send something without my knowledge?
No. Email, publication, contract changes and payments are subject to the agreed human approval, and until that approval is given, output remains a draft. An approval points to an immutable version, is checked server-side against the role and scope, and has a consequence that can occur only once, so the same approval cannot accidentally cause two emails to be sent. The time, actor, reason and any revocation are recorded.
Does knowledge from different customers remain separate?
Yes, that is a product principle: customer brains are logically isolated, there is no unrestricted exchange between brains and customer content is not used for training. The TheSEO superbrain can offer generic improvements as a proposal, but a customer brain never adopts them silently across the organisation, and customer content does not flow back automatically. The negative test set intended to demonstrate this in production is still outstanding as a release gate, and we explicitly state that alongside it.
How much does it cost and how do I get started?
The current plans are Brain Start 9 euros, Brain Solo 29 euros, Brain Team 99 euros and Brain Business 249 euros per month including tax. The amount remains the same everywhere and only the tax component within it varies by country. The proposition is still being validated, so these amounts are not yet final. There is no metering of prompts, tokens or calls. You can get started in this order: first the free business brain template, then the interactive demo without an account, and then choose a plan and start by yourself. A conversation is available for business arrangements and bespoke work. The automatic payment route remains disabled until all release gates are green.
And now
If you take one thing away from this page, let it be this: the difference between a smart AI and a useful organisational brain lies not in the answer, but in the question you can ask afterwards. Where did this come from, who approved it, and since when has it applied? As long as you cannot answer those questions, AI remains a tool for drafts. Once you can answer them, you can entrust it with work for which someone is responsible.